• In less than a year’s time, memory stocks have gone from just another way to play the AI trade to arguably its biggest beneficiaries from a return perspective. 
• Micron, Samsung and SK Hynix now rank among the world’s top 20 most valuable stocks, each with market capitalizations well above $1 trillion. 
• Widespread shortages across HBM, conventional DRAM, LPDDR5 and NAND SSDs are affecting the industry, putting immense pricing power in the hands of memory companies. 
• The top AI processor companies have greatly increased memory content across their systems to meet the changing needs of frontier model developers. 
• While memory makers are pointing to a prolonged shortage, there are multiple key risks to stay aware of as it relates to the memory trade. 

Over the past 10 months, memory chip stocks have gone from being solid beneficiaries of the AI boom to capturing a massively outsized piece of the return pie. 

The inflection in Micron’s performance demonstrates this. From the beginning of 2025 to the end of August 2025, Micron added around $36 billion to its market capitalization, rising to $133 billion for a strong 37% gain. Since then, returns have exploded, with Micron (MU) soaring more than 900% from August 2025, and up over 1600% since the April 2025 low. 

Perhaps the most striking figure is that over these 10 months, Micron added more than $1 trillion to its market capitalization, which now sits near $1.35 trillion. Along with this, industry watchers expect the memory market to far exceed $1 trillion in revenue by 2027. 

This chart is a comparison of Micron’s stock performance since the April 2025 low, showing a gain of over 1,600%, far outpacing Nvidia, the semiconductor ETF, and the Nasdaq-100, highlighting the strength of the AI memory-driven rally.

The sheer velocity of the move reflects how quickly investors have repriced memory’s role in AI infrastructure. What was once viewed as a cyclical, commoditized segment of semiconductors is now becoming one of AI’s most drastic bottlenecks, as shortages spread across HBM, conventional DRAM, LPDDR5X and NAND SSDs. 

The I/O Fund has been covering this dynamic for nearly three years. We first explored it in our deep dives on AMD’s AI acceleration strategy and Lam Research’s leadership in HBM and DRAM equipment in the summer of 2023. In December 2023, we expanded on this theme in our analysis of the 2024 memory and PC rebound, highlighting the shift from cyclical demand to AI-driven secular growth “that is strong enough to transform commoditized hardware into a secular trend,” with HBM and high-performance DRAM “in the early stages of a multiyear growth cycle.” 

This thesis led us to make memory stocks some of our highest allocations of 10%+ in 2026, even as many market participants feared memory had topped for good. 

The question now is whether the memory trade has already run too far, or whether shortages and the upside in memory pricing support more upside. Below, you’ll see pricing power is still intact, with meaningful new supply not arriving until 2028, or later, with the only overhang being how pricing power flows to memory suppliers under long-term agreements.  

The discussion is data-driven, but also nuanced, because there may be no bigger debate in the market today than whether memory can continue its historic run. 

What Triggered the AI Memory Supply Crunch? 

First off, it is worth taking a step back to understand what brought about the shortages seen today. Traditionally, memory demand has been very cyclical, with much of the market driven by consumer spending on mobile phones and PCs. H2 2022 saw the memory market enter its worst cyclical downturn since the Great Financial Crisis. Sales and earnings plummeted from pandemic highs, with Samsung’s operating profit falling by 95% YoY in Q1 2023. 

This led to significant production cuts, with Micron and SK Hynix announcing reductions of 15%–25% and cutting capex by 40%–50%. Samsung also implemented meaningful cuts to conventional DRAM and NAND output in 2023. At the same time, both Samsung and SK Hynix began aggressively expanding HBM capacity for 2024, albeit from a relatively small base—with estimates placing HBM at just 8% of total DRAM sales in 2023. 

mid

Amid this, the AI market had its watershed moment: the release of ChatGPT in November 2022. Nvidia’s data center revenue would go on to soar 217% in its fiscal year 2024 (roughly calendar year 2023), its fastest growth rate during the AI era. 

This helped memory demand improve, but companies supported this demand by drawing down inventory from extremely high levels. Notably, at the beginning of 2023, Micron’s days of inventory outstanding, excluding write-downs, were 235. Without adjusting for impairments, it would have taken the company nearly eight months to sell its inventory. By the beginning of 2024, that figure had fallen drastically to 160. 

The decisions to cut capacity, reduce capex, and allow inventory levels to drop were logical, considering that these firms were fresh off their worst decline in a decade and a half. However, as data center demand continued to explode and HBM capacity remained relatively low, those decisions set the stage for the drastic supply shortages we are seeing today. 

This chart is a comparison of AI memory stock performance since the August 2025 low, showing U.S. memory up over 1,200% and South Korean memory over 700%, significantly outperforming the semiconductor ETF and Nasdaq-100.

HBM Shortage: AI Infrastructure’s Core Memory Bottleneck 

Memory shortages are surfacing across essentially all product types. However, perhaps the most easily identifiable shortage is in high-bandwidth memory (HBM). This comes as HBM is the type of memory packaged directly alongside AI accelerators, from NVIDIA and AMD GPUs to Google TPUs. HBM, a DRAM form factor, uses advanced packaging to stack up to 12 (and soon 16) DRAM chips, offering higher bandwidth, capacity, performance, and power efficiency compared to conventional DRAM. 

Micron, Samsung, and SK Hynix are the three companies that control the HBM market. Due to massive hyperscaler demand, all three are sold out of their HBM capacity for 2026. 

Rapid Market Growth Outpaces Supply Increases 

Looking back a few years, we can see how fast the HBM market has grown. Bloomberg Intelligence placed the size of the HBM market at $4 billion in 2023. Meanwhile, at the beginning of 2026, SK Hynix cited data from Bank of America placing the HBM market at $34.6 billion in 2025. Taking the $3.9 billion 2023 estimate, the HBM market grew by an astounding 198% CAGR from 2023-2025. Notably, BofA forecasts additional growth of 58% YOY to $54.6 billion in 2026. 

Considering this growth rate, making capacity investments at the scale required for supply and demand to balance was antithetical to the position of memory suppliers in 2023. Even if they wanted to, actually achieving this would not have been feasible, given that increasing production capacity is a lengthy and expensive process. Notably, these three companies combined for free cash flow of -$18.5 billion in 2023. Looking at conventional DRAM, a knock-on effect from HBM imbalances is exacerbating shortages there. 

HBM Reallocation Is Tightening DRAM Supply 

In conventional DRAM, which centers around double-data rate 5 DRAM (DDR5), the situation is somewhat similar but has different mechanics. DDR5, and increasingly low-power DDR5 (LPDDR5), are high-performance memory chips paired with AI CPUs. Nvidia uses 480 GB of LPDDR5X per Grace CPU, and will more than triple that figure to 1.5 TB in its Vera CPU. 

Meanwhile, AMD uses standard DDR5 in its current generation EPYC Turin CPUs. The company plans to first support LPDDR5X in its next generation EPYC server CPU "Verano," which is expected to become available in 2027. 

DRAM Pricing Surges 

Related to this, SK Hynix noted all the way back in October 2025 that its conventional DRAM, NAND, and HBM capacity was all sold out for 2026. Micron and Samsung have not said their conventional DRAM capacity is sold out, but we can see based on price increases that the shortage is very significant. As commodity products, conventional DRAM sales take place at monthly/quarterly contract prices or spot prices, while memory suppliers are allocating HBM capacity through long-term contracts. 

In Q3 2025, DRAM contract prices soared by 171.8% YoY. Meanwhile, in Q1 2026, TrendForce estimates that conventional DRAM contract prices increased by 93%-98% QoQ and projects another 58%-63% QoQ increase during Q2 2026. 

One of the key factors contributing to conventional DRAM tightness is memory makers relocating capacity away from these products and toward higher-margin HBM. Importantly, this move from conventional DRAM to HBM does not translate into a 1:1 shift in bit supply. 

Why HBM Production Reduces DRAM Supply 

Current generation HBM3E requires approximately 3X the wafer capacity per GB compared to DDR5. This makes the strain on conventional DRAM exponentially worse, as reallocating wafer capacity toward HBM disproportionately reduces the wafer supply available for conventional DRAM production.  

Furthermore, Micron noted in May 2026 that this ratio will continue to grow as suppliers transition from HBM3E to future generations in HBM4 and HBM4E. Notably, Nvidia’s upcoming Rubin generation and AMD’s upcoming Instinct MI450 accelerators will use HBM4, while Google’s TPU v8 will use HBM3E. 

SSD Shortages Add to the AI Memory Crunch 

NAND flash SSDs, which store large amounts of data beyond what DRAM can store at a given time, are also facing a supply crunch. Kioxia has a joint venture with SanDisk in operating SSD fabs. Near the beginning of 2026, Shunsuke Nakato, Managing Director of Kioxia's Memory Business Unit, said that the company’s capacity was sold out for the year. Notably, 60% of the capacity from the joint venture goes to Kioxia. Furthermore, while interviewing SanDisk’s CEO, Bernstein analyst Mark Newman estimated that the company’s ASP per GB increased by 140% QoQ in Q1. 

Perhaps even more striking are statements made by Everpure (formerly known as Pure Storage) CEO Charles Giancarlo in an open letter to customers. Everpure, which makes NAND-based storage systems, says its “input costs of many high-volume semiconductor components have surged between 300 percent and 900 percent (4x to 10x) since mid-2025.” Additionally, HDD makers Seagate and Western Digital have said their capacity is sold out for 2026. 

How Long Will the Memory Shortage Last? 

Looking ahead, memory suppliers are all saying that shortages will continue but are providing differing statements about how long. 

SK Hynix Signals Prolonged AI Memory Shortage Into the Next Decade 

SK Hynix has outlined a particularly long path to normalization. At Computex in June, Chairman Chey Tae-won reiterated his stance that shortages would persist into 2030. This comes even as the firm plans to nearly double its monthly DRAM wafer capacity from 550,000 today to 1 million by 2030. By 2034, the company expects to triple DRAM capacity, a timeline that is 10 years ahead of its previous plan. 

Pursuant to its investment plans, SK Hynix is said to be in the final stages of listing its American Depository Receipts (ADRs) on the NASDAQ. The current expectation is that the offering will represent around 2.5% of the firm’s outstanding shares. This would imply a gross proceeds value near $26 billion based on recent prices and exchange rates. That would be very significant, potentially increasing its cash by 72% from $36.1 billion last quarter to around $62 billion.  

Micron, Samsung, and SanDisk Expect Tight Supply Through 2027+ 

Micron is also adding fabs, with initial wafers expected at its Idaho 1 facility in mid-2027, and with several others to follow. Related to this, Micron's VP of Marketing, Christopher Moore, said in a January interview, “you're not really gonna see real output, meaningful output by the time we get all the qualification done and customers are accepting it and you get the tools, everything up and running until 2028.”  

In a May interview with Bloomberg, Micron CEO Sanjay Mehrotra echoed this, saying, “we see that meaningful new supply in the industry doesn't really start ramping until 2028 timeframe.” In its latest earnings report, Micron added “Even as we expect industry supply to improve gradually in 2028, we currently do not have line of sight as to when memory supply will be able to catch up with increasing demand." 

Image showing Micron’s planned fab expansions. The first leading edge DRAM and HBM site in Idaho is not expected to come online until mid-2027, with the second not coming online until late 2028. The first two leading-edge DRAM sites in New York do not come online until 2030, while shipments from the leading-edge DRAM and HBM site in Japan are not expected until 2028. Source: TrendForce TrendForce 

Meanwhile, in its latest earnings call, Jaejune Kim, EVP of Samsung’s Memory Business, said, “And unlike previous years, customers who are concerned about supply shortages are actually bringing forward their demand for 2027 already. So currently, just based on prebooked demand alone, the supply-demand gap is looking to widen further in 2027 versus this year.” 

Specific to NAND, SanDisk’s CEO said at the JPMorgan Technology Conference, “We see this market undersupplied for a long period of time.” More specifically, he noted that in 2025, the company had a “clear point of view” that the market would become undersupplied through 2026. He added, “And I think we can say through the end of '27, we have that same level of conviction now.” 

Across these statements, we can see that executives from top memory companies are all indicating that shortages will continue until at least some part of 2028. Importantly, Samsung indicated that shortages would intensify in 2027, while Micron doesn’t expect meaningful output increases until 2028. Given this, it may not be so far-fetched to think supply and demand will not balance until near the end of the decade. 

HBM and DRAM Demand Surge as AI Models and Infrastructure Scale 

While the memory shortage is clearly in place today, it is important to understand the underlying factors within AI models and infrastructure driving this shortage and contributing to its continuation. 

Model Complexity and KV Cache Requirements on HBM 

Model complexity and the KV cache are two primary drivers of increased HBM demand, especially as it pertains to inference deployments. Increasingly complex models are being trained and deployed for multi-step inference or agentic tasks, requiring larger context windows.  

Context windows represent the amount of information that the model can remember at a given time to execute tasks, with window length increasing dramatically over time, such as for OpenAI’s models. For example, according to Artificial Analysis, OpenAI’s GPT-3.5 Turbo, released in 2023, had a context window of just 4k tokens. This increased to 128k tokens in GPT-4.5 Preview in early 2025, while OpenAI’s latest model, GPT-5.5 (xhigh), has a context window of 922k tokens, a 230X increase in the span of three years.  

Chart showing the context windows of three OpenAI models. GPT-3.5 Turbo’s context window is 4k tokens, increased to 128k tokens in GPT-4.5 Preview, while OpenAI’s latest model, GPT-5.5 (xhigh), has a context window of 922k tokens. Source: Artificial Analysis 

The reason this sharp increase in context windows is important for the memory thesis is because context windows define the potential size of a model's KV cache, or the actual working memory that a model continually references during inference. HBM is particularly important here as the goal is to keep as much of the KV cache on HBM — the fastest memory available — as possible.  

However, if HBM capacity is not large enough to hold the entire KV cache, that remaining portion can be offloaded to slower conventional DRAM or SSDs, introducing latency during inference or leaving expensive GPUs (or other accelerators) underutilized.  

We can roughly put in perspective potential memory requirements for frontier models, using OpenAI’s GPT-4 with an estimated 1.8T parameters and a 128K context window as a benchmark. At FP8 precision, storing the model weights would require 1.8TB of HBM capacity (at 1 byte per parameter), while a 25% activation buffer would add 450GB.  

On a single Nvidia GB200 NVL72, this would leave roughly 11.1TB of HBM capacity free for the KV cache. Assuming 120 layers and a hidden size of 16,384, KV cache requirements per token would be ~3.9MB at FP8, meaning one NVL72 could in theory support maximum tokens of ~2.85 million, or around 22 concurrent requests at the max 128K context window. 

This problem becomes further multiplied as inference demand grows, resulting in more requests from many concurrent users. Consider that OpenAI has dozens of production models available and over 900 million weekly active users, implying that at peak usage it could be handling tens of millions of concurrent requests, each consuming KV cache memory.    

HBM Content Soars Over GPU Generations 

Notably, Nvidia’s 8-GPU DGX H100 server contained 640 GB of HBM, or 80 GB per chip. The B200 moved to 1.44 TB of HBM in an 8-GPU configuration, resulting in 180 GB per chip, or 125% higher than the H100. The B300 contained 288 GB of HBM per chip, 60% more than the B200, and 3.6X more than the H100. 

Overall, the latest system available, the 72-GPU GB300, supports up to 21.7 TB of HBM. This results in rack scale deployments that contain nearly 34X more HBM content than the DGX H100 server. This shift came over approximately three years, with the H100 entering full production in September 2022, while the GB300 entered full production in August 2025. This rapid increase in HBM content over a short period is another significant contributor to HBM shortages. However, it is important to note that Rubin will remain at 288 GB per chip, but use HBM4 rather than the HBM3E in Blackwell to provide higher bandwidth. 

Chart showing the increase in HBM content per Nvidia GPU from H100 to Blackwell Ultra. HBM content starts at 80 GB in H100 and moves progressively higher to 141 GB in the H200 and 192 GB in Blackwell to 288 GB in Blackwell Ultra, equating to 3.6X increase across these GPU generations. Source: Nvidia Nvidia 

This shift is similarly evident at AMD. The company’s Instinct MI250 GPUs, launched in November 2021, supported 128 GB of HBM per chip. Meanwhile, the company’s MI450 series will deliver nearly 3.4X HBM capacity than MI250 at 432 GB per chip. The “Helios” MI450 72-GPU rack scale system delivers 31 TB of total HBM4 content—or 1.5X higher than the GB300 NLV72. Shipments are expected in the second half of 2026. 

Conventional DRAM Pressure Points: Vera Content Triples as CPU Demand Increases 

As noted, Nvidia’s Vera CPUs will support more than triple the LPDDR5X per chip of the Grace CPU, which is likely to put additional pressure on conventional DRAM demand. However, this comes down to more than just per-chip DRAM content. 

Nvidia has announced that it will launch a standalone Vera rack containing 256 CPUs—or 7X more than the 36 CPUs in the Vera Rubin NVL72. This is due to the increasing importance of agentic AI, which is expected to increase CPU demand significantly. With this, LPDDR5 demand could be further pressured from two sides: higher content per CPU and higher overall CPU sales. 

Notably, TrendForce cites the rising importance of CPUs as a rationale for increasing its 2026 DRAM market forecast to $618.7 billion, representing 303% YoY growth. The firm projects DRAM growth of another 46% YoY in 2027, and for the overall DRAM and NAND market to hit $1.28 trillion. This would equate to a 5.7X increase in just two years versus the $225 billion market in 2025. 

To learn more about the emerging CPU bottleneck, read I/O Fund’s June 2026 article: AMD, Nvidia, Arm, Intel: Inside the $120 Billion CPU Gold RushAMD, Nvidia, Arm, Intel: Inside the $120 Billion CPU Gold RushAMD, Nvidia, Arm, Intel: Inside the $120 Billion CPU Gold Rush 

Risks to the Memory Thesis: TurboQuant and Long-Term Agreements 

There are two major risks investors should watch as the memory trade matures, which is whether shortage-driven pricing surges will continue to flow disproportionately to memory suppliers, and whether software-based efficiency gains could reduce the intensity of memory usage.  

In recent earnings calls, memory suppliers have discussed hyperscalers and AI infrastructure customers seeking longer-duration contracts of up to five years. There are some positives to these agreements, which is they smooth out the cyclicality that many investors fear given these agreements guarantee any inventory will be quickly absorbed, resulting in more stability.  

On the flip side, memory stocks are surging precisely because pricing is uncapped right now, therefore this introduces a new era for many memory stocks to where they transition to becoming secularly certain, yet the epic volatility in both directions is more muted.  

Notably, there are various deal structures for these agreements, yet in most instances, they lock-in demand for many years in exchange for some kind of cap in memory component pricing.  

The second risk is model and system efficiency through improvements such as Google’s TurboQuant. TurboQuant is a compression method that directly addresses the KV cache bottleneck. Google says that TurboQuant can reduce KV cache memory size by 6X while simultaneously preserving model accuracy and accelerating speed by up to 8X. 

