The stocks selected for the Q1 report passed stringent tests for technical positioning, competitive advantage, and underlying fundamentals. Before presenting the list, I also revisit the trends driving the AI market—as you’ll see, much has changed in just three months. 

Despite these trends being long-term bullish, the team at the I/O Fund fully accepts the inevitable downturns that characterize not only technology, but the growth markets that historically drive a disproportionate share of returns. 

For example, last year Nvidia reported mild returns, and did not even beat the broader semiconductor sectors SMH and PHLX, despite continuing to offer some of the strongest fundamentals the market has ever seen (the keyword here is “continuing”). The market is clearly not static and requires a level of discipline that comes naturally to the analyst team at the I/O Fund.  

On the topic of investment discipline, what you have in your hands is a 61-page report totaling over 24,000 words – nearly a novel. What drove this report is a mind-numbing amount of due diligence on the stocks included in the report, but also those we passed on. 

Below are the I/O Fund’s Top 15 AI Stocks for Q1 2026 and the trends driving the AI market forward.  

Reference our Q4 2025 AI Stocks list here and our Q3 2025 AI Stocks list here. 

Top 3 Emerging AI Trends for 2026-2028 

#1 Networking Shifts with Rubin, Yet Importance Remains 

As you’ll recall from our previous coverage, Blackwell and Blackwell Ultra are fundamentally a networking problem. We began to form this thesis nearly a year ago with non-stop AI networking coverage, and we have ample evidence the thesis is playing out.  

Nvidia’s networking segment surged again this past quarter to 162% growth YoY and was up 13% QoQ for $8.2 billion in revenue. We began to see initial signs last quarter from Nvidia with 78% growth YoY and was up 46% QoQ.  

This represents an acceleration of 84 percentage points from 78% YoY growth in Q2, driven by NVLink scale-up, Spectrum-X Ethernet and Quantum X-InfiniBand. For Nvidia’s systems, there is a 75% attach rate which leaves about 25% for smaller networking vendors – therefore, even though this growth is driven by Nvidia’s proprietary networking stack, the growth rates are directionally aligned with smaller players, as well. 

As discussed in last quarter’s Top 15 AI Report, Nvidia’s Blackwell architecture drives a new growth trajectory for AI networking, as it requires 5× to 9× more networking components for 72-GPU and 36-CPU systems to operate as a single node. Because these systems are now shipping in volume, the current networking stack largely reflects the companies capturing this demand. As a result, it is reasonable to expect growth rates among the highest-growth networking stocks to remain healthy over the next one to two quarters.  

Rubin Redefines AI Networking as a Bandwidth-First Constraint  

Inside the Rack: The Copper-to-Optics Boundary 

However, as we turn our attention to the Vera Rubin generation, there is a notable shift in the networking stack. While copper-based links remain essential for short-reach, low-latency connections—particularly within NVLink scale-up domains—the expansion of Ethernet fabrics, higher port counts, and the adoption of co-packaged optics are driving an inevitable shift toward optical content.  

Blackwell and Blackwell Ultra are fundamentally focused on solving scale-up problems, where the primary challenge is binding large numbers of GPUs into a single coherent node using ultra-dense, low-latency NVLink fabrics.  

Rubin, by contrast, is primarily focused on assisting higher bandwidth requirements, as the focus is now on sustaining inference and training workloads at scale without bottlenecks forming beyond NVLink. The limiting factor is how efficiently bandwidth can be delivered and distributed across racks and fabrics, resulting in higher port counts, faster link speeds (800G now and moving toward 1.6T).  

Further necessitating a need for higher bandwidth is Rubin’s “extreme co-designed” nature, as CEO Jensen Huang puts it, where “GPUs, CPUs, networking, security, software, power delivery, and cooling are architected together as a single system rather than optimized in isolation” to deliver substantial performance upgrades for inference, such as a 5X increase in FP4 performance with just a 1.6X increase in transistor count on Rubin’s GPU.  

The increasing amount of computing nodes (especially as Nvidia pushes towards the NVL576 with Rubin Ultra) along with increasing amount of interconnects means that bandwidth must also increase, from 400G to 800G and now to 1.6T, to ensure that low-latency, high-throughput communication remains across the entire platform. 