However, the increased usage of HBM in Google’s own systems is a telling point that pushes back on TurboQuant fears. The company’s latest TPUs, the 8t and 8i, support 216 GB and 288 GB of HBM per chip, respectively. These figures are 13% higher and 50% higher than the 192 GB of HBM capacity offered by Ironwood v7. Thus, Google, which developed TurboQuant, is itself substantially increasing HBM capacity in spite of the efficiency gains. 

Conclusion:

Micron has been one of our largest positions in 2026, and for good reason. Over the past 10 months, the company added more than $1 trillion to its market capitalization, which now sits near $1.35 trillion. That move reflects not only Micron’s execution, but also the market’s growing recognition that memory demand is entering a much larger cycle. 

Investing in memory has been anything but easy. Many market participants feared the cycle was topping at the start of 2026, but the I/O Fund’s disciplined process kept us in the position at a high allocation, frequently above 10%, for year-to-date returns of 277%. 

This analysis is a small sample of what we do behind our paywall. We are not simply stating memory is an important bottleneck, but rather we show, with data, how the shortage is pressuring every layer of the memory stack, including HBM for accelerators, LPDDR5X and DDR5 for CPUs, and NAND SSDs for storage and agentic inference.  

As Q2 wraps up, the I/O Fund is preparing to identify the next wave of AI winners in our upcoming Top 15 AI Stocks for Q3 2026 report. Previous reports identified Micron as a major beneficiary, along with names such as Bloom Energy, up over 1800% since our April 2025 entry, and lesser-known AI networking stocks up over 600% since our November 2025 entry. 

We publish more than 100 paywalled articles each year on AI stocks, hold weekly 1-hour webinars, and offer an actively managed portfolio with real-time trade alerts. 

Don’t miss the next wave of the AI trade
Join I/O Fund

Please note: The I/O Fund conducts research and draws conclusions for the company’s portfolio. We then share that information with our readers and offer real-time trade notifications. This is not a guarantee of a stock’s performance and it is not financial advice. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis. Beth Kindig and the I/O Fund own shares in MU at the time of writing and may own stocks pictured in the charts.   

Leo Miller, AI and Semiconductor Investment Writer at I/O Fund, contributed to this analysis. 

👉🏻 Share with a Fellow Investor
Heare with a Fellow Investor
Help someone else benefit from this insight.

Recommended Reading:

• Nvidia, CoreWeave, and Nebius: Inside the Circular Financing of the GPU Boom
• AMD, Nvidia, Arm, Intel: Inside the $120 Billion CPU Gold Rush
• Google TPU v8 vs Nvidia: How Inference Is Rewriting the AI Market
• The AI Networking Stock That Beat Nvidia by 7X YTD for Returns of 135% YTD 

• Neoclouds are seeing massive hyperscaler demand as companies race to scale AI infrastructure, resulting in rapid revenue and backlog growth. 
• Leaders like CoreWeave and Nebius enable this through access to the latest Nvidia GPU’s while also optimizing compute utilization.  
• However, the bearish argument behind hyperscaler demand lies in their desire to offload their capex spending and shift costs to the operating expense line. 
• CoreWeave’s and Nebius’ growth is far from profitable, as they seek to capture AI demand with limited cash flow and soaring debt loads in an increasingly tough macro backdrop.  
• Circular financing, demonstrated by Nvidia’s investments and financial backstopping, is another key item to monitor closely 

Neoclouds are one of the more hotly debated AI business models, with CoreWeave and Nebius being the two most widely recognized names. These companies have seen their sales, backlog, and share prices soar, differentiating themselves through quick access to the latest GPU compute and GPU utilization advantages that allow hyperscalers to rapidly add efficient compute capacity. 

Notably, CoreWeave and Nebius have each secured 3.5 GWs of contracted power capacity; while these power footprints are key considering power is a hindrance to data center expansion, the vast majority of their contracted power capacity has yet to come online. CoreWeave is targeting 1.7 GW of active power by the end of 2026, while Nebius is targeting 800 MW to 1 GW of connected power. 

In turn, they are quickly working to convert their contracted power to active power, and thus convert large backlogs into revenue. Yet doing so is extremely expensive, and neoclouds do not have the same cash nor operating cash flow profiles of Big Tech. This is leading neoclouds to employ unique and circular financing structures, raising some red flags. 

In this analysis, I dive into the two public neoclouds that are riding Nvidia equity, hyperscaler contracts, and GPU-backed debt to fund the buildout, and what it means for the durability of the surge. 

Microsoft and Meta’s $120B+ Bet on Neoclouds 

The size of hyperscaler-neocloud partnerships compared to their current revenue is astounding. Microsoft has struck the most neocloud deals, with approximately $60 billion worth of commitments between CoreWeave, Nebius, and other private players such as Nscale. Meanwhile, Meta has committed $35.2 billion to CoreWeave in total after its recent $21 billion expansion, and an up to $27 billion deal with Nebius for a total commitment of up to $62.2 billion. Along with Meta, OpenAI is one of CoreWeave’s two largest customers, while CoreWeave also has a multi-year compute agreement with Anthropic.  

Alone, Microsoft and Meta’s total commitments extend up to $122.2 billion – for perspective, that is ~90% of the TTM revenue of AWS being allocated towards neoclouds over long-term capacity deals. When factoring in hyperscaler-backed deals from OpenAI and Anthropic (although exact deal value is unknown), total potential commitments surpass $145 billion.  

Keep in mind, CoreWeave’s FY2026 estimated revenue is $12.6B and Nebius FY26 revenue is expected to be $3.4B – therefore, these partnerships are leading to commitments that are an order of magnitude higher than current sales.  

The reason hyperscalers are willing to allocate this capital to a relatively new business model in the neoclouds is three-fold – quick access to leading GPU generations, optimized compute utilization, and the added benefit of not having to recognize capex on the balance sheet – we look at each of these drivers below. 

Neocloud Advantage is Offering Quick Access to GPUs 

At its root, neocloud demand is a product of hyperscalers' insatiable demand for compute capacity. However, neoclouds can often add compute capacity much faster than hyperscalers can through internal builds, offering a key value proposition for Big Tech. As hyperscalers spend hundreds of billions a year on AI compute, minimizing the lag between data center expenses and revenue generation is critical to maximizing their return on investment. 

Supporting the argument around neocloud’s advantage lying within time to deployment, commercial real estate giant JLL notes, “Neoclouds can deploy high-density GPU infrastructure within months compared to multi-year builds for hyperscale data centers, providing crucial time-to-market advantages for businesses needing rapid AI development.”

mid

In CoreWeave’s S-1 Registration filing, it lists “Faster access to the latest AI infrastructure advancements” as one of its key benefits to customers. Specifically, CoreWeave says “we were among the first to deliver NVIDIA H100, H200, and GH200 clusters into production at AI scale, and the first cloud provider to make NVIDIA GB200 NVL72-based instances generally available. We are able to deploy the newest chips in our infrastructure and provide the compute capacity to customers in as little as two weeks from receipt.”  

Nebius makes a similar statement in its Annual Report, noting its “consistent track record of being one of the first to deploy the latest generation of NVIDIA GPU chips.” 

CoreWeave and Nebius' relationship with Nvidia is key to acquiring the latest GPUs ahead of others. Nvidia recently invested $2 billion in both CoreWeave and Nebius. Under these partnerships, CoreWeave and Nebius will each look to deploy more than 5 GW of data center capacity by 2030. 

CoreWeave recently demonstrated its ability to offer quick access to the latest chips and newest architectures to hit the market once again, being the first to have a Vera Rubin system up and running at the start of June.  This provides evidence that partnering with CoreWeave and Nebius can help hyperscalers access as much of the latest GPU compute as possible in short order. 

Beyond Hardware: Neocloud Platforms Offering Higher GPU Utilization  

Aside from raw compute access, CoreWeave and other neoclouds layer on software and additional capabilities that improve GPU utilization – a key value add for hyperscalers.  

For example, CoreWeave Kubernetes Service (CKS) helps coordinate the allocation of workloads across thousands of GPUs, while its SUNK service helps optimize GPU utilization by allowing training and inference workloads to run on the same cluster. CoreWeave Tensorizer enables high-speed model loading, reducing GPU idle time. 

Combining these software and optimization capabilities with rapid fault detection and remediation services, CoreWeave believes it can offer higher GPU utilization rates than hyperscalers, based on the model FLOPs utilization (MFU) metric. The “MFU gap” is a metric that describes the gap between compute capacity and usage, which today often ranges between 30% to 40%. 

The MFU gap can become quite costly as it represents a more realistic way to measure the performance of GPUs — rather than only taking into account if a GPU is sitting idle or not. According to Trainy AI: “GPU Utilization is only measuring whether a kernel is executing at a given time. It has no indication of whether your kernel is using all cores available, or parallelizing the workload to the GPU’s maximum capability.”  

Chart comparing theoretical model FLOPS utilization (100%) with observed performance (35%–45%), illustrating a significant efficiency gap in AI workloads. Source: CoreWeave CoreWeave 

When going public, CoreWeave published its MFU rate at 35% to 45%, stating it is 20% higher than competitors, which means other AI data centers had MFU rates more in the 30% range. However, in a March 2025 blog post, CoreWeave noted that it was achieving an MFU of >50% on Hopper GPUs. This ability to stand up next-generation GPU hardware in short fashion combined with improved utilization rates is where the neoclouds’ advantage lies.  

Behind the Balance Sheet: Why Hyperscalers Are Leasing Neocloud Capacity 

By leasing compute capacity from neoclouds, hyperscalers shift their cost timeline from being a large upfront capex outflow to an operational expense outflow spread over long-term contracts. The need to spread costs is becoming increasingly evident due to the massive spending hyperscalers are engaged in.  

Although this is the “bear” case on why hyperscalers work with neoclouds—contrasting this with the rationale behind GPU access and utilization is key because one could argue that hyperscalers are quite capable of software optimizations and GPU utilization on their own (in fact, they are the longstanding incumbent here with deep expertise in cloud operations and workload optimizations). 

Take Meta for example. Analysts are currently expecting the company to generate $136 billion in cash from operations in 2026. With its stated capex guidance of $125 billion to $145 billion, the company could easily be free cash flow negative during the year. However, as noted, Meta also has up to $62.2 billion in neocloud agreements. If Meta built the equivalent value of capacity itself, the firm would recognize that spending as balance sheet capex, weighing further on its already pressured free cash flow.  

On the other hand, neocloud agreements add nothing to Meta’s capex, as the costs are recognized as operating expenses over the life of the contracts. Notably, Meta’s contracts with CoreWeave and Nebius extend through 2031-2032, meaning that opex payments could average less than $10 billion annually. 

Looking at Microsoft, we can see a similar situation. In calendar year 2026, the company is guiding for capex of $190 billion, while analyst forecast $200 billion in cash from operations over the same period. If these figures materialize, the company would consume 95% of its OCF on capex. The $60 billion in neocloud agreements, recognized as operating expenses over many years, expands its capacity while keeping that spend off its cash flow statement. 

As hyperscalers offload their capex, neoclouds are the ones taking that capex on—resulting in their massive funding needs.  

Circular Financing: Nvidia’s Role as an Investor, Supplier, and Demand Backstop 

Both Nebius and CoreWeave lend some of their advantage to Nvidia, as it is this partnership with the GPU leader that offers them that ability to be among the first providers to stand up and deploy next-gen platforms such as Blackwell Ultra and now Rubin.  

Having Nvidia as a partner also could play a role in helping CoreWeave and Nebius secure funding at much better terms, extending presence and support beyond the hyperscalers to another investment-grade firm with a strong balance sheet and cash flows. Nvidia’s LTM free cash flow was $119 billion, the second highest of any company in the world, only behind Apple. The downside, however, is that Nvidia’s relationship with the two is one of the most identifiable instances of circular financing.  

This stems from the multi-billion-dollar investments that Nvidia has made in CoreWeave and Nebius. Notably, Nvidia’s latest $2 billion investments in each company were not its first. Nvidia’s Q1 2025 13F filing revealed a CoreWeave stake worth $896.7 million at the time, while its Q4 2025 13F revealed a $33 million stake in Nebius. Thus, the investment relationship between Nvidia and these firms extends well beyond one year. 

Furthermore, in the case of CoreWeave, Nvidia has also provided a significant financial backstop against unsold GPU capacity. Under the agreement with an initial value of $6.3 billion, “in instances where [CoreWeave’s] datacenter capacity is not fully utilized by its own customers, NVIDIA is obligated to purchase the residual unsold capacity through April 13, 2032.” In other words, Nvidia is committed to purchasing unsold GPU capacity if CoreWeave is unable to find another buyer. With an initial value of $6.3 billion, there is the potential that the arrangement could become larger over time. 

As Nvidia makes these investments, CoreWeave and Nebius are going right back to Nvidia to purchase large volumes of GPUs – a clear representation of circular financing. By providing a relatively small amount of equity funding, Nvidia secures relationships with these neoclouds that intend to purchase tens of billions' worth of GPUs.  

Nvidia could see long-term benefits by supporting CoreWeave and Nebius through their ramp-up phases where cash flow is deeply negative. If the firms can eventually become self-sustainable, Nvidia would have two large-scale customers that it can continue selling its latest systems to for years to come. However, for the neoclouds, the concern is whether they have to continually raise cash into the foreseeable future to build new infrastructure and when that would level out, as revenue lags capex 2:1. 

How Neoclouds Are Funding AI Expansion: Debt, Equity, and Circular Financing 

Both CoreWeave and Nebius are eyeing rapid ramps in active power – CoreWeave currently has 1GW of its 3.5GW contracted power pipeline active, but it aims to convert the majority of that over to active capacity by the end of 2027, while Nebius similarly has 3.5GW of contracted power and a goal of reaching up to 1GW of connected (active or can be activated upon GPU installation) by the end of 2026.  

However, as with all AI buildouts right now, the keywords are “active power” as energy constraints are intensifying across the board.  

CoreWeave’s Balance Sheet Challenged, Debt Quickly Rising 

CoreWeave’s balance sheet is in a difficult position, as the company looks to rapidly expand its active power footprint at a rate that is not supported by its cash balance and its operating cash flow. 

Revenue of $2.08 billion rose by 112% YoY in its latest quarter. However, operating cash flows (OCF) came in at $2.98 billion, compared to capex of $7.7 billion, leading to free cash flow of -$4.71 billion. This mismatch led to the firm’s cash balance falling by $890 million, or 28.3% QoQ to $2.27 billion. Meanwhile, debt increased by nearly $3.5 billion, or 16.1% QoQ to $24.86 billion – this is set to rise further in Q2 as CoreWeave just announced a $3.5 billion senior note raise on June 11. 

Chart showing CoreWeave’s quarterly capex rising sharply to approximately $7.7 billion, while revenue reached around $2.07 billion over the same period. Source: YChartsYCharts

For the full-year, CoreWeave expects to spend $31 billion to $35 billion on capex, or $33 billion at the midpoint. This implies capex spending for the remainder of the year of $25.3 billion. Analysts currently estimate that the company will generate $8.68 billion in operating cash flow in 2026, or just $5.7 billion for the rest of the year. Given CoreWeave’s $2.27 billion cash balance, this creates a huge funding gap of $17.33 billion. In practice, CoreWeave is likely to raise more than this to avoid further decreasing its already somewhat thin cash cushion. 

CoreWeave has used equity issuance in the past as a funding source, but debt issuance far outweighs this. Looking at its first five earnings reports since going public, its total equity issuance is only $3.5 billion, while debt issuance was more than 5X higher at $18.81 billion. Thus, a further increase in debt is likely to be the primary way that CoreWeave continues to fund its capex plans while already having a net cash position of -$22.6 billion. Looking into its unique funding structures shows that debt will continue to be a key lever that the firm pulls. 

Nebius: Stronger Balance Sheet but Ongoing Funding Needs 

Nebius is comparatively in a much better position, with $9.37 billion in cash to $8.45 billion in debt, for a net cash balance of $920 million. Revenue rose 684% YoY to $339 million in its latest quarter, while operating cash flow was $2.26 billion, rising by 170.7% QoQ due to significant customer prepayments. Capex came in at $2.47 billion, resulting in FCF of -$214.9 million.  

However, Nebius is also looking to rapidly expand its active power footprint, with the firm’s midpoint capex guidance for the full year at $22.5 billion. This implies $20 billion in spending over the remainder of the year. Including the company’s cash and contractual commitments of approximately $6.9 billion, Nebius currently needs to draw $6.3 billion in additional funding to support the midpoint of its capex forecast. 

Like CoreWeave, Nebius has also leaned heavily on debt rather than equity issuance to fund itself, although to a lesser extent. Since Q4 2024, Nebius’ total equity issuance was approximately $3.92 billion when including the $2 billion in pre-funded warrants Nvidia recently purchased. Over the same period, its debt issuance was $8.32 billion. In its latest earnings call, Nebius noted asset backed financing, corporate debt, and equity issuance as options for raising capital.  

Notably, Nebius’ undeployed 25 million share at-the-market equity program could go a long way toward bridging its 2026 funding gap. At a $200 share price (around 10% below the stock’s current level), fully utilizing this program would generate gross proceeds of $5 billion while diluting shareholders by approximately 8%. However, given past trends, asset backed and corporate debt are likely to be the primary path forward. 

Overall, this breakdown of CoreWeave and Nebius’ funding requirements for 2026 is just one stage of a much larger push to convert its contracted power into active power. After all this spending, CoreWeave aims to have just under 50% (1.7 GW) of its contracted power active. Meanwhile, Nebius hitting the upper bound of its connected power target would account for less than 30% of its contracted power, which includes power that is either active or can be activated once GPUs are installed.  

In turn, the companies will continue to need to find more and more funding to scale until CFO converges with capex. With the spread between these figures still very wide, the likely result is further increases in debt loads and/or shareholder dilution over several years. 

GPU-Backed Debt: Inside CoreWeave’s Funding Engine for AI Infrastructure 

CoreWeave relies heavily on GPU-backed delayed draw term loans (DDTLs), having closed six separate facilities. Under DDTLs, CoreWeave draws down funds intermittently as it uses them to pay for different stages of data center buildouts.  

Notably, the company’s $8.5 billion DDTL 4.0, closed in March, was the first of its kind to receive an investment-grade credit rating. As of Q1 2026, CoreWeave had only drawn $1.26 billion worth of DDTL 4.0. This is the only portion of the $8.5 billion that currently shows up in CoreWeave’s total debt. Thus, as the firm draws down more of DDTL 4.0 over time, its debt will also increase.

Table showing CoreWeave’s debt structure with total debt of approximately $25.1 billion, including multiple delayed draw term loan (DDTL) facilities and senior notes. Notably, the DDTL 4.0 facility totals $8.5 billion, but only $1.26 billion has been drawn, indicating significant future debt expansion as capital is deployed. Source: CoreWeaveCoreWeave 

CoreWeave notes that the investment-grade rating is “supported by a long-term customer contract with an investment-grade AI enterprise," which is presumably tied to Meta’s latest contract. Essentially, the contract that CoreWeave has signed with the investment-grade customer, as well as the value of the GPUs it buys, are collateral for the debt. This is why the facility can achieve an investment-grade credit rating despite CoreWeave itself having a poor balance sheet, allowing for much more favorable interest rates that CoreWeave could not otherwise receive. 

Still, CoreWeave's ability to receive better interest rates than peers relies on backing from investment grade customer contracts. Notably, DDTL 5.0, closed in May (and is thus not included in the table above), was backed by two non-investment-grade customer contracts. This resulted in the facility not receiving an investment grade rating and thus having a higher interest rate. 

Interest Rate Pressure: A Growing Risk to Profitability 

Increases in general rates apply further upward pressure on the rates that CoreWeave and other neoclouds can receive in future funding rounds. The fixed rate tranche of DDTL 4.0 is tied to U.S. Treasuries with an average weighted maturity of 3.14 years, plus a 2% premium. This portion of the yield curve has seen rates rise significantly since the beginning of the year from less than 3.6% to nearly 4.2%.

Chart showing the 3-year U.S. Treasury rate rising from below 3.6% in early 2026 to approximately 4.16% by June, reflecting a sharp increase in short- to mid-term interest rates. Source: YCharts.YCharts.

Notably, CoreWeave’s interest payments are already elevated, coming in at $536 million in Q1. This equates to 25.8% of its $2.08 billion in revenue, and 46.3% of its $1.157 billion in adjusted EBITDA. The company is guiding for midpoint revenue of $2.525 billion next quarter, and midpoint interest expense of $690 million—which would push its interest to revenue ratio up to 27.3%. With this, interest expense is expected to become an even more relevant line item while already putting significant pressure on profitability. 

The Neocloud Race: Balancing Surging AI Demand With Rising Debt and Circular Risk 

Overall, neoclouds clearly have significant growth momentum, with revenues and backlogs spiking, while attracting interest from investment-grade hyperscalers such as Microsoft and Meta, and AI labs including OpenAI and Anthropic. Access to leading Nvidia systems, and GPU utilization advantages make neoclouds an option for hyperscalers looking to quickly scale AI compute capacity. 

At the same time, the mismatch between operating cash flow and capex is causing debt levels to rise rapidly, which is a dynamic that is unlikely to change in the near term. Elevated interest rates remain an external risk, while circular financing raises questions around the degree to which neocloud growth depends on Nvidia’s capital support, and the extent to which Nvidia’s GPU demand is increasingly tied to the neocloud model. 

As Q2 wraps up, I/O Fund is preparing to identify the next wave of AI winners in our upcoming Top 15 AI Stocks for Q3 2026 report, with coverage across AI networking, memory, energy, custom silicon, and the infrastructure bottlenecks driving the next leg of the trade. 

Premium Members will also receive upcoming thematic reports on the latest shifts in AI networking and a new catalyst we believe could become one of the more important opportunities in the second half of the year. 

We publish more than 100 paywalled articles each year on AI stocks, supported by an actively managed portfolio and real-time trade alerts.

Don’t miss out on the AI trade 👉 Join I/O Fund Premium today.oin I/O Fund Premium today.

Please note: The I/O Fund conducts research and draws conclusions for the company’s portfolio. We then share that information with our readers and offer real-time trade notifications. This is not a guarantee of a stock’s performance and it is not financial advice. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis. Beth Kindig and the I/O Fund own shares in NVDA at the time of writing and may own stocks pictured in the charts.  

Leo Miller, AI and Semiconductor Investment Writer at I/O Fund, contributed to this analysis. Leo Miller owns shares of NVDA and META.

👉🏻 Share with a Fellow Investor
Heare with a Fellow Investor
Help someone else benefit from this insight.

Recommended Reading:

• AMD, Nvidia, Arm, Intel: Inside the $120 Billion CPU Gold Rush
• Google TPU v8 vs Nvidia: How Inference Is Rewriting the AI Market
• The AI Networking Stock That Beat Nvidia by 7X YTD for Returns of 135% YTD
• Bloom Energy — Our 2026 Top Pick Was the Best Performing Stock in April

Three months ago, GPUs were all the rage, and the idea that CPUs could challenge GPUs when it comes to AI budgets was unfathomable. The shift in this perception is evident not only in management commentary, but also in CPU design companies and OEMs raising forecasts that are now 2X+ higher, as many of the largest players have stated they did not foresee the magnitude of the surge in CPU demand from agentic AI. 

In just six months, AMD has issued a massive increase to its server CPU market forecast, nearly doubling its expected CAGR to 35%—estimating that the market will eclipse $120 billion by 2030. Arm made a similar announcement in March, projecting that the total addressable market (TAM) for data center CPUs will grow to over $100 billion by its fiscal year 2031 (roughly calendar year 2030). This would represent a more than 4X increase over its current TAM estimate of $24 billion, equating to a 33% CAGR. 

An important shift is driving these forecasts as the AI market transitions away from chatbots, which saw a CPU-to-GPU ratio that was heavily weighted toward GPUs from 2023-2025. As we move into agentic AI, an Intel and Georgia Tech paper has stated that “tool-dominated agentic AI workloads are significantly bottle-necked" with CPUs consuming up to 88% of the end-to-end latency. The paper further concludes that “with better quality GPUs, the bottleneck can swiftly shift more towards CPUs.” 

What Intel and Georgia Tech are referring to, is that to scale agentic AI efficiently, CPU orchestration capacity will need to catch up to GPU reasoning capacity to minimize latency and prevent GPU underutilization. The answer to this problem is increasing the CPU-to-GPU ratio in AI clusters to keep token costs down.  

Below, I break down why CPUs are positioned to take a larger share of AI cluster bill of materials (BOM) and the explosion in demand we are already seeing. I examine server CPU forecasts that indicate this market will continue to grow rapidly over the coming years. Lastly, I look at the competitive dynamics and key players in this space, and how Nvidia is playing both sides of the CPU-GPU equation, and what front runners Intel and AMD are doing to maintain their lead.  

Ultimately, CPUs have gone from an afterthought to becoming the AI trade’s next great bottleneck – and with AMD, Nvidia, Arm and Intel circling a market that is doubling nearly overnight, the only question left is which company walks away with the lion’s share.

Why Agentic AI Is Driving a Massive Shift to CPUs 

Agentic workloads are structurally different from non-agentic workloads like chatbot queries, which is what has dominated the AI trade up to this point. Chatbots respond to simple requests and provide an output, moving at the pace of the human on the other side. Agents are far more complex, handling hundreds of concurrent tasks autonomously and reasoning through a problem to reach a conclusion, often with limited direction from humans. 

The Intel and the Georgia Tech paper highlights why CPUs are becoming increasingly important as agentic AI proliferates. Researchers noted that while CPU-GPU systems are needed to serve the diverse responsibilities of agents, the “majority of the external tools responsible for agentic capability either run on or are orchestrated by the CPU.” This is not the case in non-agentic workloads, where GPUs are the workhorses that CPUs feed data to. 

Why CPUs Handle Orchestration in AI Workloads 

The key bottleneck this creates on AI infrastructure is orchestration—or the need to call tools, direct API requests, and coordinate tasks between dozens of independent agents. Orchestration is where CPUs thrive. GPUs continue to handle inference reasoning, but CPUs tell GPUs where, when, and how to allocate their resources. 

As AI progresses over the next few years, inference demand is expected to explode—largely driven by agentic AI. Goldman Sachs estimates that by 2030, agentic AI will drive a 24X increase in total token consumption versus today to 120 quadrillion tokens per month. Its forecast shows agentic workloads accounting for over 80% of token consumption in 2030—dramatically higher than their share today.

mid

TrendForce notes that today, the CPU-to-GPU ratio in AI data centers sits between 1:4 and 1:8. For agentic AI applications, TrendForce sees the CPU-to-GPU ratio moving “to between 1:1 and 1:2, significantly boosting market demand for CPUs.”

Other forecasts, like those from Arm, rely on the CPU core count per GW metric. This measures the number of CPU cores per unit of data center power, regardless of the discrete number of CPUs. It is the more accurate way to measure the shift in CPU demand as chip density is increasing, with upcoming generations featuring higher core counts per chip. 

Notably, Arm CEO Rene Haas sees agentic AI driving CPU core demand as much as 4X higher to 120 million cores per GW, compared to around 30 million cores per GW today. Aside from the raw increase in core demand, packing more cores into each chip is a margin expansion opportunity for CPU designers. 

CPU Shortages: Supply Constraints and Pricing Power 

We are already seeing the CPU bottleneck start to play out through worsening CPU server shortages. Reuters reported in February that Intel has a substantial backlog of unfulfilled CPU orders, and that delivery times stretch as long as six months. It also noted delivery times for some AMD products of between eight and ten weeks. KeyBanc issued upgrades on Intel and AMD in January, noting that both firms were nearly sold out of CPU servers for 2026. At the time, KeyBanc noted ASP increases of 10% to 15%. 

Intel and AMD Backlogs and Lead Times 

It appears that the situation has become even more dire since, based on several reports from late May. Reuters now says that TikTok parent company ByteDance is working to accelerate its in-house CPU efforts, as Intel and AMD have raised prices by between 10% and 35% QoQ. ByteDance’s move suggests that it sees a prolonged CPU shortage, leading it to lean into this early-stage initiative. This adds weight to the structural increase in CPU demand implied by AMD’s forecast and shows the pricing power that CPU vendors are exerting.

Electronic equipment distributor Fusion Worldwide says that Intel distributors are only fulfilling around 40% of their yearly backlog allocations. It highlights lead times of 8 to 22 weeks domestically, with Asian customers waiting as long as 8 months. Overall, the firm estimates that Intel is under-shipping real demand by 20% “at best." It notes that AMD’s EPYC CPUs are effectively sold out in 2026, with delivery windows stretching more than 30 weeks. 

The Elec, a South Korean electronics industry trade publication, notes won-denominated price increases as high as 3X for some x86 (Intel and AMD) CPUs. This comes as Intel and AMD prioritize supply for U.S. hyperscalers—leaving little capacity for other customers. The Elec also said that the expected timeline for mass production of Intel’s next-gen Xeon 7 “Diamond Rapids” CPU has been delayed, moving from the second half of 2026 to the middle of 2027.

This data points to a shortage that is intensifying, putting pricing power into the hands of CPU vendors as they seek the highest-margin opportunities. 

AMD Sees Record CPU Server Sales, TAM Estimate Doubles to $120B 

The cause of these shortages is the rapid growth in server CPU demand seen at top players like AMD, and expectations that this market will grow much faster than it traditionally has over the coming years. AMD released its Q1 2026 results in early May, posting its fourth consecutive quarter of record server CPU revenue. Sales rose more than 50% YOY, with both cloud and enterprise end markets up over 50%. AMD expects growth to accelerate significantly in Q2, projecting server CPU revenue growth above 70% YOY, “with robust growth continuing through the second half of 2026 and into 2027.” 

Citing this acceleration in demand and the structural increase on CPU compute requirements that agentic AI is putting on data center infrastructure, AMD has doubled its server CPU TAM estimate. Per CEO Lisa Su, the company anticipates that this will be an over $120 billion by 2030—growing by a 35% CAGR. In November, AMD’s server CPU growth TAM CAGR forecast was just 18%. AMD’s decision to double its market growth forecast and add $60 billion to its TAM in just seven months demonstrates how rapidly current demand signals are translating to long-term confidence among industry leaders. 

Beth Kindig of the I/O Fund discussed in 2024 why AMD would be a winning AI stock and surpass Nvidia's returns over a 3-year time frame. Since then, Nvidia returned 80% and AMD has returned 220%

Thinking about margins going forward, AMD noted at the Bank of America 2026 Global Technology Conference that two-thirds of its server CPU growth in Q1 and expected growth in Q2 are coming from unit increases. Thus, units rather than ASPs are the primary growth driver. Given the worsening supply and demand gap, it’s possible that ASPs could drive an increased share of growth—providing a further lever for margin expansion. 

Server CPU Market Growth Forecasts Surging TAM 

Notably, server CPU TAM forecasts among several Wall Street banks line up with AMD’s forecast. For reference, AMD’s forecast implies a 2025 TAM of just under $27 billion. 

UBS projects that the market will grow from $31 billion in 2025 to $170 billion in 2030, or a 40.6% CAGR. It sees AI CPUs driving the vast majority of this growth, with the TAM increasing from $7 billion to $125 billion, or an 88% CAGR. Their forecast also includes a 56% increase in AI CPU ASPs over this period—implying a significant margin expansion opportunity. 

CPU TAM Revised Higher by Analysts 

Bank of America forecasts a TAM expansion from $43 billion in 2026 to $125 billion in 2030, or a CAGR of 30.6%, recently raising its 2030 estimate from $110 billion. While BofA’s growth rate is lower than AMD’s, this is likely because it accounts for the particularly high growth rates already being seen in 2026. 

Citi breaks down its forecast into three buckets: general purpose CPUs, AI head nodes, and agentic CPUs. It sees the overall market growing from $29.3 billion in 2025 to $132 billion in 2030, or a 35% CAGR. Within this, general purpose CPUs grow by a 20% CAGR to $50.9 billion, and AI head nodes grow by a 21% CAGR to $21.1 billion. Citi estimates that agentic CPU growth will drastically outpace the rest of the market, hitting $59.4 billion in 2030 for a massive 185% CAGR. Overall, the estimates from these three banks circle around the 35% CAGR that AMD outlined. 

Why Growth Rates Are Unprecedented for CPUs 

These very high CAGR forecasts highlight why server CPU shortages are escalating. This market has historically experienced single-digit annual growth rates. Thus, the supply chain was not necessarily prepared for a scenario where customers suddenly look to procure CPUs at a drastically higher pace, and long-term expected growth rates soar in a matter of months. 

AMD’s Goal: 50% Server CPU Market Share 

As AMD looks to increase its share of the CPU market to over 50% by 2030, it is targeting all three of the CPU categories Citi described. This will come through its Venice family of EPYC CPUs, including Verano, its first EPYC CPU purpose-built for AI infrastructure. AMD has begun to ramp production of Venice, while it plans to launch Verano in 2027. 

With this, AMD clearly expects CPUs to be a core growth driver over the coming years. If AMD achieves a 50% market share in the server CPU market, it would imply $60 billion in annual revenue. With server CPUs representing around half of data center revenue, this side of AMD’s business generated approximately $2.9 billion in revenue last quarter, or nearly a $12 billion run rate. Thus, hitting its $60 billion target would require a 5X increase in server CPU sales by 2030—an ambitious goal.

AMD vs Intel: x86 Market Share Dynamics 

There are two ways to think about market share in server CPUs. Mercury Research is one of the key authorities that estimates share in this space, with their estimates often centered around the x86 market. 

AMD is already very much in range of a 50% market share within x86. At its Investor Day, AMD noted that based on metrics from Mercury Research, its share of the server CPU market was around 40%. This lines up with Mercury’s estimate of AMD x86 market share of 41% at the time. Since then, AMD has gained considerable ground on Intel. Mercury Research estimates that AMD controlled 46.2% of x86 server CPU revenue share in Q1 2026 to Intel’s 53.8%. At this pace, AMD is well on its way to achieving a 50% market share in x86.

At Investor Day 2025, Lisa Su said AMD has a clear path to capturing more than 50% of server revenue market share, up from 40% today, alongside a 50% data-center CAGR and a goal of 40% PC revenue share.

AMD, Intel and Arm Market Share Dynamics 

Arm estimates that in terms of chip value, it held 20% of the cloud compute market share at the end of its fiscal year 2025, which ended in March 2025. Considering Mercury’s estimates on x86, or the 80% of the market that is not Arm-based, these figures imply overall market shares of 43% for Intel, 37% for AMD, and 20% for Arm. However, note the figures from Arm are stale. 

Thus, AMD would need to increase its market share by around 2.6% annually through 2030 to achieve its 50% goal. At least in the x86 market, AMD has cleared a much higher bar over the past several years. In Q2 2023, the company’s server CPU revenue share was just 25.1%, meaning that AMD increased its share of the x86 market by more than 7% annually through Q1 2026.  

While this historical pace is encouraging, it shows AMD’s progress only against Intel—not including Arm’s traditional IP business, nor its move to become a CPU designer through its Arm AGI CPU. Additionally, Nvidia is pushing more aggressively into the CPU market, largely through its standalone Vera racks. However, the Diamond Rapids delay is one factor that could give AMD a leg up in continuing to take share from Intel. 

AMD has its hands full as some of the world’s strongest IP and chip-design companies are targeting the same TAM – including the incumbent Intel, mobile-IP superstar Arm, and the newest entrant, Nvidia. 

Nvidia’s CPU Strategy: Expanding Beyond GPUs 

Within AI-specific infrastructure, CPUs have been traditionally deployed as AI head nodes paired with GPUs in the same rack. A critical development to track is the emergence of standalone CPU racks as this marks a significant shift in architecture.  

Customers will increasingly be able to deploy full CPU racks independently without automatically having to increase GPU counts—one of the key arguments for why CPUs can increase their BOM share in AI clusters. 

Nvidia Vera Standalone CPU Rack Overview 

For example, Nvidia’s Vera rack marks the first time that it will market a standalone CPU rack; a clear signal of the current opportunity in this space. With 256 CPUs in the standalone Vera rack, customers can deploy nearly 7X more CPUs in one rack compared to the 36 CPUs in the Vera Rubin NVL72, which also contains 72 GPUs. Total CPU cores sit at 22,528 for the Vera rack versus 3,168 for the Vera Rubin NVL72. Note that Vera is based on Arm architecture, rather than x86 architecture. 

Arm specifically mentioned the standalone Vera rack as a reason why its 4X CPU core count growth estimate is likely conservative. Arm CEO Rene Haas said, “we probably have undercalled the CPU demand in terms of the transition here. We talked about a 4x increase. We could get our heads around a bigger number than that.”  

Haas went on to say that the number of CPU cores “probably will” exceed the number of GPU cores, even though the number of CPU chips may not exceed the number of GPU chips. 

Nvidia Quickly Eclipses AMD with $20 Billion in CPU Revenue in 2026 

Importantly, Nvidia said on its latest earnings call that the Vera rack opens up a $200 billion CPU TAM for the company—dramatically larger than AMD’s +$120 billion estimate. Within this, Nvidia says it has visibility into generating nearly $20 billion in CPU revenue this year, primarily for standalone Vera racks. Meanwhile, about 50% of AMD’s data center revenue comes from server CPUs, putting this at $2.9B or about a $12B run rate. I expect that to change but allows for a baseline comparison, which is that Nvidia is entering the market aggressively. 