As a result, it’s expected that optics move closer to the switch, as copper and AEC content becomes constrained by reach and signal integrity. The result is a networking stack where silicon photonics capture incremental value, even though copper remains relevant and intact at the shortest distances.  

With Rubin, Nvidia is doubling NVLink scale-up bandwidth over Blackwell with its sixth-gen NVLink 6 interconnect, offering 3.6 TB/s of bidirectional bandwidth for GPU-GPU communication and 1.8 TB/s for GPU-CPU communication.  

This is accomplished with 36 NVLink 6 switches, deployed as a full all-to-all fabric across the NVL72 rack, delivering 2X throughput for inference at scale with total bandwidth of 260 TB/s per rack, versus Blackwell’s 130 TB/s. 

Source: Nvidia 

Packing more NVLink switches per rack (18 in Blackwell to 36 with Rubin) and doubling bandwidth emphasizes Nvidia’s goal of maximizing scale-up bandwidth to deliver increasing throughput and inference performance gains.  

However, the content opportunity for copper and AECs may slowly erode at the copper-to-optics boundary as the design goal with Rubin is to bring Ethernet closer to the switch, in a shift that favors silicon photonics over time, first through shorter electrical reaches and earlier optical transitions, and eventually through architectures such as CPO, while leaving copper relevant at the shortest distances.  

The industry remains favorable on copper in the near future, with Broadcom CEO Hock Tan saying that the industry will “try to do scale-up within a rack in copper as long as possible” but the shift to SiPho at the electrical-optical transition point appears to be inevitable. Therefore, copper is not going away, rather it faces a lower attach rate. This is due to copper’s reach limitations and needing GPU systems to scale further with a low power, low latency and high bandwidth solution. 

For Credo, the company is expanding its presence into optics as well with its ZeroFlap Optical DSPs and transceivers, though it faces potential decreasing AEC content. 

Scale-Out: CPO Signifies the Shift Toward SiPho 

For scale-out networking, Nvidia announced its new NVIDIA Spectrum-X Ethernet Photonics switch, which it says will deliver 10X greater reliability with co-packaged optics (CPO), bringing 1.6T silicon photonics (SiPho) optical engines directly onto the switch. Maximum bandwidth is also doubled to 102.4Tb/s per ASIC, matching Broadcom’s new Tomahawk6 switch, though Nvidia is also offering the industry’s first four-ASIC design, delivering 409.6Tb/s bandwidth.   

The push towards the new Spectrum-X Ethernet switches will reduce reliance on traditional pluggable transceiver designs. There are a few main advantages this architectural shift: it eliminates the need for digital signal processing (DSP) retimers, reducing latency, and it reduces network power, driving up to 5X better power efficiency with a lower cost versus pluggable transceivers.   

It also will drive increasing 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, ultra-high-power (UHP) lasers for external light source (ELS) modules, fiber array units, optical interconnects, and more.  

PCIe Remains Relevant from Nvidia’s “Extreme Co-Design" 

Inside the server, PCIe remains firmly intact. Growth should persist as PCIe continues to serve as the foundational interconnect for intra-server connectivity between GPUs, CPUs, DPUs, NICs, and NVMe SSD storage. 

This directly ties back to Nvidia’s “extreme co-design” philosophy. As Rubin brings multiple compute, networking, and memory components together into a single, tightly integrated platform, the need for rapid, low-latency data movement within the server increases—therefore, it is our understanding PCIe as the connective tissue does not decrease.  

This extends further with Nvidia’s move to PCIe Gen6 alongside expanded CXL support on its Vera CPU, up from PCIe Gen5 on Grace. CXL enables low-latency, high-bandwidth memory and cache sharing between CPUs, GPUs, and attached memory devices, reinforcing PCIe’s role at the heart of the system architecture. PCIe fabric switches are also expected to remain critical for backend GPU-to-GPU communication and for linking CPUs, NICs, and storage at scale. 