When asked whether CPUs are cannibalistic to GPUs, Nvidia CEO Jensen Huang did not offer a direct yes or no, but framed CPUs as additive to GPUs. He argued that more AI agents require more orchestration—increasing CPU demand—but that more agents also require more inference—increasing GPU demand. This lines up with AMD CEO Lisa Su’s statements that CPUs are largely additive/incremental to their overall TAM.  

Additionally, the standalone Vera rack is just one of four ways that Nvidia targets the CPU market—and is the only one that could substantially change its current CPU-to-GPU ratio. Its other markets include selling head node CPUs paired with Rubin GPUs at a 1:2 ratio in the Vera Rubin NVL72. Nvidia also sells Vera alongside its ConnectX-9 SuperNICs for both storage and confidential computing use cases. 

Still, with the standalone Vera rack, Nvidia is indicating that it expects the CPU-to-GPU ratio to shift—creating a need for the product. Ultimately, while the mix of AI BOM should move toward CPUs, Nvidia is still positioned to capture growth from both chip types. It can benefit from the increasing size of the overall pie, rather than CPU spending going up at the expense of GPU spending. 

Arm Makes Historic Move into Merchant Standalone CPU Racks 

Arm is also forwarding the CPU rack approach through its AGI CPU. Leveraging Arm’s history of delivering high performance with low power requirements for mobile devices, the new AGI CPU is designed to offer a similar balance between high performance and low power consumption. 

The AGI CPU was co-developed with key partner Meta, the chip’s first customer, who revealed they turned to Arm almost two-and-a-half years ago to see if there was a CPU option that fit Meta’s needs: “put in a lot more cores per watt, but we do not want to compromise on the performance piece.” Meta had only been finding options satisfying one of the two criteria: meeting the performance but with too much power, or meeting the power but with too little performance. 

Arm CPUs Deliver Higher Performance Per Watt vs x86 

One of the main advantages that Arm touts is higher performance per watt. Based on internal estimates, the firm says the Arm AGI CPU can provide up to 2x greater performance per watt vs. Intel and AMD’s x86.  

Higher performance per watt is a key value proposition for hyperscalers, allowing more power to be dedicated to compute or networking equipment. 

Bar chart comparing the performance of Arm’s AGI CPU against x86 CPUs (with SMT enabled and disabled) across three metrics: sustained performance per thread, sustained threads per rack, and performance per watt. Arm’s AGI CPU leads in all three categories, with roughly 1.2× higher performance per thread, up to 1.8–2.0× higher thread density per rack, and approximately 2× better performance per watt–highlighting Arm’s efficiency advantage in AI data center workloads, particularly for agentic AI applications where CPU orchestration, scalability, and energy efficiency are critical. Source: Arm Arm 

For more details on Arm, see my analysis from April: Arm Stock Could Win as Agentic AI Shifts the Bottleneck to CPUsArm Stock Could Win as Agentic AI Shifts the Bottleneck to CPUsArm Stock Could Win as Agentic AI Shifts the Bottleneck to CPUs 

In an air-cooled rack, Arm can pack 30 blades (or 60 CPUs) for a total of 8,160 cores in a 36kW power envelope, saying this configuration can deliver up to 2X the performance per rack versus x86 chips based on its internal estimates. Arm says this 30-blade design is “setting records for air cooled” racks that is not feasible with other systems, as power consumption is too high. 

Arm is taking this a step further with a fully-liquid cooled, 200kW open-standard rack in partnership with Super Micro, packing 168 blades, or 336 CPUs, delivering a total of up to 45,696 cores. Arm EVP of Cloud AI Mohamed Awad stated that while it is a “200-kilowatt rack. We actually will consume about half that much power. We ran out of space. That’s why we couldn’t put more cores in there.” 

This is one of the key advantages – it is not just about offering 2X the performance of x86 chips, but providing that performance boost while freeing up power for more compute or for more networking. 

Intel’s Fight to Maintain Server CPU Market Leadership 

Intel is positioning itself to be a stronger competitor in the server CPU market as it relates to agentic AI, announcing its intention to deploy rack-scale CPU systems at Computex. Intel’s new racks look to have quite an edge over Arm when it comes to core density, benefitting from its lead in cores at the individual chip level.  

Intel revealed two blueprints for its upcoming rack-scale CPU systems, with one design targeting maximum density and the other targeting latency-sensitive agentic AI workloads. The two designs can support 128 of Intel’s Granite Rapids Xeon 6 or Clearwater Forest Xeon 6+ chips, which will provide either 16,384 or 36,864 cores based on the chip of choice, alongside up to 384 TB of DDR5 memory per rack. 

Intel also leads on core counts at the individual chip level. Intel’s Xeon 6+ offers 288 cores per chip, slightly beating out AMD’s Venice at 256 cores and Arm’s AGI CPU at 136 cores; AMD has yet to release Verano’s core count. Despite the lead in core count, Intel’s Xeon 6+ only packs 288 threads whereas Venice offers up to 512 threads via multi-threading, allowing each core to handle two sets of instructions, reducing core idle time and increasing efficiency.

Bar chart comparing core and thread counts of major data center CPUs in 2026. AMD’s EPYC Venice leads with 256 cores and 512 threads, while Intel’s Xeon 6+ and Xeon 7 offer higher core counts at 288 but fewer threads due to no SMT. Nvidia Vera, AmpereOne, and Arm AGI CPUs have lower counts, while hyperscaler chips from AWS, Google, and Microsoft range from 64 to 192 cores. Source: TrendForce 

Intel vs Arm: Power Efficiency Battle 

Where Intel could find its edge in the rack-scale systems is power, with the blueprints fitting inside a 100kW power envelope. Compared to Arm, Intel is offering more than 368 cores per kW, while Arm’s AGI CPU is offering 228 cores per kW. This can also be viewed at the 200kW envelope of Arm’s AGI CPU, at which Intel could theoretically offer 73,728 cores across two 100kW racks, or more than 60% of Arm’s rack.  

This advantage stems from Intel’s 18A node, which in general offers up to 15% better performance per watt and up to 30% better density versus the Intel 3 node. Manufacturing on more advanced nodes is how x86 is fighting back against Arm – AMD’s Venice is the first CPU to ramp on TSMC’s 2nm (N2) process, which is designed to deliver 10%-15% higher performance at the same power level, or a 25%-30% reduction in power at the same  performance level.  

While Clearwater Forest just launched at the start of June, the most important part of Intel’s story is that its next-gen Xeon 7 ‘Diamond Rapids’ is rumored to be delayed. The new chip was originally expected to launch in the later part of 2026, yet is now expected to launch in 2027, giving AMD a bit more of a head start with Venice. 

CPUs Are the Next Major Bottleneck in AI Infrastructure 

It would be a mistake to think the AI trade begins and ends with Nvidia’s GPUs. Although GPUs are still the heart of AI compute, agentic AI is placing additional emphasis on making sure those accelerators do not sit idle while the rest of the system catches up. 

The addressable market is expanding overnight, as this no longer about adding a few more CPUs as head nodes next to GPU clusters. The bigger opportunity is the move to standalone CPU racks, which is a major architectural change that allows more orchestration capacity to be added without adding more GPUs at the same attach rate. Nvidia’s Vera rack is leading the way, with CEO Jensen Huang projecting roughly $20 billion in standalone CPU revenue this fiscal year, yet AMD, Intel, and Arm are not going to concede the market. 

This is the same framework the I/O Fund has used to identify massive AI winners across memory, networking and energy with CPUs now becoming the next bottleneck. Behind our paywall, we publish over 100 articles on the AI trade per year alongside portfolio allocations and real-time trade alerts.  

For example, we identified lesser-known AI winners, including Bloom Energy, up 1600% since our initial entry last year, a networking player that has delivered roughly 7X Nvidia’s returns YTD and an optical networking stock up more than 810% since November. 

Don’t miss out on the AI trade.
Join NowJoin Now

Please note: The I/O Fund conducts research and draws conclusions for the company’s portfolio. We then share that information with our readers and offer real-time trade notifications. This is not a guarantee of a stock’s performance and it is not financial advice. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis. Beth Kindig and the I/O Fund own shares in AMD and NVDA at the time of writing and may own stocks pictured in the charts. 

Leo Miller, AI and Semiconductor Investment Writer at I/O Fund, contributed to this analysis. Leo Miller owns shares of NVDA.

👉🏻 Share with a Fellow Investor
Heare with a Fellow Investor
Help someone else benefit from this insight.

Recommended Reading:

• Google TPU v8 vs Nvidia: How Inference Is Rewriting the AI Market
• The AI Networking Stock That Beat Nvidia by 7X YTD for Returns of 135% YTD
• Bloom Energy — Our 2026 Top Pick Was the Best Performing Stock in April
• Inside Nvidia’s $4B Optical Strategy—and Why CPO Changes Everything

• Google announced that it will begin selling TPUs to select third-party data center operators, marking the company’s formal entrance into the merchant AI accelerator market where Nvidia dominates
• The share of AI inference workloads is increasing; the shift toward inference is making the economics of custom silicon increasingly difficult for hyperscalers to ignore, and Nvidia may be facing a Rubin delay—three converging factors opening the door for TPUs
• The large coherent shared memory of TPU pods is a key feature that Google is banking on to differentiate from Nvidia systems

Google blew the doors off with its latest earnings report—cloud growth rapidly accelerated, margins expanded, and backlog soared 400% YoY to $462 billion. However, the quarter’s most pivotal development wasn’t in the financials, rather it came from a strategic announcement.

In April, Google announced it would begin selling its TPUs to select third-party data center operators, which is something the market has anticipated for nearly a decade. The TPU-versus-Nvidia-GPU debate has long fueled both bulls and bears; yet it may finally carry real stakes. Google’s announcement is far from a coincidence—it is driven by several converging factors that make now the right moment to move.

As hyperscalers look to monetize their models, AI workloads are expanding from training to inference. This changes the focus away from accumulating expensive compute to a very different goal, which is lowering cost per token in order to scale inference economically.

In a previous article covering Google’s TPUv7, we stated: “[…] custom silicon’s cost advantages and ability to drive lower inference serving costs at scale creates a strong value proposition for Big Tech.” Building on this, Nvidia may be facing a Rubin delay, which opens a window of opportunity for Google to make the case for diversification beyond a single vendor for AI accelerators.

Below, we look at how Google’s entrance into the merchant AI accelerator market sits at the center of three converging trends – and how the newly released TPU v8 generation positions custom silicon to meet the moment, giving Google a fighting chance against Nvidia.

The Shift from AI Training to Inference: Why It Opens a Window of Opportunity for Google

To understand why the market is opening up for more players, we should first discuss why inference is becoming the dominant AI workload—and what this means for Nvidia.

Training frontier models is a discrete, multi-month event with a clear beginning and end. By contrast, inference is the revenue-generating phase, and thus, runs continuously with no ending point. Both training and inference workloads will continue to grow as labs build better models and monetize them. However, the always-on nature of inference will result in inference being the higher volume workload over time.

According to industry analysts, inference could take the larger share as soon as 2027.  McKinsey estimates that in 2026, 31.2 GWs of data center demand will be allocated to training, and 31.2 GWs will be allocated to inference—an even 50/50 split. However, by 2027, inference becomes the larger share. By 2030, inference accounts for 93.3 GWs of demand, compared to training’s 62.2 GWs—or a 60/40 split.

Google TPU v8 Explained: 8i vs 8t and the Inference Advantage

At Cloud Next in late April, Google unveiled its latest TPU v8 in two configurations—the training-optimized 8t and the inference-optimized 8i. Notably, the Ironwood TPU v7 was the first TPU optimized for inference, but v8 marks the first time that the architecture has been split for two distinct purposes. As Google looks to capitalize on inference becoming the primary AI workload, splitting the v8 into two separate chips allows it to target this part of the market more effectively.

TPU v8 Architecture: Why Google Split Training and Inference

With the 8i, Google is positioning itself to beat out Nvidia on one key aspect – coherent shared memory, a key anchor in improving inference efficiency.

While the 8i’s pod size only scales 4.5X over Ironwood’s 256-TPU pod to 1,152 TPUs per pod, pod-level HBM capacity increases by 7X to 331.8 TB versus 49.2 TB with Ironwood. Yet the key here is that this HBM capacity is coherent across the pod, across all 1,152 chips.

This is arguably the most critical point to understand surrounding Google’s architectural advantage with the 8i, that this 331.8 TB of memory is shared across the entire pod over Google’s inter-chip interconnect (ICI). ICI is similar to Nvidia’s NVLink—with both allowing for the fastest chip-to-chip memory access within a pod. Compare this to Nvidia’s NVL72, where true memory coherency only extends at rack-scale across 72 GPUs and just 20.7TB of HBM. Scaling out to 1,152 of Nvidia’s GPUs would span 16 racks, yet memory does not become a unified pool shared across the entire cluster.

By keeping the maximum amount of memory in a shared domain with the TPU 8i, large frontier models with long context windows can run with minimal latency.

How TPU v8i Lowers Cost Per Token: SRAM and Boardfly

Several other key decisions reinforce the 8i’s inference capabilities—pursuant to the ultimate goal of increasing inference efficiency by reducing latency, helping reduce cost per token as inference and agentic AI expand. These include boosting SRAM capacity per chip, and introducing a new networking topology, dubbed Boardfly.

SRAM capacity is where Google is driving latency improvements at the chip level, increasing on-chip SRAM by 3X to 384MB for the 8i. SRAM is the fastest memory available to a chip, and the larger pool allows more of the chip’s working memory, or KV cache, to stay on the fastest tier possible. In doing so, latency falls as the KV cache does not have to be retrieved from slower HBM. With 1,152 chips, the pod’s total SRAM capacity is 432 GB.

Google’s new Boardfly topology is its second lever in reducing latency. With Boardfly, Google connects ‘building blocks’ of four TPUs into boards, consisting of eight building blocks, that are then fully linked together as one pod. This is achieved via direct optical long-haul links, flattening the topology and reducing networking hops for any chip-to-chip communication from 16 hops to just seven. Google says this reduction in hops drives a “50% improvement in latency for communication-intensive workloads.”

As stated, the result of these improvements is lowering the 8i’s cost per token. In line with this, Google notes that TPU 8i delivers up to an 80% performance-per-dollar improvement over the Ironwood TPU, particularly at low-latency targets for large MoE models. The 8i’s deployment would compound the already significant serving cost reductions Google achieved in 2025. Last quarter, Google’s CEO stated there was a 78% reduction in Gemini serving unit costs in 2025.

As chips spend less time sitting idle, Google—or any other TPU operator—can process more tokens at the same price. This strikes at the core of inference economics—minimizing the cost per token.

Deploying agentic AI within enterprises dramatically increases the need for memory in comparison to chatbots. Agents can act autonomously, performing complex multi-step tasks, drawing from organization-specific workflows, policies, and data—all of which require increased memory. Overall, Nvidia notes that agentic systems consume up to 15X more tokens than traditional AI applications. As token consumption vastly increases, lowering cost per token is critical to scaling agentic AI efficiently.

mid

Nvidia Prepares to Answer on Inference

While Google is deploying an inference-optimized TPU that warrants attention, from its ability to offer 331.8TB of shared coherent memory at pod level alongside other topology and architectural optimizations to improve inference efficiency, Nvidia remains the world’s best chip designer, and will not simply lay down and concede the inference market.

On that note, Nvidia is moving quickly with a different approach via its 256-chip Groq LPX rack, leveraging Groq’s SRAM-based design to accelerate inference-based workloads via ‘disaggregation’ at rack scale. As covered in our free newsletter, Nvidia Stock to See New Growth Catalyst; 35X Faster AI with Groq 3 LPX, disaggregation refers to splitting up the two-step process of token generation, prefill and decode, and allocating each step to the hardware best designed for the task – prefill goes to compute-heavy Rubin GPUs, and memory and KV-cache intensive decode to the LPX rack.

Nvidia CEO Jensen Huang stated that combining the two co-designed racks can deliver up to 35X higher throughput per MW on trillion-parameter LLMs, with these throughput gains most evident on high token rate applications, such as real-time AI agent communication.

Naturally, there will be architectural differences between custom silicon and GPUs, such as the TPU 8i leveraging on-chip SRAM, yet the key takeaway is that Nvidia is moving ahead with a new strategy. The strategy, in a nutshell, is to offer seven co-designed chips that offload tasks to specialized hardware and optimize inference at the rack/system level versus the chip level.

Nvidia is the world’s best AI chip design company, and all the above plus other incoming rapid changes to the company’s product roadmap is something to keep a close eye on.

For more information on why Nvidia’s CUDA moat matters less with inference, read our analysis here: Nvidia’s $20 Trillion Thesis in Intact, my 2026 Allocation Isn’t.Nvidia’s $20 Trillion Thesis in Intact, my 2026 Allocation Isn’t.

How Lower Token Costs Are Driving Google Cloud Growth and Margins

In Q1 2026, Google Cloud put up a hallmark performance. Revenue came in at $20 billion, with growth accelerating to 63% YoY. This was nearly double the 32% growth seen in Q2 2025 and 15 percentage points higher than the 48% growth seen in Q4 2025. Cloud backlog also hit $462 billion, up 400% YoY and 90.3% QoQ, signaling both the massive scale and acceleration of demand.

However, just as important was the huge expansion in Cloud operating margin. The figure moved up to 32.9%, a 15.1 percentage point expansion YoY and a 2.8 percentage point expansion QoQ.

Gemini vs GPT vs Claude: Token Pricing Comparison

Connecting back to the TPU discussion, lowering token costs is key to Google Cloud’s success. By keeping costs low, Google can attract more developers to Gemini, generating more cloud revenue. Gemini 3.1 Pro Preview, Anthropic’s Claude Opus 4.7, and OpenAI’s GPT-5.5 are widely considered frontier models—but data from Artificial Analysis indicates that Google has a very significant cost advantage.

The blended price per 1M tokens that customers pay on Gemini 3.1 is approximately $1.74. This is around 58% lower than Claude 4.7 and 60% lower than GPT 5.5. Additionally, this difference comes even as Google increased the per-token cost of Gemini 3.1 Pro Preview by 30% over Gemini 2.5 Pro.

Bar chart compares the blended price per 1 million tokens across leading AI models from Google, OpenAI, and Anthropic. Google’s Gemini 3.1 Pro Preview is priced at approximately $1.74 per million tokens, making it roughly 58% cheaper than Anthropic’s Claude Opus 4.7 ($4.10) and 60% cheaper than OpenAI’s GPT-5.5 ($4.35). Earlier models such as Gemini 2.5 Flash ($1.34) and GPT-5.4 Mini ($2.18) are also included for historical context. Source: Artificial Analysis

By leveraging Ironwood TPU v7 and TPU v8, Google can attract more developers while balancing operational leverage—creating a perfect storm for the growth acceleration and margin expansion we are seeing today. Furthermore, Google Cloud’s 33% operating margin and the large expansion in this figure provide evidence that the company is not deeply subsidizing its token costs to gain share.