#2 AI Energy: AI's Biggest Bottleneck 

The AI market has moved from being compute-constrained to being energy constrained. Hyperscalers have access to GPU supply, making the limiting factor how quickly those GPUs can be energized and deployed.  

As we’ve discussed in our analysis, Why Power is Critical for Data Centers and their Hyperscaler Customers, every month that GPUs sit idle waiting for power delays – revenue, profits and market share can be affected. This is especially true given GPU generations refresh annually and is driving significantly higher power requirements. 

For the energy section, I break down both the problem and the solution — each central to how we plan to position heading into 2026. Many of these energy solutions have existed for decades yet are now experiencing a resurgence in product-market fit driven by rapid AI data center expansion. For that reason, even if we have already covered the scale of AI data center investment, it is critical to double-click on why energy has become the bottleneck nearly overnight.  

The Problem: 

Nvidia’s Blackwell lineup is bringing a significant increase in power consumption, nearly double the H200’s 70 kW at 120 kW for the GB200 NVL72 and 140 kW for the upcoming GB300 racks.   

Beyond Blackwell, Nvidia’s future design lineup shows continual increases in power consumption. Its Vera Rubin generation is expected to boost thermal design power (TDP) by 50% over Blackwell at up to 180 kW to potentially 230kW per rack, with the Rubin Ultra boosting this to 600kW by late 2027.  

In its largest configuration, the Vera Rubin NVL576, dubbed the ‘Kyber’ rack, could draw as much as 600 kW (0.6 MW), or 5x that of the GB200 NVL72 in just a two-year design timeframe. These figures do not include networking, interconnects, cooling and other hardware, which will further boost power draw per rack. 

Existing data center infrastructure is largely incompatible with next-gen AI. Nearly 70% of data centers were built for 4-9kW racks with fewer than 2% able to handle even 50kW, which is forcing new construction and major retrofits. 

Furthermore, there exists a significant disconnect between when hyperscale and colocation developers expect to have site power, and when utilities expect to be able to deliver said power. Connecting new data centers to the grid in quick fashion may not be the most feasible option for hyperscalers looking to deploy gigawatts of capacity quickly, and instead, alternative power sources may be in higher demand. 

For example, across the board, developers are expecting to have power delivered by late 2026 to early 2027 on average, with most regions seeing expectations as early as late 2025. This is likely driven by consistent strong demand for AI infrastructure services, as new capacity will allow hyperscalers to meet more demand and drive more revenue. 

Yet, utilities do not expect to be able to meet these delivery timelines in most of these primary and secondary markets, with many projecting late 2027 through 2028, with major hub Northern Virginia seeing one of the longest timelines at nearly 2029. 

Most importantly, the AI race is not merely a battle between companies like Google, Amazon and Microsoft. Rather, it is a battle among global powers. While the news has latched onto China-fears such as DeepSeek, tariffs or rare earth materials, and H200 bans (that are later lifted), the true challenge lies in the fact that China has significantly more power than the United States.  

In a recent Fortune article, energy experts stated China’s reserve margin has never dipped below 80% to 100% nationwide, meaning it’s at 2X the capacity the country needs. Meanwhile, the United States is at a 15% reserve margin. The article states, “The gap in readiness is stark: While the U.S. is already experiencing political and economic fights over whether the grid can keep up, China is operating from a position of abundance.” Specifically, the article calls out that large-scale infrastructure projects depend heavily on private investment, yet returns can take years and up to a decade to pay off. Meanwhile, private investors greatly prefer software with returns realized on a much shorter timeline.  

Therefore, there are dual forces placing outsized pressure on this trend – not only is the AI data center expansion physically dependent on power availability and it is now the bottleneck, but one could argue that United States global dominance is also highly dependent on this sector. The United States undeniably has the world’s best design companies with Nvidia, AMD, Broadcom and soon TSM will be on our soil. We also have the best software companies – from Big Tech to the entrepreneurial culture of our country with startups coast to coast.  

What the United States doesn’t have is enough power.doesn’t have is enough power. 

The Solution: 

The bottleneck has shifted from compute supply to energy. As a research firm, we want to be early in providing you analysis on the companies that solve this problem, as energy determines the timing and economics of AI deployment. 

The disconnect described above is driving demand for behind-the-meter power, on-site natural gas turbines, fuel cells, nuclear and SMRs in the long-term and retrofitted Bitcoin mining sites. Each GW of AI data center capacity costs roughly $30 to $38 billion all-in, which puts total required capex into the trillions this decade.  

Compute still remains the bulk of the data center spend (i.e., the overall pie), however, energy is growing its slice of the pie. Breaking it down on a MW basis, Alpha Matica, an AI consultant company that specializes in AI, states electrical systems are 50% of the initial construction costs which range from $900M to $1.5B per 100MW.  

There are multiple different ways that hyperscalers, neoclouds and developers can get power to data centers to meet upcoming demand growth over the next few years, each offering its own benefits and drawbacks.  

Grid interconnection: This is when data centers connect to the power grid under standard service, providing access to flexible power needs with no additional capex and a wide range of power generation options, including renewables. However, grid interconnection requests are often the longest time to power, ranging from three to seven years for hyperscale data centers in most key markets. 

Behind-the-meter: How Power is Contracted can offer a time to power advantage 

BTM refers to when data centers connect directly to the power source and bypass the retail grid (meter) and associated tariffs, which can offer significant time advantage with stand-up times often in the range of several months to a year, along with cost savings from buying power direct versus at retail price.  

BTM arrangements also provide greater control over power supply and reduce exposure to grid outages. These deals can be structured across multiple power sources, including solar, wind, nuclear, and natural gas.  

On-site power generation: Where Power is located can offer time to power benefits and is increasingly becoming attractive to AI data centers 

With on-site power, data centers will install their own power source within the facility grounds or adjacent, also offering a relatively quicker time to power of a few months to over a year as this bypasses the need for a grid connection and transmission upgrades. 

These solutions are modular and designed for rapid installation, as they can be manufactured off-site and be built in parallel with a data center. This also offers the benefit of foregoing regulatory risks and delays that come with grid connections. Controlling the timeline can be a significant asset for the reasons described above in terms of when increasingly-power-hungry GPUs are shipped.  

In addition, grids are designed for lower, flatter power loads and gradual increases. With Blackwell Ultra, Rubin and Rubin Ultra on the product road map through 2027, data centers need ultra-dense loads and to scale power locally without overwhelming the grid. In addition, as discussed above, time-to-power is technically time-to-revenue, and thus, circumventing the grid as much as possible is the aim. 

Lastly, on-site power is seen as more reliable as it’s not subject to grid failures. This is why on-site backup power sources are also becoming a major growth market. 

On-site power can come in many forms, such as Bloom’s fuel cells, natural gas turbines or generators such as those from GE Vernova or Caterpillar, and in the 2030s and beyond, potentially small modular nuclear reactors. These power sources are discussed more below. 

• Natural gas turbines/generators: Behind-the-meter and on-site

NG is a widely available fuel source with a broad pipeline in the US, offering continuous power to data centers. Turbines can come in a range of sizes and be easily deployed, such as Caterpillar subsidiary Solar’s SMT-130 turbines that xAI is using, or GE Vernova’s LM2500XPRESS that Crusoe is using, scaling up to 1GW capacity. Notably, NG turbines could help meet substantial future demand, as GE Vernova is expanding manufacturing in South Carolina to be able to ship 20 GW worth in 2027. Large (>225MW) turbines are reportedly sold out over the next three years.

• Fuel cells: Behind-the-meter and on-site 

Similar to NG, fuel cells can be quickly deployed (in as little as three months per Bloom and Oracle’s deal), and provide continuous power for operations. Due to being a relatively newer tech, SOFCs can come at a higher cost than NG, but without the related emissions. Bloom is planning to double its SOFC manufacturing capacity to 2GW in 2026 to meet rising on-site power demand.

• Small modular reactors: Not behind-the-meter right now (will be around 2030) and on-site or near-site

SMRs are drawing more interest for future demand needs, as commercialization at scale is not likely until 2030 or beyond. Google is working with Kairos to bring 0.5 GW of SMR capacity online from 2030 through 2035, while Oklo and NuScale are progressing with commercialization plans and a long-term combined ~20 GW backlog.