The up to 80% reduction in performance-per-dollar from Ironwood to 8i can allow Google to continue lowering its own costs—benefiting margins further. Additionally, with token costs still much lower than other frontier models, Google could choose to boost margins through price increases.

The distinction here is that Gemini is served exclusively on TPUs, while Claude and GPT-5.5 are not (although TPUs are part of Anthropic’s infrastructure stack). As we isolate this variable across the frontier model providers, we can reasonably assert that the fundamentally different architecture that Gemini runs on—TPUs—are a key driver of Google’s lower cost per token.

Anthropic’s TPU Bet: What It Signals for AI Infrastructure

Anthropic’s large partnership with Google provides further evidence of TPU competitiveness. Anthropic has been growing at a breakneck pace, with recent estimates suggesting that the company’s ARR increased from $9 billion at the start of 2026 to now over $44 billion. This clearly positions Anthropic as scaling inference and monetization, and the firm is making long-term commitments with Google – which sends a clear message. Anthropic has reportedly expanded its partnership with Google, agreeing to a 5 GW TPU deployment over the next five years, with additional GWs possible. This is a notable expansion of its previously announced agreement for 3.5 GWs.

One reason for this move is the fact that a rapidly growing AI lab like Anthropic simply needs to secure additional compute capacity. Anthropic has also announced compute capacity expansions that run on Nvidia hardware—including an up to 1 GW deal with Azure and an over 0.3 GW deal with SpaceX. However, the scale of these agreements is clearly much smaller than the TPU deal, which could indicate that Anthropic is benefiting from Google’s TPU advantages in lowering token costs.

Anthropic’s Compute Strategy Across Google, AWS, and Azure

Today, Amazon is Anthropic’s primary cloud provider, utilizing the firm’s Trainium chips. This comes as the bulk of Anthropic’s TPU capacity will not start to come online until 2027. Anthropic has also committed $100 billion over the next ten years to AWS, allowing it to secure up to 5 GW of new capacity. However, one report suggests that it's commitment to Google Cloud is worth $200 billion over the next five years—or double the spending in half the time. This is another data point implying that Anthropic sees TPUs as highly competitive with both Nvidia and Amazon hardware.

With Anthropic being one of the preeminent companies pushing the AI world into the inference phase, its support of TPUs validates the thesis that Google can drive forward merchant sales. Notably, Google is already providing evidence of its ability to drive merchant sales, launching an AI cloud joint venture with Blackstone that aims to deliver its first 0.5 GW of TPU capacity in 2027.

Nvidia Rubin Delay: A Strategic Opening for Google TPUs

Lastly, the reported one-quarter delay of Nvidia’s Rubin ramp, officially scheduled for Q3 2026, could offer a strong argument for diversification across AI accelerators. Notably, TrendForce revised its estimate of Rubin’s contribution to Nvidia’s total high-end GPU shipments for 2026 down from 29% to 22% to account for such a delay.

Factors contributing to the reported delay and TrendForce’s revision include “the time required to validate the newer HBM4 memory used by the chips, challenges with the migration to Nvidia's faster ConnectX-9 NICs, the system's higher overall power consumption, and the more advanced liquid cooling requirements.”

While Nvidia has not lent credence to delay rumors itself, statements made on the company’s latest earnings call provide clues into the trajectory of the Rubin ramp.

Joshua Buchalter, TD Cowen

“Colette, I believe, in your prepared remarks, you mentioned GB300 is sort of the fastest ramp in the company's history. How should we think about Vera Rubin against this benchmark?”

Colette Kress, Nvidia CFO

“Yes. Well, we've indicated for a while that we will be launching Vera Rubin in the second half. We will start in Q3. That will be our initial pieces together. And then once we get to Q4, we're probably going to start to see our ramping continue… It's hard to say at this point what will be a faster ramp… But yes, we're going to start in Q3 and continue to ramp into Q4. And Q1 of next year certainly is going to be very big as well.”

If we take what the CFO stated, Rubin systems meaningfully ship Q4-Q1. Specifically, it was noted that in Q3 Nvidia would bring together the “initial pieces” and that the ramp would “probably” continue in Q4. This is far from a definitive statement that the Rubin ramp will take off in Q3. If anything, Kress seemed to position Q1 2027 as the large ramp—adding weight to the delay rumors.

Delays in Nvidia’s roadmap have happened before, such as the two-quarter delay experienced with Blackwell. What’s different now is that a merchant alternative optimized for inference is available through Google.

Final Thoughts: Why Google May Be Nvidia’s Strongest AI Challenger Yet

Google’s inference-optimized TPU 8i is targeting the fastest growing segment of the compute market, with meaningful advantages in lowering cost per token. Google Cloud growth is accelerating, operating profitability is compounding, and leading AI labs like Anthropic are validating the merchant TPU thesis. As Google steps into the AI accelerator arena, it’s one of the few legitimate challengers to Nvidia’s dominance.

Meanwhile, Nvidia iterates and improves its systems at an unusually fast pace. It may not be long before the AI juggernaut responds with a much stronger answer to Google’s v8 series.

Regardless, our thesis is that neither Google nor Nvidia is likely to offer the highest returns in the AI trade from here. Instead, we think the best opportunities will come from the companies that supply the world’s most valuable firms with networking, energy, memory components, and other critical AI infrastructure.

The I/O Fund has excelled at shifting our thesis when presented with new evidence while others stick to what is familiar. For example, we identified lesser-known AI winners, including Bloom Energy, up 1100% since our initial entry last year, a networking player that has delivered roughly 7X Nvidia’s returns YTD and an optical networking stock up more than 790% since November.

We publish more than 100 paywalled articles each year on AI stocks, supported by an actively managed portfolio and real-time trade alerts. Don’t miss out on the AI trade.
Learn more here

Please note: The I/O Fund conducts research and draws conclusions for the company’s portfolio. We then share that information with our readers and offer real-time trade notifications. This is not a guarantee of a stock’s performance and it is not financial advice. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis. Beth Kindig and the I/O Fund own shares in GOOGL and NVDA at the time of writing and may own stocks pictured in the charts.

Leo Miller, AI and Semiconductor Investment Writer at I/O Fund, contributed to this analysis. Leo Miller owns shares of GOOGL and NVDA.

👉 Share with a Fellow Investor
hare with a Fellow Investor
hare with a Fellow Investor
Help someone else benefit from this insight.

Recommended Reading:

• The AI Networking Stock That Beat Nvidia by 7X YTD for Returns of 135% YTD
• Bloom Energy — Our 2026 Top Pick Was the Best Performing Stock in April
• Inside Nvidia’s $4B Optical Strategy—and Why CPO Changes Everything
• Nvidia’s $20 Trillion Thesis Is Intact. My 2026 Allocation Isn't

April 2026 was a historically strong period for the stock market. The S&P 500 rose by 10.43%—its best monthly return since April 2020, when the market rebounded from COVID-era lows. The Nasdaq-100 achieved the same feat, rising over 15%. 

Against this backdrop, I/O Fund performed exceptionally well, as we owned 4 of the market’s 10 best-performing large caps in April. This includes the market’s top large cap gainer, Bloom Energy, which soared 112.81%. In February, my firm called out Bloom as our Top AI Stock Pick for 2026.  

Adding to AI energy's strong performance was networking stocks, a subsector that most investors shy away from due to the complexity of the products, in addition to the supply chain moving lightning fast with immense volatility in both directions. The reason networking sees immense volatility is straightforward: much of the market is tied to a single customer, Nvidia; and Nvidia is rolling out new architectural iterations at an unusually fast pace these days. 

AI networking stock Lumentum is among the key I/O Fund winners in 2026. We allocated heavily to LITE in January—a month before Nvidia backed the company. While most investors couldn’t stomach taking a stake in this stock that soared 339% in 2025, I/O Fund built a 9% position that has since paid off in spades. Overall, in just five brief months, our Lumentum position delivered a return of 135.4%, or 6.8X higher than Nvidia’s 19.9% return since the end of January. 

For investors new to this name, Lumentum recently received significant validation from the world’s most valuable company—Nvidia—with the dominant force in AI infrastructure investing $2 billion in LITE. However, the importance of this goes far beyond the investment itself. The real story is Lumentum’s central position in Nvidia’s multi-year networking roadmap, and the broader AI market, which is affording Lumentum the opportunity to grow its business several times over. 

Below, we break down the key dynamics currently benefiting Lumentum, the structural factors supporting continued margin and EPS growth, and our perspective on the key question: “Is it too late?” 

Nvidia–Lumentum Partnership: CPO Growth and Optical Transceiver Market Expansion 

Nvidia’s partnership with Lumentum includes multi-billion-dollar agreements on two fronts: the investment and a purchase commitment for the company’s ultra-high-powered lasers (UHPs). 

Currently, Lumentum is the sole supplier of UHPs for Nvidia’s co-packaged optic (CPO) networking switches—which are expected to undergo a step function in demand over the coming years. Nvidia has already taken up nearly all of Lumentum’s UHP capacity, leaving little for other customers. 

The $2 billion investment is key to expanding Lumentum’s existing UHP capacity in San Jose, its Caswell fab in the United Kingdom, and bringing online its recently acquired fab in Greensboro, North Carolina. 

This is all due to the dramatic ramp-up of CPO demand that Nvidia is preparing for. Lumentum expects to generate its first $100 million in CPO-related revenue in the final quarter of calendar 2026, but the longer-term opportunity is much larger. More on this later. 

Overall, Nvidia sees Lumentum as a vital partner in this ramp-up and is making significant commitments to ensure capacity once CPO takes off. This ties Lumentum directly to the world’s preeminent AI infrastructure company over a multi-year period. And, even as CPO has yet to penetrate significantly into data centers, Lumentum is posting extremely strong financial results. This is driven by insatiable demand for high-speed optical transceivers.

Get $275 off Advanced Market Signals for our 6-year anniversary sale. Access real-time trade alerts, weekly webinars, and 100+ research articles on the AI trade per year. Claim OfferGet $275 off Advanced Market Signals for our 6-year anniversary sale. Access real-time trade alerts, weekly webinars, and 100+ research articles on the AI trade per year. Claim OfferClaim Offer

Optical Transceiver Market Growth Forecast (2025–2026) 

TrendForce notes that "the global market for AI-focused optical transceivers has entered a phase of rapid growth” and projects the market will expand from $16.5 billion in 2025 to $26 billion in 2026—growing 57.6% YoY. 

Chart illustrating the rapid expansion of the AI optical module market, with revenue projected to increase by 57.6% year-over-year from $16.5 billion in 2025 to $26 billion in 2026, driven by rising demand for high-speed optical transceivers such as 800G and 1.6T. Source: TrendForce (April 2026).

800G modules are the primary growth driver, with shipments of 1.6T units ramping in mid-to-late 2026. TrendForce also predicts that “optical transceivers shipments of 800G and higher will hit 24 million units in 2025, then jump by 2.6 times to nearly 63 million units in 2026.” Within this, Yole Group forecasts over 10 million 1.6T module shipments in 2026.  

Lumentum is a critical player here, generating growth on two sides of the transceiver coin. Lumentum sells its own 800G and 1.6T transceivers, and is a supplier of key components to other transceiver makers. 

Lumentum Financials: Record Revenue Growth and Precipitous Margin Expansion 

Lumentum’s latest results showed that the company is indisputably firing on all cylinders. It’s Q3 FY2026 ended in March, with Lumentum posting revenue of $808.4 million. The figure missed analysts' estimates very slightly (0.2%), but sales still grew by 90.1% YoY. This marked the fastest YoY growth rate in Lumentum’s history and was a large acceleration over 65.5% YoY in Q2. 

The company expects growth to accelerate further next quarter. It projects sales of $980.5 million at the midpoint, implying growth of 104.9% YoY. Additionally, after QoQ growth decelerated from 24.7% in Q2 to 21.5% in Q3, its midpoint guidance projects solid consistency with 21.8% in Q4. 

Lumentum’s Margin Expansion Across Gross, Operating, and Net Income 

The margin story was equally impressive, driven by improved manufacturing utilization, favorable product mix, and operating leverage. 

• FQ3 adjusted gross margin improved by 12.7 percentage points YoY to 47.9%, supported by utilization gains. 
• FQ3 Adjusted operating margin rose by 21.4 percentage points YoY to 32.2%, benefiting from gross margin improvements and operating leverage. 
• FQ3 adjusted net income margin rose by 18.3 percentage points to 27.9%. 

Adjusted net income expansion was moderately less than operating margin expansion, largely due to higher income tax provisions. However, higher taxes are simply the cost of doing business when adjusted net income soars 184.8% YoY to $225.7 million. In Q4, management projects further adjusted operating margin improvement, with the figure moving up to 35.5%, or a 3 point QoQ gain. 

Key Growth Drivers Powering Lumentum’s AI Networking Business 

Lumentum is achieving this growth without large sales from the Nvidia UHP partnership, as UHP shipments have yet to ramp significantly. Instead, electro-absorption modulated (EML) lasers, narrow linewidth and pump lasers, and optical transceivers are driving growth. 

EMLs are lasers used within optical transceivers for scale-out networking applications, with the company selling them as components, and using them in its own transceivers. Notably, Lumentum saw record EML shipments in FQ3. 100G shipments drove this, but 200G revenue also more than doubled QoQ. 

Narrow linewidth and pump lasers are used in scale across applications—connecting geographically separated data centers. Pump laser sales grew rapidly by 80% YoY, and narrow linewidth lasers saw their ninth consecutive quarter of growth, with sales rising 120% YoY. EML's, narrow linewidth lasers, and pump lasers helped the company’s Components revenue rise by 77.3% YoY to $533.3 million, accounting for 66% of total revenue. 

This strong growth comes even though Lumentum is capacity constrained across all three components. The firm is working to expand EML capacity at its Japan fab, expecting to increase EML units by over 50% by December 2026 versus a December 2025 baseline. 

When it comes to pump and narrow linewidth lasers, Lumentum says it is “effectively sold out for the foreseeable future." Notably, pump lasers are even more constrained than EMLs. 

Cloud Transceivers and Systems Revenue Expansion 

Cloud transceivers grew 40% QoQ with record shipments, likely driven mostly by 800G units. Cloud transceivers represent most of Lumentum’s System sales, which rose 121% YoY to $275.1 million, or 34% of total revenue. As EMLs are used in transceivers, the company is also facing significant capacity constraints here. 

The takeaway is that Lumentum is shipping these various products at a rapid and, in many cases, record pace, and still under-shipping the market. Demand is pent up, putting pricing leverage on Lumentum’s side, and creating future growth opportunities. All the while, demand specific to the Nvidia relationship has yet to meaningfully kick in. 

Indium Phosphide (InP): The Chokepoint Material in Optical Interconnects 

Across its business, Lumentum’s indium phosphide (InP) processing capacity is the unifying constraint holding back laser output. InP is the specialized semiconductor material that all of the discussed products are built on, with ideal properties for optical communication. 

Thus, ameliorating the constraints in InP wafer processing is key to meeting customer demand. Notably, InP capacity constraints come even as Lumentum leads the market, saying “We probably have more [indium phosphide] capacity than any company on the planet." 

Lumentum’s InP Capacity Expansion Plans 

Lumentum is making strides to increase its InP processing capacity. From the last quarter of calendar 2025 to the last quarter of calendar 2026, the company plans to increase its InP capacity by 50% while already having the industry’s largest base. This is a meaningful increase over the company’s past statements of expanding capacity by 40%. 

Despite all of this, the company is still under-shipping drastically, by more than 30% as of FQ3. Furthermore, as its UHP business scales, Lumentum expects the gap between supply and demand to widen. 

While this is a negative for unit growth, supply and demand imbalances can provide significant benefits to margins and EPS. The memory chip market shows how companies that control undersupplied AI infrastructure products are in a very favorable position. 

Supply Constraints Driving Pricing Power 

With InP imbalances expected to grow, Lumentum’s margins and EPS can be prime beneficiaries of this dynamic. 

Lumentum CEO Michael Hurlston substantiated the company’s pricing power recently, stating negotiations are on “very favorable terms” with non-Nvidia buyers. This comes as Nvidia will soak up much of its InP capacity, creating a “little bit of a feeding frenzy” among other players. 

It is important to note that InP constraints extend beyond the wafer processing layer. InP substrates are the most upstream input for InP-based products, and a set of concentrated suppliers controls this layer. Japanese firms Sumitomo Electric and JX Nippon Mining, as well as AXT (U.S.-headquartered, Chinese manufacturing), are the top names. China and the United States have created geopolitical risks at this level. 

Geopolitical Risks in InP Supply Chain 

Per AXT’s 10-K filing, China placed InP substrates on its export control list in February 2025, requiring an export permit for every order. Meanwhile, in March 2025, the United States placed 70% tariffs on Chinese products, including substrates. Supreme Court rulings have invalidated certain tariffs, but others remain in place. 

These factors have had a significant impact on AXT. The company saw full-year revenues fall 11% YoY in 2025, and North America fell from 10% of total sales to 1% in its latest quarter. This adds pressure to companies like Lumentum looking to ramp up InP wafer processing, with Hurlston noting, “the thing that keeps me up at night most is substrates.” 

However, Lumentum has worked to mitigate this risk through long-term supply agreements, signing a 7-year substrate supply deal with a non-Chinese firm that extends through the mid-2030s. Hurlston says Lumentum worked with this partner to “corner the supply of their indium phosphide substrates," presumably securing a very significant share of their capacity. 

With this agreement, Lumentum says it is in “pretty good shape on substrates." However, it needs to continue securing supply as laser output is going to have to take a “massive” step up in 2027, given CPO demand. 

Future Catalysts for Lumentum: 1.6T, OCS and CPO Growth 

Beyond products driving current results, Lumentum has two near-term catalysts layering on: the 1.6T transceiver ramp alongside the insourcing of continuous wave (CW) lasers. Lumentum is set to ramp its higher bandwidth 1.6T transceivers in FQ4. Combining its strong pricing power on 800G modules and ramping already higher-margin 1.6T modules should allow for further margin expansion. 

At the same time, Lumentum will insource more of the CW lasers used in its transceivers, expecting insourced CW lasers to be in ~20% of transceiver modules in FQ4. Reducing its reliance on third-party CW laser suppliers should benefit transceiver gross margin and alleviate some external supply constraints. 

Optical Circuit Switching (OCS) Growth Opportunity 

Optical circuit switching (OCS) is another key demand driver. Lumentum has recently signed a multi-year, multibillion-dollar order with an OCS customer. Currently, Lumentum holds a $400 million OCS backlog, which is “very much on track to be shipping” in the second half of calendar 2026. In calendar 2027, Lumentum expects OCS revenue to ramp above $1 billion. 

Importantly, Lumentum addressed concerns that this was simply a “bubble order" or a large one-time deal that will not be repeated. The firm definitively said that was not the case, and instead stated, “We would expect to see significantly more business across calendar '27 on the OCS.” This strongly suggests that the firm expects additional OCS orders from current or new customers in 2027. 