• Retrofitting existing infrastructure, i.e. Bitcoin mining: How power is contracted and delivered is nuanced (see below); effectively on-site

Miners leverage existing infrastructure with secured power to the building, offering quick delivery times as short as a few weeks to a year, depending on cooling, flooring or other upgrades needed. Overall, the value proposition of Miners is that they are cheaper and faster than new, greenfield data center sites. 

Bitcoin mining is not behind-the-meter in a strict sense, but it is effectively behind-the-meter because miners secure direct, wholesale power through upfront contracts that are not renegotiated.  

They are considered on-site power as there is minimal transmission dependency due to co-locating near a mix of power sources (near gas plants, augmented by wind, solar, water/hydro). In most cases, even if a utility meter exists, the power system is purpose-built for that site and is not shared retail infrastructure.  

While this method can offer quick time to power for >100MW sizes with low latency, low electricity costs and cooling expertise compared to greenfield projects, miners are capital constrained and may be unable to build-out capacity beyond what is currently in their pipelines. For example, they are not suitable for training a frontier model. 

Miners have been attracting substantial deal activity, primarily from neoclouds, from an ability to deliver larger chunks of power quickly, with capex costs well below greenfield builds. 

Click here for our most recent full write-up on Why Power is Critical for Data Centers and Their Hyperscaler Customers. The I/O Fund first covered this topic  in June of 2024, many quarters before the problem became well-known. We furthered this by investing early in a Bitcoin miner and one of the year’s highest-performing AI energy stocks.Why Power is Critical for Data Centers and Their Hyperscaler Customers. The I/O Fund first covered this topic  in June of 2024, many quarters before the problem became well-known. We furthered this by investing early in a Bitcoin miner and one of the year’s highest-performing AI energy stocks.

#3: The Incoming AI Inference Market (i.e., the AI Boom Hasn’t Happened Yet) 

As you’ve likely noticed, we have been writing 10,000-foot level analysis on the AI inference phase, which is synonymous with the Monetization Phase for AI. The last few years have been marked by intense R&D and high compute costs, yet the economic reality is that training large language models is an initial research stage, and this is not the stage to expect recurring revenue and expanding profits.  

Rather, it is my assertion that AI development is nearing a crux where peak capex spending intersects steeply with low ROI. However, to call this a bubble or to claim that AI does not drive enough revenue in exchange for the hundreds of billions being spent on data center expansion is to assume we are in the final stages of AI rather than the early stages. The AI market will take off when the inference phase fully arrives – my estimates are 2027-2028 for this. However, as you know by now, the I/O Fund has no intention of being late to this trend.  

The aggregators and distributors for AI – whether that’s Big Tech, best-of-breed software companies, or enterprises that are already using AI to increase profits will be able to fully leverage LLMs and AI automation once AI inference becomes faster and cheaper. It is a mix of hardware and software that will achieve this, and we will want to assess this market carefully as the inference opportunity is expected to exceed the training market in both size and velocity. You can expect to hear extensively from us on this trend over a 2-3 year time period.  

For now, what is most important is to track some of the hardware companies that are unlocking this opportunity. We want to understand the “why” behind Nvidia’s Vera Rubin architecture, the “why” behind AMD’s Helios and the “why” behind Broadcom’s rise in custom silicon. The breadcrumbs are crucial for positioning correctly into H2 2026 and 2027.   

We will also highlight select software opportunities with the understanding that timing may be early. In these cases, price action will play a decisive role. That means if we see a breakout or strength in the chart align with product, we will move accordingly. 

Memory: 

Memory's medium-term thesis is based on the shift from training to inference. Inference workloads require only a forward pass, making them significantly less compute- and power-intensive than training. However, to achieve low latency, especially at small batch sizes, models must remain resident in device memory, which shifts the primary bottleneck from compute over to memory capacity and bandwidth. 