CPO Runway and Revenue Potential 

However, CPO demand, anchored through Lumentum’s Nvidia partnership and expected to broaden across hyperscalers over time, is the potential game changer. The company describes the CPO opportunity as being in three phases. Phase 0 is scale-out CPO, where Lumentum expects to generate $100 million in revenue in the final quarter of calendar 2026. The company will then deliver on a multi-hundred-million-dollar scale-out commitment in H1 2027. 

Lumentum’s market in scale-up CPO is drastically larger. Phase 1 CPO involves connections between racks in scale-up pods and is 3X-4X larger than its Phase 0 scale-out CPO opportunity. The opportunity in Phase 2 scale-up CPO, which involves links within each rack as copper gets displaced over even shorter distances, is 10X larger than Phase 0.

Graphic illustrating the evolution of co-packaged optics (CPO) deployment across AI data centers. Phase 0 begins with scale-out architectures using single-rack clusters. Phase 1 expands to multi-rack clusters with inter-rack optical links, increasing CPO connections by 3X to 4X. Phase 2 advances to intra-rack optical connectivity, significantly increasing link density and enabling larger-scale compute clusters. Source: Lumentum.

Phase 1 scale-up shipments are expected to begin in H2 2027 and ramp significantly in 2028 and beyond. The Phase 2 opportunity is expected to inflect in late 2029 and 2030. Increasing InP capacity is critical to meeting scale-up demand, which is exactly what Lumentum’s newly acquired facility in Greensboro is designed to accomplish. 

After retrofitting the site for InP-based devices, Lumentum expects to ramp UHP production in mid-2028. Greensboro is the ‘moonshot’ opportunity for Lumentum, with the company targeting $5 billion of incremental annual revenue capacity through the facility. Compared to Lumentum’s last 12 months' revenue of $2.488 billion, the full ramp-up of Greensboro alone has the potential to triple the size of its business.  

Greensboro will also come with semiconductor-like margins—which are structurally higher than transceiver margins. This provides Lumentum with another opportunity to significantly improve its profitability profile. 

Lumentum’s AI Networking TAM Expansion 

Through the combination of four networking markets: scale across (narrow linewidth and pump lasers), scale-out (CPO, transceivers, and transceiver components), scale-up (CPO), and OCS, Lumentum sees its optical AI total addressable market (TAM) expanding massively. From 2025 to 2030, Lumentum forecasts a 5X TAM increase, surging from $18 billion to $90 billion. 

Chart illustrating Lumentum’s projected expansion in the optical AI total addressable market (TAM), growing from approximately $18 billion in 2025 to over $90 billion by 2030, representing ~40% CAGR. Growth is driven by scale-out networking, scale-up architectures, and optical circuit switching (OCS), with 2030 bandwidth demand split roughly between 55% scale-out and 45% scale-up. Source: Lumentum.

The Ultimate Question: “Is it Too Late?” 

Clearly, Lumentum has a huge opportunity ahead of it to not only grow revenues but also expand margins. Fundamentally, this opportunity rests on the supply and demand dynamics in the optical networking market. Lumentum is already under-shipping the market by over 30%, despite having market-leading InP capacity today. 

Even as Lumentum increases capacity by 50% from December 2025 to December 2026, it expects the gap to widen. This setup puts more power in Lumentum’s hands over time, benefiting pricing, margins, and EPS. The dramatic rise in both Lumentum’s sales and margins today shows this dynamic is already playing out, and the factors driving it are getting stronger, not weaker. 

Still, investors have taken notice of this, as well as Lumentum’s very strong financial performance. Based on FY 2027 earnings estimates, Lumentum trades at a forward P/E ratio of 48.67X, approximately 109% above its average forward P/E of 23.25X since the start of 2023. 

Chart illustrating Lumentum’s forward price-to-earnings (P/E) ratio over time, showing a recent rise to 63.2x, significantly above its three-year median of 33.8x. The valuation expansion reflects strong investor expectations for continued growth in AI networking, optical transceivers, and co-packaged optics (CPO) demand. Source: YChartsYCharts

However, analysts forecast a dramatic rise in the denominator, reflecting continued transceiver growth and the ramp-up of scale-out CPO, OCS, and the early stages of scale-up CPO, boosting revenues and margins. 

EPS Growth and Forward Valuation 

Estimates place Lumentum’s NTM adjusted EPS at roughly $15.84. In Lumentum’s FY 2028 (the 12 months ending calendar Q2 2028) that estimate rises to $28.12. Using this figure puts Lumentum’s forward P/E at 31.65X, drastically closer to its average since the start of 2023. And, critically, this comes before Phase 2—Lumentum's largest CPO opportunity—gets underway. 

All the while, Lumentum benefits from having the largest base of InP capacity and expanding that capacity significantly, leaving the firm in a prime position to service scale-up demand. When considering these factors, Lumentum’s valuation starts to look much more reasonable. 

Final Thoughts: Looking Beyond Nvidia 

Networking stocks are notoriously difficult to navigate. The products are complex and supply chains shift quickly, particularly for those supplying Nvidia. This can create the kind of volatility that pushes most investors to the sidelines. Lumentum’s 339% run in 2025 is a perfect example: most funds allocated heavily to Nvidia instead of recognizing that run would likely continue into 2026. Five months later, and Lumentum surged another 135.4%. 

The broader AI trade problem illustrates this well, as many funds treat Nvidia as their entire AI allocation. Meanwhile, the AI infrastructure buildout is expanding across power, networking, memory and more, and the best-performing names are increasingly not Nvidia. 

The I/O Fund owned 4-of-10 top large cap performers in April, which illustrates the importance of looking more broadly at the AI trade. Our large Lumentum position is up 135.4% YTD – or 6.8X what Nvidia returned in the same window. Lumentum is one example; along with Bloom Energy and another networking stock up 750% since November, leading to YTD returns of 48% – nearly 3X the Nasdaq and 5X popular ETFs. 

To celebrate six years of the I/O Fund, we're offering our deepest discount of the year: $275 off Advanced Market Signals through May 27. 🥂

Members get our Q2 Top 15 AI Stocks report that runs over 70 pages and identifies the 15 stocks I expect to lead the AI market this quarter. The report is built on the same investment discipline that drove our YTD outperformance and surfaced Lumentum, Bloom Energy, and other major AI winners. 

Since May 2020, our audited portfolio has returned 326% cumulatively — a track record that would place I/O Fund #1 versus hedge funds and #3 versus tech ETFs or mutual funds, before including this year's 48% YTD return.
Sign up here

Leo Miller, AI and Semiconductor Investment Writer at I/O Fund, contributed to this analysis. Leo Miller owns shares of NVDA.

Please note: The I/O Fund conducts research and draws conclusions for the company’s portfolio. We then share that information with our readers and offer real-time trade notifications. This is not a guarantee of a stock’s performance and it is not financial advice. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis. Beth Kindig and the I/O Fund own shares in LITE at the time of writing and may own stocks pictured in the charts.

Recommended Reading:

• Bloom Energy — Our 2026 Top Pick Was the Best Performing Stock in April
• Inside Nvidia’s $4B Optical Strategy—and Why CPO Changes Everything
• Is Nvidia Stock a Buy? Why Semiconductor Strength May Signal a Market Top
• Nvidia’s $20 Trillion Thesis Is Intact. My 2026 Allocation Isn't

Last month was the S&P 500’s best month in six years, marking the biggest rally since the Covid lows in April of 2020. The S&P 500 rose 10.43%, while the Nasdaq gained more than 15%. Yet the single best-performing large-cap stock in that historic month was not Nvidia, Microsoft, Meta, or another obvious AI leader. Rather, it was Bloom Energy, which rose roughly 109%. 

Bloom was our 2026 Top Stock pick, published on February 27th when shares were at $160.90. However, my firm’s history on the stock began one year ago when we first identified AI energy as the next bottleneck, with initial buys during the April lows at $16.64 and $17.04. For most of the last 12 months, we’ve held Bloom at a high allocation of 10% or higher, with real-time trade alerts sent to our Research Members. Today our returns from those entries are roughly 1300%. Many of Wall Street's most renowned firms eventually followed the I/O Fund much later and entered at significantly higher prices. 

Earlier this year, I designated Bloom as our Top 2026 Stock Pick on February 27 when shares were at $160.90 (about 10X on our cost basis).  

The decision to place Bloom as our Top Stock pick required strong conviction — not only in Bloom Energy's positioning, but in the sheer pressure from AI's primary bottleneck to believe the stock could see a repeat year of strong performance. Repeat years are especially rare after a big run-up as early investors typically book gains and move on. 

Below, I’ll walk you through why Bloom outperformed in the strongest rally tech has seen in six years, why the recent Q1 2026 results confirm the fundamentals beneath the rally, and why I believe the setup still holds – even after the stock rose 109% in April, just two months after we named it our 2026 Top Stock pick. 

Why an Energy Stock — Not Software or Semiconductors — Led Tech’s Biggest Rally in 6 Years 

Investors should take note that tech’s biggest month in six years was not led by a Mag 7 stock, a semiconductor, or a software platform like it was in 2020. Although many of these sectors were deservedly ranked in the top 10, the month’s biggest outperformer was centered around power availability.

The reason for this is straightforward as companies like Microsoft, Google and Meta are spending hundreds of billions annually on AI, with tens of billions allocated to Nvidia’s GPUs and custom silicon like Google’s TPUs. These systems risk being delayed if Big Tech cannot energize and deploy them quickly. Meanwhile, the market has already penalized these companies for outsized spending on AI infrastructure. The effects of low immediate ROI only compound with a timing risk as GPUs sit idle, while competitors who do have power amplify the consequences of a delay. 

Despite power being the primary bottleneck, the market is hyper-focused on whether Big Tech can monetize AI. My contention in my original article on Bloom Energy is that the market is missing the point. the market is missing the point. Monetization has never been Big Tech’s weakness as explosive revenue growth and high margins have defined their businesses for decades. While execution risk always exists, these companies remain the world’s most reliable operators at scale.   

Instead, the real risk to the AI economy lies in the physical constraints of scaling these AI ambitions — not in compute availability from companies like Nvidia or Broadcom, and certainly not in Big Tech’s software capabilities, but in power availability, thermal management, and infrastructure that were never designed for this magnitude of demand. 

Bloom Energy Solves AI’s Most Critical Bottleneck: Time to Power 

Over the next two years, Nvidia’s GPU systems are expected to require a 5x increase in power per rack from what was needed in the first half of 2025 as we move across GPU generations from Blackwell to Rubin Ultra. As stated, if hyperscalers cannot energize these systems quickly, billions of dollars of AI capex can sit idle, especially critical now that the AI market is shifting toward generating inference revenue. 

Therefore, due to the rapidly increasing power requirements for AI systems, it is not enough to say the AI economy needs more power, but rather it needs power urgently. These are two entirely different matters; for example, the first could be supported by the expansion of nuclear power and the electrical grid, but the latter cannot. In fact, combining these two is something very few companies can do. 

Get $275 off Advanced Market Signals for our 6-year anniversary sale. Access real-time trade alerts, weekly webinars, and 100+ research articles on the AI trade per year. Claim OfferGet $275 off Advanced Market Signals for our 6-year anniversary sale. Access real-time trade alerts, weekly webinars, and 100+ research articles on the AI trade per year. Claim OfferClaim Offer

Behind‑the‑Meter Fuel Cells vs Grid‑Dependent Power 

Bloom Energy offers onsite power generation through solid oxide fuel cells that are behind the meter to reduce dependency on the grid.  By providing behind-the-meter generation, Bloom reduces reliance on utility infrastructure and accelerates time-to-power for customers. An added benefit is the United States is the largest producer of natural gas, therefore, Bloom does not struggle to secure supply given the United States has large, well-developed gas supplies and pipeline infrastructure. 

Our primary message has been “time to power” for Bloom, and the company continues to stand out for this very reason as it is finding strong product market fit in AI data center power needs. This is a key advantage as on-site power is becoming more of a necessity as grid constraints and connection timelines rise.   

Regarding grid constraints, PJM has already fallen short of reliability requirements in its last two capacity auctions, including a roughly 6.6 GW shortfall for the 2027/28 planning year, while ERCOT’s interconnection queue has surged to about 226 GW, including roughly 165 GW from data center projects targeting approval by 2030. Against that demand, ERCOT added only 23 GW of new capacity in 2024–25, underscoring why time-to-power is becoming a central bottleneck for AI data centers. 

This further validates Bloom’s positioning by enabling new data center projects to come online sooner with on-site, behind the meter power without sitting in interconnection queues for years at a time. Bloom has already proven that it can quickly establish data center power solutions in a rapid manner, completing shipments to Oracle Cloud Infrastructure in just 55 days of its 90-day delivery request.   

Its fuel cells are also fuel-flexible and can run on natural gas, biogas, or hydrogen, and provide continuous power with 99.9-99.999% reliability metrics. They are also modular in nature and can scale from 20 MW to 500 MW+, allowing flexibility in deployments and ease of scaling. Bloom is also continuously improving on price-performance, stating that its fuel cells have seen double digit YoY cost reductions each year for the past ten years, and a 10X increase in power production in the same footprint versus ten years ago.   

Oracle’s Project Jupiter Sends Bloom’s Stock Soaring in April 

We’ve covered previously that Bloom delivered a fuel cell system to Oracle in 55 days, standing out among the longer to deploy solutions in the market.  In April, Bloom Energy announced an expansion with Oracle for a total of 2.8GW of fuel cell capacity with 1.2GWs shipping now.  

Following the capacity announcement, Oracle announced Project Jupiter yesterday stating the company will utilize up to 2.45GWs “to fully power the AI data center campus” located in New Mexico. This is an important development as it means the AI data center will not use gas turbines and the diesel generators as originally planned. According to the press release, nitrous oxide emissions will be cut by 92% compared to the previous gas turbine plan.   

The following was stated about the new deal: “It will be 100% Bloom. When completed, it will be one of the largest islanded microgrid power facilities in the world. Oracle pivoted to Bloom only solution for 2 main reasons: first, be a responsible corporate citizen and partner by being responsive to resident concerns about air quality, water use, noise and increasing electricity rates.   

Second, to stand up their grid independent and clean AI factory with even greater reliability and speed. Bloom is the cleanest commercially available on-site power generation option for such data centers and the most water efficient. Even Blooms community-friendly attributes, Oracle should be able to energize the campus materially faster than any other available alternative solution in the market.” 

The new deal with Oracle has sent Bloom’s stock soaring on a 1-month basis (on top of the already strong 1-year returns). We’ve discussed in-depth the product market fit for the stock as being “time to power,” yet the company’s value proposition has actually improved since last quarter as the Jupiter deal serves as an important proof of concept for the company.  

My understanding, as an analyst how has tracked this stock longer than any research site on record, is the Jupiter deal will mark the first time an AI data center will be powered entirely by Boom Energy’s solutions – an important moment in Bloom’s history. 

At minute 5:43, Beth Kindig discusses why Bloom Energy is her top stock pick for 2026. 

Why AI Inference Will Drive the Next Wave of Power Demand 

The inference market will require more gigawatts than training, yet an additional constraint is location (or geography). Inference sits at the edge, close to users for latency, which means demand will be coming from dense metros instead of less populated areas, such as where training data centers are located (rural areas). 

The underlying trend is significant. Boston Consulting Group projects inference power demand to grow at a 122% CAGR through 2028, compared to 30% for training. McKinsey forecasts inference rising from ~21 GW today to ~91 GW by 2030 at a 35% CAGR, accounting for the majority of AI's incremental power draw through the end of the decade. As frontier models stabilize and the AI economy shifts toward serving users at scale, inference will overtake training as the dominant compute workload, which translates to inference also consuming more power. 

Bloom’s Advantage in Dense Metro Deployments 

Bloom could see more demand from the inference market compared to training as it offers a combination of low emissions, minimal water use and a small footprint – which is ideal for dense areas. Its solid oxide fuel cells (SOFCs) are electrochemical rather than combustion-based, producing minimal emissions. They are also modular, scaling from 20 MW to 500+ MW, and operate quietly enough to be placed in urban environments. Lastly, they run on natural gas pipelines that already exist in metros, bypassing grid interconnection queues entirely. For an inference-driven AI buildout, these details matter. 

Notably, the newly appointed CFO is a signal that Bloom is positioning for this shift toward inference, as Simon Edwards, was the former CEO of Groq, a leading developer of inference infrastructure and LPUs (recently acquired by Nvidia). Edwards is likely a deliberate, strategic hire ahead of the inference scale-up. 

Q1 2026 Revenue Surged 130% YoY — Strongest Growth in Bloom Energy’s History 

Bloom Energy reported Q1 2026 revenue of $751.1 million, up 130.4% YoY and beating estimates by 39.1% — the strongest growth in the company's public history. Product revenue reached $653.4 million, up 208.4% YoY, and now represents ~87% of total revenue. 

Management raised full-year 2026 guidance to $3.4 billion–$3.8 billion, implying 77.9% YoY growth at the midpoint, up from $3.1–$3.3 billion. 

Gross and Operating Margins Expanded Sharply YoY 

GAAP gross margin was 30%, up 280 basis points YoY. Adjusted gross margin was 31.5%, up from 28.7% in Q1 2025. Management raised full-year 2026 adjusted gross margin guidance to 34%, up from 32%. 

GAAP operating margin was 9.6%, up from (5.8%) in Q1 2025. Adjusted operating margin was 17.3%, up from 4%. Management raised FY2026 adjusted operating income guidance to $675 million at the midpoint, up from $450 million. 

Bloom Energy EPS Crushed Estimates by 242%  

Adjusted EPS was $0.44, crushing estimates of $0.13 by 242.4%. GAAP EPS was $0.23 versus estimates of ($0.02). Management raised full-year 2026 adjusted EPS guidance to $2.05 at the midpoint, implying 169.7% YoY growth, up from $1.405. 

Positive Cash Flow Inflection 

Bloom reported the first positive Q1 operating cash flow in company history at $73.6 million (9.8% of revenue), versus an outflow of ($110.8 million) a year ago. Free cash flow was $47.4 million. Cash of $2.49 billion and debt of $2.60 billion at quarter-end. 

Conclusion: 

Bloom Energy’s April rally validated what we have been writing since June 2024, which is that power is the leading constraint on the AI buildout. Of this, time-to-power is the variable that matters most, and Bloom is one company that can deliver mission critical power solutions for the incoming inference market. 

The harder question, and one that matters more than entering Bloom at $17 in April 2025, is what comes next. The investors who outperform from here will not be the ones who pile into the trade that already worked. They will be the ones positioned for the next bottleneck long before the market sees it. 

Get $275 Off on our Advanced Plan. April was the Nasdaq-100’s best month in six years, and few portfolios participated like I/O Fund. Bloom Energy was the top-performing large-cap stock during the April rally, yet I/O Fund also held four of the top 10 large-cap stocks at allocations of 7% or higher. 

Since May 2020, our audited portfolio has returned 326% cumulatively. Based on our published comparisons, that would place I/O Fund #1 versus hedge funds and #3 versus tech ETFs or mutual funds — before including our 48% YTD return in 2026. 

Join during our 6-year anniversary sale and save $275 on our Advanced Plan 🥂

Members receive real-time trade alerts, I/O Fund’s portfolio, 100+ research articles on the AI trade per year, and weekly one-hour webinars. 

Offer ends soon.
Claim OfferClaim Offer

Please note: The I/O Fund conducts research and draws conclusions for the company’s portfolio. We then share that information with our readers and offer real-time trade notifications. This is not a guarantee of a stock’s performance and it is not financial advice. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis. Beth Kindig and the I/O Fund own shares in BE at the time of writing and may own stocks pictured in the charts.

Recommended Reading:

• Inside Nvidia’s $4B Optical Strategy—and Why CPO Changes Everything
• Is Nvidia Stock a Buy? Why Semiconductor Strength May Signal a Market Top
• Nvidia’s $20 Trillion Thesis Is Intact. My 2026 Allocation Isn't
• Bitcoin 2026 Price Prediction: Why the Dollar, Global Liquidity and Volume Signal More Downside Ahead

• Over the coming years, CPO is poised to see a dramatic uptick in demand, as data center operators push to expand the limits of AI. 
• CPO provides key benefits over the two networking systems that dominate today: copper and optical transceivers. This includes up to 5x power efficiency versus transceivers and much higher bandwidth. 
• Nvidia and Broadcom are huge players in CPO, and firms that gain qualification in their supply chains can be massive beneficiaries

Nvidia’s Rapid Networking Roadmap Is a Key Driver for AI Stocks

Within the AI investment theme, there is nowhere that the supply chain shifts faster than in networking, leading companies to gain content on new platforms or lose incremental share. 

The reason is straightforward: much of the market is tied to a single customer, Nvidia; and Nvidia is rolling out new architectural iterations at an unusually fast pace. When it comes to networking, two of the most important architectural advancements are the increase in pod and cluster sizes and the transition to 200G per lane. 

Last month, Nvidia made $2 billion equity investments in two separate optical component suppliers: Coherent and Lumentum. Nvidia is securing its supply chain as it ramps its co-packaged optics (CPO) roadmap and writing big checks to do so. These targeted moves signal that CPO, the next major architectural shift in AI networking, is moving from theory to reality. 

Below, we break down why this transition is taking place and the key companies involved in the secular trend toward CPO. 

Nvidia’s Move to Larger Pods: Scale-Up from NVL72 to NVL576 and Beyond 

With Blackwell and Blackwell Ultra, Nvidia was fundamentally focused on solving scale-up problems, where the primary challenge is binding large numbers of GPUs into a single coherent node with a unified memory using ultra-dense, low-latency NVLink fabrics. This led to NVL72, which packed 72 GPUs into one rack, acting as one giant GPU. 

However, with Rubin and Rubin Ultra, the company is pushing this concept further. Nvidia will offer Rubin in NVL72, NVL144, and NVL576 configurations, connecting two and eight racks respectively into a single NVLink scale-up domain. With NVL576, an eight-rack pod behaves as a single, massively larger GPU. 

Rubin also doubles NVLink scale-up bandwidth versus Blackwell — 3.6 TB/s of bidirectional GPU-to-GPU bandwidth on the sixth-generation NVLink 6 interconnect, with 36 switches per NVL72 rack delivering 260 TB/s of total bandwidth versus Blackwell's 130 TB/s. 

As these pods grow in size and require higher bandwidth, copper hits physical limits. Each step up in bandwidth degrades signal integrity faster, shortening the effective length of copper cables. A chart from Marvell illustrates this. At 100G per lane, the speed that now dominates deployments, copper can stretch around 5 meters using range-extending AECs. At 200G, the speed that will be used in Rubin Ultra, the effective length of AECs falls to just 3 meters. 

Comparison of passive direct attach copper (DAC) and active electrical cable (AEC) reach at increasing lane speeds. As bandwidth scales from 50G to 200G per lane, copper cable length degrades significantly, with AEC reach falling to roughly 3 meters at 200G. Source: Marvell estimates.

According to Supermicro, one GB300 NVL72 rack is 0.6 meters wide. Rubin Ultra NVL576 will place eight racks side by side, resulting in a width of nearly 5 meters; too long to connect the entire pod at 200G using AECs. 

In turn, Nvidia will use CPO for rack-to-rack connections in Rubin Ultra NVL576, although copper will still be used for connections within each rack. This is why Huang said that customers will be able to buy Rubin Ultra in “copper, or copper plus CPO." Copper plus CPO will be used in NVL576, while only copper will be used in smaller configurations. Huang went on to say, “two years from now, at [NVL]1152, it's all CPO because there's a limit to how far it could take copper.” 

Co‑Packaged Optics as a Structural Shift for Nvidia's Stock

With pod sizes and bandwidth only increasing, the transition from copper to optics in scale-up is structural, not cyclical. CPO is positioned as the eventual endpoint of that transition.

Notably, companies in the supply chain are moving to reflect this. Credo recently acquired DustPhotonics to diversify away from AECs, the product the company has built its name on. Through this deal, Credo adds silicon photonics to its portfolio, with the company expecting to generate $500 million in optical revenue in FY2027. For reference, Credo reported its Q3 FY2026 results in March. This will aid the company in bridging the gap between AEC content and optics content. 

Marvell acquired Celestial AI as it looks to offer CPO solutions. During its Q4 FY2026 results in March, Marvell projected its CPO revenue reaching a $500 million annualized run rate in Q4 FY2028 before doubling to $1 billion by Q4 FY2029. Nvidia and Marvell also recently announced a strategic partnership, connecting Marvell to Nvidia’s AI factory ecosystem through NVLink Fusion. Customers can easily pair Marvell products, including custom XPUs, certain scale-up networking, and silicon photonics, with Nvidia’s rack-scale AI compute and other components using NVLink. Additionally, Nvidia has invested $2 billion in Marvell. 

Scale-Out CPO: Boosting Performance and Efficiency Versus Transceivers 

Scale-out networking poses a different challenge for CPO adoption, as companies look to connect larger pods into massive clusters with 1 million AI accelerators. Copper has already been largely phased out of scale-out, as distances are far too long. This has led to optical transceivers becoming a key solution.  

Optical transceivers take electrical signals sent through copper traces in ASIC switches and convert them into optical signals. These signals then flow through fiber optic cables, which can stretch kilometers at high bandwidths without losing integrity. 

However, using optical transceivers also comes with significant drawbacks. Most notably, they consume much more power than copper and are more expensive. This is the trade-off that data center operators are increasingly having to accept in exchange for longer cable lengths and/or higher bandwidth.

mid

CPO offers something closer to a best-of-both-worlds solution, allowing for both long cable lengths as well as better power efficiency, higher bandwidth, and lower latency compared to transceivers. CPO provides better power efficiency by drastically shortening the distance signals flow through copper before conversion to light.  

In most cases, CPO eliminates the need for power-hungry DSPs, which clean up the degraded electrical signal before sending them to transceivers. This comes as CPO embeds optical engines in the same package as the switch. A visual from Nvidia illustrates this difference clearly. The orange line (copper) is much shorter in the CPO diagram, and the DSP is gone, allowing the electrical signal loss to be significantly lower.  

Comparison of a traditional pluggable switch architecture and Nvidia’s co‑packaged silicon photonics design. In pluggable systems, electrical signals travel across the PCB, connectors, and port cage before reaching an external transceiver, resulting in roughly 22 dB of signal loss and requiring DSPs and multiple lasers. Co‑packaged optics integrate silicon photonics alongside the switch ASIC, shortening the electrical path to the substrate, reducing loss to about 4 dB, and improving power efficiency at 1.6 Tb/s. Source: Nvidia

CPO also offers higher bandwidth and lower latency versus pluggables. As inference workloads rise, largely driven by agentic AI, improving these variables is key. Automating workflows in enterprise environments means higher data rate requirements compared to the use of chatbots. 

LLM developers will compete on how fast their models can execute tasks, making latency reduction paramount. Reducing latency is particularly relevant going forward, as many expect inference to overtake training as the dominant AI workload over the coming years.  

McKinsey projects that by 2030, inference will account for 93 GW of data center demand, versus 62 GW for training. It sees inference demand rising by a CAGR of 35% through 2030, significantly faster than training’s 22% CAGR. 

In summary, as AI workloads continue expanding, power efficiency, bandwidth, and latency improvements are vital to increasing performance while limiting costs. CPO is a key solution that allows for these advancements. 

CPO Adoption: Gated by Low-Cost Copper and Reliability Concerns Near Term 

Despite these benefits, CPO faces constraints that limit its adoption today. Copper and optical transceivers are generally sufficient at today’s bandwidth levels and cost less than CPO upfront. With hyperscalers already spending hundreds of billions on AI infrastructure annually, staying on lower-cost solutions makes more sense for now. In line with this, Broadcom CEO Hock Tan said that the industry will “try to scale up within a rack in copper as long as possible.” Echoing this, Jensen Huang said, "We should scale with copper [as far as] we can, as long as we can." 

CPO reliability is another hurdle that developers are tackling. Theoretically, CPO should be more reliable than pluggables, as it consolidates many otherwise separate parts, creating fewer points of failure. However, because CPO has not been deployed at scale, there is a lack of real-world evidence to support this idea. 

This is key, as when a CPO chip fails, servicing costs are much higher. Pluggable transceivers can be easily swapped out when they fail, but this is not possible when optical engines are embedded in the switch package. CPO servicing requires removing the full switch to have a complex repair performed or replacing it entirely. 

To accelerate adoption, CPO providers must demonstrate strong reliability of the technology. On this front, Broadcom recently made a significant step forward. In a study conducted with Meta, the company showed a 5X improvement in serviceable failures compared to pluggables. The study also found no unserviceable CPO failures after 15 million hours of device testing. This provides solid initial evidence of CPO reliability. 

Still, these tests were performed in a lab environment, not in actual data centers. This underscores the need for more CPO reliability testing in real-world environments before adoption hits an inflection point. The industry has an opportunity to generate this data through early CPO deployments in 2026 and 2027, setting the stage for increased adoption thereafter. 

Nvidia and Broadcom Are Leading the Push Into CPO Networking

Nvidia and Broadcom are the two market leaders in CPO, as both are leaders in switching ASICs. Nvidia has the largest networking business in the world, with revenue hitting $11 billion. Meanwhile, one-third of Broadcom’s $10.7 billion in total AI revenue, or approximately $3.6 billion, came from networking last quarter. 

Broadcom has been developing CPO since 2021 and is now shipping its third-generation scale-out product, the Tomahawk 6 – Davisson switch, which delivers 3.5x better power efficiency than pluggables. Broadcom is currently developing its fourth-generation CPO product, which will double the per-channel bandwidth compared to Davisson. 

Meanwhile, Nvidia will use CPO for scale-up NVL576 approximately a year from now. For scale-out networking, Nvidia has its Spectrum-X Ethernet Photonics switch, which it says will deliver 10X greater network resiliency with CPO, bringing 1.6T silicon photonics (SiPho) optical engines directly onto the switch. 

Maximum bandwidth doubled to 102.4Tb/s per ASIC, matching Broadcom’s Davisson, though Nvidia is also offering the industry’s first four-ASIC design, delivering 409.6Tb/s bandwidth. Notably, Spectrum-X Ethernet switches drive up to 5X better power efficiency with a lower cost versus pluggable transceivers. 

CPO adoption should bring substantial benefits to Nvidia and Broadcom. However, companies that gain qualification within Nvidia and Broadcom’s CPO supply chains are poised to be among the biggest winners from this networking shift. I/O Fund specializes in identifying these types of lesser-known networking players. 

Subscribe to Advanced Market Signals to receive I/O Fund’s best thinking on the stocks with massive potential as the shift to CPO progresses, including real-time trade alerts for every entry and exit, portfolio access, and weekly live webinar access. Sign up hereAdvanced Market Signals to receive I/O Fund’s best thinking on the stocks with massive potential as the shift to CPO progresses, including real-time trade alerts for every entry and exit, portfolio access, and weekly live webinar access. Sign up here

Optical and CPO Stocks: Coherent, Lumentum, and Corning

Coherent, Lumentum, and Corning are optical and CPO beneficiaries to be aware of. All three are benefiting from scale-out transceiver adoption today and are positioned to benefit from CPO gradually replacing copper in scale-up over the coming years. Below, we break down what each company supplies, their opportunity ahead, and how the market is valuing them today. 

What Each Company Supplies in the CPO and Optical Ecosystem

Coherent and Lumentum: Lasers, Silicon Photonics, and Nvidia’s CPO Supply Chain

Coherent and Lumentum make pluggable optical transceivers and high-powered lasers, critical components within transceivers. While CPO will replace transceivers in certain instances, it also drives higher content for the SiPho-laser ecosystem and CPO photonics components, as SiPho will serve as the backbone for the CPO switches. This extends beyond the photonics ICs to include CW lasers and ultra-high-power (UHP) lasers for external light source (ELS) modules.  

Coherent and Lumentum expect to be leading suppliers of these components within Nvidia's CPO rollout. Nvidia has rubber-stamped its supply chain relationship with both firms, investing $2 billion in each to fund manufacturing capacity expansions.

Corning’s Role in CPO

Corning plays a different role in the ecosystem as one of the top fiber optic cable makers. CPO adoption will translate into much more fiber optic cable usage in data centers. According to Marvell, this increase will be very significant. They say CPO will enable "tens of thousands of fiber per rack, no longer just a few thousand." Marvell believes the increase in fiber usage will be so large that the industry must create new innovations to manage it.

Optical Demand Is Inflecting Across AI Data Centers

Demand is already inflecting for these companies. Lumentum's revenue rose by over 65% YOY to $665.5 million in its latest quarter, and adjusted operating margin expanded by 1,730 basis points. Lumentum expects growth to accelerate to around 90% YOY next quarter and an approximately 500 basis point sequential operating margin expansion.  

Coherent saw revenues rise by 34% YOY in its data center and communications segment last quarter, driven by growth in 800G and 1.6TB transceivers. The company's data center book-to-bill ratio exceeded 4X, showing how dramatically demand is outstripping supply. 

Meanwhile, Corning's Enterprise business, which captures sales inside data centers, grew 61% YOY in 2025 to $3 billion, with the hyperscale data center portion growing significantly faster. 

Optical and CPO Market Outlook Through 2030

The market ahead of these firms is substantial. Corning has made very strong statements around its opportunity to benefit from scale-up CPO adoption. The firm believes that its scale-up CPO opportunity is at least 2-3X larger than its Enterprise business, implying an incremental opportunity of $6 billion to $9 billion. Management believes it could be even larger as it spends more time with partners in the ecosystem. Compared to Corning's 2025 core sales of $16.41 billion, this incremental market is very significant. 

Coherent estimates that its serviceable addressable market (SAM) in CPO will be more than $15 billion by 2030. This compares to Coherent's LTM revenue of $6.29 billion. Notably, SAM estimates represent just the portion of the total addressable market (TAM) that a company believes it can realistically serve.  

Related to this, Lumentum estimates that its current optical AI TAM is $18 billion today. It sees this figure increasing by more than 5X to over $90 billion in 2030. These forecasts help illustrate the huge opportunity that exists for smaller players in the optical and CPO market.

CPO Shipments Are Set to Gain Share

Importantly for Coherent and Lumentum, TrendForce estimates that both transceiver and CPO shipments will rise greatly over the coming years, although CPO will increasingly take share. Forecasts show optical transceiver shipments continuing to rise from around 50 million in 2026 to nearly 200 million by 2030. Simultaneously, CPO shipments exceed 50 million by 2030, and increase their penetration rate within optical networking from less than 1% to more than 35%.

TrendForce forecast for CPO penetration in AI data centers from 2025 through 2030. Total optical shipments continue to rise, while co‑packaged optics scale rapidly from negligible adoption in 2025 to more than 35% penetration by 2030. The data highlights a structural shift toward CPO as data center bandwidth and power efficiency requirements increase. Source: TrendForce, March 2026.

Valuation 

How Much CPO Upside Is Already Priced Into Networking Stocks

The market has already moved to reflect much of the optical and CPO prospects for these stocks. All three are trading at or very close to their all-time high forward P/E ratios, with these multiples being 2.2-2.6X higher than their median levels over the past three years. Since the end of June 2025, Lumentum has delivered a return of 950%, while returns exceed 280% and 210% at Coherent and Corning, respectively. 

Chart showing the current and three-year median forward P/E ratios of Coherent, Lumentum, and Corning. Coherent’s forward P/E is 54.2x versus a median of 24.7x, Lumentum’s forward P/E is 73.9x versus a median of 33.1x, and Corning’s forward P/E is 52.4x versus a median of 20.1x. Source: Koyfin

Conclusion

CPO Is a Multi‑Year Structural Tailwind for AI Infrastructure 

The AI networking stack is moving secularly towards optics and away from copper. Nvidia's pod-scaling roadmap clearly demonstrates this. From NVL72 in Blackwell today to NVL576 with Rubin Ultra, NVL1152 with Feynman, and potentially beyond, the physical limits of copper cannot be engineered around. With CPO emerging as the preferred optical form factor, adoption will continue to increase. 

How quickly the transition moves is up for debate. Broadcom and Nvidia management teams have both said that copper will be used for scale-up as long as possible, and TrendForce estimates that CPO penetration will remain in the low single digits through 2027 before inflecting. Reliability validation in real-world deployments over the next 12-24 months will be a key factor in determining the pace, while there are bridge solutions, such as NPO and LPO. 

Regardless, the supply chain is already shifting. Component makers and switch vendors are positioning for a networking stack that looks considerably different from today's. For investors, the CPO transition is a clear multi-year theme in AI infrastructure, with implications that extend well beyond the handful of names the market is focused on. 

My updated Q2 Top 15 AI Stocks report was just released. The report runs over 70 pages and identifies the 15 stocks I believe will lead the AI market this quarter. The report is built on the same investment discipline that identified massive winners like Bloom Energy and Lumentum early in their AI cycle.

Since our inception in May 2020, the I/O Fund has delivered a cumulative return of 326% — outperforming the Nasdaq-100 by 152 percentage points. Subscribers receive the Top 15 AI Stocks report, real-time trade alerts, full portfolio access, and weekly one-hour webinars. Sign up now.

Please note: The I/O Fund conducts research and draws conclusions for the company’s portfolio. We then share that information with our readers and offer real-time trade notifications. This is not a guarantee of a stock’s performance and it is not financial advice. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis. Beth Kindig and the I/O Fund own shares in NVDA at the time of writing and may own stocks pictured in the charts.

Recommended Reading:

• Is Nvidia Stock a Buy? Why Semiconductor Strength May Signal a Market Top
• Nvidia’s $20 Trillion Thesis Is Intact. My 2026 Allocation Isn't
• Bitcoin 2026 Price Prediction: Why the Dollar, Global Liquidity and Volume Signal More Downside Ahead
• 2026 Stock Market Outlook: Cycle Convergence & What's Next

Recently, I've reiterated my $20 trillion market cap thesis for Nvidia, which implies upside of roughly 310% over the next four years. However, my thesis does not hinge on Nvidia reaching that milestone through hardware alone. Instead, the thesis hinges on software advancements and the recurring revenue that will inevitably come from Nvidia's lead in robotics and simulation. I have emphasized the growing importance of Nvidia's software business relative to hardware since 2023. 

The distinction cuts both ways. By arguing that software is central to the $20 trillion thesis, I am also implying that Nvidia's hardware moat becomes less effective over time. Seven years ago, when Nvidia was still a roughly $100 billion company trading near $3.15 on a split-adjusted basis, my original thesis for why it could become the world's most valuable company centered on the CUDA moat. At the time, I wrote: "Developers will self-regulate the number of competitors for processing units due to a need for a universal platform that supports all frameworks." 

My firm, the I/O Fund, has held Nvidia through the full seven-year journey, sometimes at an allocation as high as 20%, through both remarkable upside and equally remarkable downside (may be hard to remember, but the stock was down 60% in 2022 when I publicly defended the stock). 

The thesis on Nvidia's hardware moat has played out exceptionally well, but that also highlights one of the biggest risks investors face, which is becoming emotionally attached to a winning stock. While I still believe Nvidia will reach $20 trillion by 2030, I believe much of that 310% return is likely to be back-half weighted in the years of 2028-2030. This is what separates investors from AI enthusiasts. While an AI enthusiast can sit back, relax and discuss specifications and other fandom, an investor must always answer — is my capital better deployed elsewhere? 

Is Nvidia Stock Still the Best AI Stock in 2026?

So far in 2026, answering the question of where to deploy capital has not been easy to answer. Nvidia's stock only recently turned positive; the QQQs are barely positive this year, as is the same for many tech-related ETFs such as IVES, GRNY and ARKK.  

In sharp contrast, the I/O Fund is up roughly 33% year-to-date, reflecting a willingness to follow the opportunities as they shift across the AI landscape. We count recent winners such as Bloom Energy, which is up 1100% since our initial entry, an optical networking name we highlighted ahead of its 2026 surge, now up nearly 300% YTD and 650% since our lowest entry in November. Plus, a photonics position we doubled down on in January with a 10% allocation that has since gained more than 130% year-to-date. 

The same framework that surfaced those opportunities is what tells me Nvidia's 2026 setup may no longer be as rewarding as what I can find elsewhere. The analytical case comes down to three things: the CUDA moat matters less with inference, custom silicon is gaining market share, and the delay in Rubin creates uncertainty at exactly the wrong moment.

On the flip side, the valuation is lower than its historic average, and in a volatile market, Nvidia could still stand out simply by continuing to post stronger earnings growth than most of large-cap tech. The company will remain the dominant system-level player in AI, and the CUDA moat will certainly not vanish overnight. 

The debate, in my view, is not about whether Nvidia stays important. It is about whether the return profile is still as compelling as what can be found elsewhere in the AI trade.  

CUDA Matters Less as AI Inference Takes Over

In 2018, my original thesis on why Nvidia can become the world's most valuable company was centered on the moat the CUDA platform provides when I stated: "Nvidia is already the universal platform for development, but this won't become obvious until innovation in artificial intelligence matures. Developers are programming the future of artificial intelligence applications on Nvidia because GPUs are easier and more flexible than customized TPU chips from Google or FGPA chips used by Microsoft […] When artificial intelligence matures, you can expect data center revenue to be Nvidia's top revenue segment. Despite the corrections we've seen in the technology sector, and with Nvidia stock specifically, investors who remain patient will have a sizeable return in the future."  

At the time, Nvidia's data center revenue was 1/6th of Intel's — whereas today, the AI juggernaut reported $194 billion in data center revenue compared to Intel's $17 billion. Although you could pontificate on the many defensible design elements of Nvidia's AI systems, one way to simply describe this historic ascent is that the mature libraries and frameworks from CUDA makes it hard for an engineer to go anywhere else. Notably, that's not a regurgitated thesis, but rather my thesis on the stock implications of the CUDA moat pre-dated the Street and AI experts by many years. by many years. That’s important because I am shelving that thesis as the inference market approaches. 

There is an incoming shift to my original investment thesis from 2018. 

Programming GPUs with the CUDA platform is primarily a training exercise as this is the phase where engineers are experimenting and need the developer ecosystem, including extensive tools like cuDNN, NCCL, debugging, custom kernel support, and CUDA's massive libraries. The ecosystem has been built for over 20 years, has over 6 million developers contributing and every ML framework is first optimized for CUDA. The switching costs today remain extraordinarily high for engineers. 

To contrast, inference is repetitive to where once a model is trained, the model is running millions of times per day. Serving platforms and inference frameworks like vLLM and TensorRT-LLM reduce dependency to develop on a specific software platform, like Nvidia's CUDA. 

Training a frontier model is a one-time, multi-month event. Inference, by contrast, is the revenue-generating phase. Every ChatGPT query, every Copilot suggestion, every Waymo autonomy decision is inference. As frontier labs reach the limits of practical model size and enterprise AI adoption scales, inference workloads are projected to grow several times faster than training workloads through the rest of the decade. The segment where CUDA's moat is strongest is becoming a smaller share of total compute, while the segment where it is weakest is becoming the larger share.

mid

There is also more of a push toward open standards for the inference phase to reduce dependency on hardware specific code for serving paths, as tools like ONNX runtime, vLLM and the compiler Triton help to export models (or compile them) to be run agnostically on any AI accelerator. 

In response to CUDA's moat weakening in the inference phase, Nvidia has pushed for their inference stack to remain proprietary by offering inference optimization software called TensorRT-LLM. TensorRT-LLM analyzes and optimizes LLMs to improve performance by fusing multiple operations into a single GPU kernel, selecting the optimal precision and optimizing memory usage for the key-value cache. Overall, Nvidia states this leads to 5X faster model performance for inference. 

However, something to consider is that Nvidia is needing to make this new attempt at preserving its ecosystem as the CUDA empire will not neatly hold as the inference market plays out. The open-source market is growing to become a serious contender to proprietary optimization software like TensorRT-LLM, as alternatives that are more community driven are available and accomplish something similar, such as vLLM and SGLang. Both have moved from research-project status to production deployment at major AI operators, with vLLM in particular now powering inference at some of the largest LLM serving workloads outside of the hyperscalers themselves. Furthermore, large inference players like Cloudflare can build their own custom engines. 

The point is not to be an alarmist, but rather to note when the piece most central to my original thesis is shifting. CUDA will remain the most popular software development platform in AI by a wide margin, however, the freedom to go elsewhere is something Nvidia has not contended with at this level. 

Custom Silicon is Undeniably Increasing in Market Share 

A few months back, the market had a brief scare around what Google's TPU v7 Ironwood might mean for Nvidia's grip on AI compute. The concern was not simply that Google had built another custom chip, but that Ironwood was introduced as the first TPU designed specifically for inference. 

At the time, Google emphasized better power efficiency and stronger "intelligence per dollar" for serving workloads. Ironwood scales up to 9,216 chips, delivers 42.5 exaflops in its largest pod, and Google has paired it with software support such as vLLM on TPU, reinforcing the idea that inference is becoming a more open and cost-sensitive market than training. We covered this more in the write-up: "This AI Stock is Set to Surge from Inference Demand." 

Although Ironwood v7 offers major headway in narrowing the performance gap with Nvidia on inference workloads, the reality is that custom silicon programs require long development cycles. Designing the chip is only the initial stages, and from there, hyperscalers need to optimize the compiler stack, optimize frameworks and also validate performance at scale. The result is a far slower product road map that typically lags Nvidia's current generation of GPUs. This lag puts additional emphasis on Nvidia delivering on time. 

Why the Advantages of Custom Silicon Outweigh Development Timelines

Nvidia's data center GPUs carry gross margins above 70%. For companies spending $50-100 billion annually on AI infrastructure, the savings from moving even 20-30% of inference workloads to in-house silicon compounds into tens of billions of dollars per year. That math is driving Google and Amazon to accept slower product cycles in exchange for architectural independence. It is also the math incentivizing Meta and Microsoft to follow suit. Perhaps most importantly, the inference market will offer a catalyst for custom silicon compared to training because workloads are more specific, and cost savings can be achieved at massive volumes. 

Below is what a few industry analysts are predicting. Although I believe these are aggressive, they help to illustrate the challenges in front of Nvidia. 

Counterpoint Research believes that by 2028, custom silicon will cross the 15-million mark to surpass GPU shipments as the top 10 hyperscalers will have deployed 40 million AI server compute ASIC chips cumulatively during 2024-2028, stating: 

"What is also supporting this unprecedented demand is AI hyperscalers building significant rack-scale AI infrastructure based on their in-house stacks, such as Google TPU Pods and AWS Trainium UltraClusters, enabling them to operate as one supercomputer." 

TrendForce is the most aggressive forecast, stating GPU-based AI servers will account for 69.7% of shipments in 2026 with ASIC-based servers rising to 27.8%. This doesn't account for GPU market share from AMD, which if you put that at 10%, would result in Nvidia's market share being 59.7%. 

With the information that I have today, these forecasts could be too aggressive. 

According to Broadcom, they'll see $100B in AI chip revenue in 2027 and we've modeled another $50B in networking. If we allocate $45B base case to AMD and go with what we know of Nvidia's stated trajectory to $1 trillion in revenue, then the split looks something more like this for 2027: 

• NVDA $500B 
• AVGO $150B to $200B (assuming mgmt team was being conservative we will use the $200B number) 
• AMD $45B 
• Total among top 3 silicon providers: $745B with NVDA at 67% market share versus the 59.7% implied above 

However, one data point that complicates things is MediaTek could see 150,000 CoWoS wafers in capacity in 2027, compared to 20,000 in 2026. Thus, the landscape is evolving in terms of the number of competitors. 

Notably, the level of erosion may be up for debate, but the most probable outcomes do not favor Nvidia continuing to dominate AI accelerator sales at the level it has in the past. In training, Nvidia represented 90% of workloads. 

There are many moving parts, but if we do assume that Nvidia sees 70% of market share, down from 90% previously, and capex grows at 60% year-over-year, then Nvidia's growth rate would be 24%.  

Here's what a sensitivity analysis looks like:

Pictured above: Nvidia revenue growth sensitivity analysis assuming 60% annual AI capital expenditure growth and varying Nvidia GPU market share in 2027. Under consensus estimates from TrendForce and Counterpoint, Nvidia’s GPU share declines to roughly 69%, or about 63% after accounting for AMD capturing 6% of the GPU market, implying a revenue growth rate of approximately 15.6%.

For the calendar year ending in January 2028, analyst estimates are at 30.1% growth. Note the numbers in the sensitivity analysis are for compute only, and does not include networking, which is growing rapidly and estimated at roughly 160% in the upcoming quarter. 

The Bull Case Hinges on Valuation 

Even with the supporting data above, I have kept a ~5% position this year in Nvidia as the growth profile combined with earnings profile is hard to beat across most tech stocks. The company is expected to see >50% growth on both the top line and the bottom line this year. This growth combined with flat price action for about a year has led to an attractive valuation. 

Pictured Above: Nvidia stock trades at a P/E ratio of 40.7 compared to the 3-year median of 55.29. Nvidia is currently trading 26% lower than the median.

Source: YChartsYCharts

Going back to my introduction, the question for a portfolio manager isn't whether Nvidia is fairly valued today. It's whether the capital compounds faster in Nvidia’s stock over the next twelve months than in the many alternatives we've identified.  

I just dropped my Top 15 List of AI Stocks — this list ranks the companies I believe will define the next year and whose fundamentals are on fire. The 70-page report is for premium Pro and Advanced members, sign up here.Top 15 List of AI Stocks — this list ranks the companies I believe will define the next year and whose fundamentals are on fire. The 70-page report is for premium Pro and Advanced members, sign up heresign up here.

Rubin Delay and HBM4: A New Risk for Nvidia Stock in 2026

In a previous free newsletter, I had stated Nvidia’s product road map is the second line of defense should the CUDA moat be breached. What happens when both are breached? That is not a scenario that I originally modeled for. 

The reported one-quarter delay on Rubin is terrible timing, to be frank, as it coincides with the timing improving for when custom silicon becomes more attractive for Big Tech (which is ultimately aligned with the incoming inference market). The delay in Rubin not only allows custom silicon one more quarter to catch up, but it also makes for a strong case for having back-up orders across Broadcom, Mediatek and/or AMD for supply chain diversification.  

HBM4 validation times have been cited as one key factor behind the delays for Nvidia’s upcoming Vera Rubin generation – we have seen in the past that these qualification tests can extend as long as 18 months, such as in Samsung’s case with HBM3e. Currently, reports suggest this HBM4-related delay could persist for one quarter. 

Reports suggest this delay stems from Nvidia pushing suppliers to “request speeds of over 11 Gb/s per pin,” well above the JEDEC standard of 8Gb/s. More evidence for a delay is surfacing, with DigiTimes reporting on April 15 that SK Hynix is “considering reducing its planned 2026 shipments of high-bandwidth memory (HBM4) to Nvidia by about 20-30%.”  

We also have another report stating SK Hynix is delaying its HBM4 production ramp until Q3, instead of its original Q2 target, with the delay said to better align with Nvidia’s schedule. Any potential delays or shipment cuts at SK Hynix also could be a key factor in a Rubin delay, as SK Hynix reportedly secured more than 70% of HBM orders for the upcoming chip; on the other hand, Micron and Samsung both have announced that HBM4 is in mass production for Vera Rubin, easing some of the supply constraints. 

Overall, I am not too concerned about 2026 revenue as Blackwell orders are likely to help backfill the Rubin delay. This is less about a revenue miss and more about the strategic shift toward custom silicon. Lastly, Rubin could be more than a one-quarter delay. To compare, Blackwell was a two-quarter delay. The unknown around exactly how long the delay will be is an additional risk that Nvidia investors will have to absorb. 

Nvidia: Seeking to Defend its Throne 

Last quarter, inventories increased more than 8% QoQ to $21.4 billion, but more importantly, Nvidia's supply-related commitments surged. We highlighted this last quarter as a key sign that the strong data center QoQ revenue inflection would continue. 

In Q4, Nvidia's supply-related commitments surged nearly 90% sequentially to $95.2 billion, a major step-up from the prior ~$28-30 billion range through late FY25 and the first half of FY26. Nvidia says it is strategically securing inventory and capacity to meet demand beyond the next several quarters, which we believe serves as a key sign that the current accelerated QoQ data center growth of ~$10 billion will likely persist as Blackwell Ultra continues ramping and as Vera Rubin also eventually ramps. 

While initially, this could be taken as evidence that Blackwell's ramp is persisting; the more likely outcome now is that it signals a Rubin delay. If this is true, the risk is that it sits on the balance sheet until Rubin ships. However, another more likely outcome is that most of these commitments could be converted to Blackwell and Blackwell Ultra. 

TrendForce data supports this theory, stating that industry watchers expect Rubin to account for 22 percent of Nvidia's high-end GPUs, down from 29 percent. As stated above, the reason is: "time required to validate the newer HBM4 memory used by the chips, challenges with the migration to Nvidia's faster ConnectX-9 NICs, the system's higher overall power consumption, and the more advanced liquid cooling requirements [are] contributing to the delays." 

In the same article, the stated assumption is that Blackwell mix rises to 71% while Hopper is down to 7% from original expectations of 10% due to China tensions. 

According to additional checks, this is aligned with Keybanc, stating 2026 supply is expected to support "5.5M-6M Blackwell GPUs, 1.5M Rubin, and 1M Hopper GPUs." KeyBanc's estimates imply higher Hopper revenue — which is what could sting slightly — as these numbers would make up roughly 69% to 71% of Nvidia's 2026 GPU output, while Rubin accounts for about 18% to 19% and Hopper about 12%. Keybanc also cut VR rack estimates by 50% to 6K, down from 12-14K. 

As stated, the Rubin delay may not result in a large impact on revenue as Blackwell is still supply-constrained. One could argue the Rubin delay could help Blackwell’s pricing remain elevated for longer to not have the next generation putting pressure on average sales prices.  

The bigger issue isn't losing the markup in the near-term but rather: (1) is the delay truly only one quarter — we've been here before with Blackwell and the delay was two quarters, and (2) Nvidia's product road map will no longer be seen as invincible.  

Nvidia Stock Long‑Term Outlook: The $20 Trillion Thesis Revisited

Our catalysts to the $20 trillion thesis remain, which is a strong product road map, analyst estimates being far too low in the 2028-2030 window, but even more importantly, my prediction is that Nvidia exits the decade as one of the largest AI software companies. We saw how quickly the company overtook Broadcom as the largest Ethernet company; something similar is what my $20 trillion thesis hinges on, but rather with Nvidia dominating a large portion of the software market across robotics and automation.

Conclusion 

Nvidia remains one of the most important companies in the AI era, and I continue to believe the stock can reach a $20 trillion market cap by the end of the decade. What has changed is not the destination, but the path. The hardware moat that powered the first phase of Nvidia’s ascent is becoming less absolute as inference grows, custom silicon improves, and the next 1-2 product cycles carries timing risk – including both Rubin and Rubin Ultra. 

My thesis hinges on Nvidia reaching $20 trillion with software as the primary catalyst. The issue more near-term is that the market is still largely valuing Nvidia through hardware, just as the durability of that moat is becoming more open to debate.  

For those who have followed me since 2018, it has been a fantastic ride. I am still looking for the same thrill of steep upward stock trajectories unique to the AI market; only in different tickers.

The I/O Fund has built a strong track record in lesser-known AI winners, including Bloom Energy, up 1100% since our initial entry last year, an optical networking stock up more than 620% since November, and one of our largest positions at a 10% allocation already up 130% year to date. We publish more than 100 paywalled articles each year on AI stocks, supported by an actively managed portfolio and real-time trade alerts. Don’t miss out on the AI trade. Learn more here. 

Please note: The I/O Fund conducts research and draws conclusions for the company’s portfolio. We then share that information with our readers and offer real-time trade notifications. This is not a guarantee of a stock’s performance and it is not financial advice. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis. Beth Kindig and the I/O Fund own shares in NVDA at the time of writing and may own stocks pictured in the charts.

Recommended Reading:

• Bitcoin 2026 Price Prediction: Why the Dollar, Global Liquidity and Volume Signal More Downside Ahead
• 2026 Stock Market Outlook: Cycle Convergence & What's Next
• Arm Stock Could Win as Agentic AI Shifts the Bottleneck to CPUs
• Nvidia Stock Prediction: The Path to a $20 Trillion Market Cap is Strengthening