Forward pass refers to taking an input and pushing it through a trained neural network. No learning occurs and no weights are updated. Rather, input tokens are embedded – which requires frequent memory access, yet this step has low compute requirements. Attention layers read fixed, pre-trained model weights, and the GPUs repeatedly read the weights from memory. From there, KV cache grows with usage to where the longer the conversation, the larger the KV cache. While reading from the KV cache reduces redundant computation, it materially increases memory capacity and bandwidth requirements. 

The point of the above paragraphs is to help illustrate the technical shift toward memory for the inference workloads, whereas compute requirements on a relative basis become reduced. 

Discovery Members Discovery Members recently received an analysis on a stock that is positioned to benefit as Nvidia tackles the context memory bottleneck and extends KV cache memory with its new Inference Context Memory Storage platform. To subscribe to Discovery with 40% off, click here to email us or email  premium@io-fund.compremium@io-fund.com and mention code DISCOVERY40DISCOVERY40

Top 15 Stocks List

Section 1: AI Accelerators 

AI accelerators are technically the #1 trend in the AI market by size, and in my opinion, offer a solid way to participate with lower risk than other AI trends over a longer-time frame. While we did not rank accelerators among our Top 3 themes, given that other areas of the market offer higher near-term growth rate, that should not be mistaken for a lack of structural strength. 

As last year illustrated, keeping up on the ins/outs can be advantageous given AMD outperformed Nvidia with 3X higher returns whereas Nvidia underperformed its sector. I review this in more detail below. 

Additionally, in a similar way that doctors check vital signs, we revisit capex ahead of earnings and immediately following Big Tech earnings as this remains a critical signal to the strength of the AI market. One day, we will be tracking enterprise AI spend and sovereign AI. However, those markets are not large enough to offset Big Tech’s investment levels for many quarters (if not years). Therefore, the customer concentration works in our favor as these particular customers must disclose their budgets in their quarterly filings. 

Capex signals from Big Tech/hyperscalers (Microsoft, Meta, Alphabet, Amazon, Oracle) are projected to be around $435 billion for 2025, while initial estimates for 2026 capex are around $583 billion, up approximately 34% YoY. On a dollar basis, this points to an initial estimate of ~$148 billion in growth, versus ~$173 billion in 2025, signaling AI demand is poised to continue.  

Keep in mind, capex estimates for 2025 started out much lower – estimates entered the year at about $320 billion, or more than $100 billion short of where the year ultimately ended. Therefore, the same could be true for 2026 to where capex ultimately ends up higher by year-end — some analysts are already penciling in the five to spend more than $600 billion next year, which could mean absolute dollar growth on a YoY basis this year surpasses 2025.  

Nvidia: Greater Emphasis on Memory 

Overview: 

There are two primary factors to track when assessing Nvidia’s path toward a potential $20 trillion market capitalization by 2030. The first is the cadence of GPU generations and the product road map, which when executed well, supports higher average selling prices and drives system-level expansion in data centers. The second is analyst estimates, which when conservative, can create opportunities for valuation upside as expectations are forced to reset.  

Jensen Huang spoiled the CY2026 alpha party by stating that management has a line of sight to $500 billion across two years from Blackwell and Rubin. Our firm had already stated we would see over $300 billion this calendar year, and that statement puts Nvidia’s revenue squarely at our estimate. Nvidia resuming H200 sales could perhaps bump that up 10% to 20% – not chump change for a stock this size yet a bit boring for I/O Fund’s purposes. 

From there, things get interesting. If we look at calendar year 2027, we see 27% growth estimated to $409 billion. If we look at calendar year 2028, we see only 8.4% growth to $443 billion. Yet in the 27% growth year we will see Rubin Ultra, a 144-GPU AI system that will shatter all previous records on training frontier models as it tests the upper limits of the amount of compute, memory and networking that can function as a single node. In terms of what is accomplished on the inference side, quite a bit depends on how memory and networking evolves over the next 1-2 years to improve efficiency at the system-level.   

Let’s say we get to CY2028 and Nvidia growth flatlines – what would cause this? While the broader market likely anticipates it will come from the pace at which compute can be monetized, yet the more likely cause would be power availability first and foremost, but also deployment complexity, as we saw from the meaningful delays in Blackwell’s 72-GPU systems.  

In the more near-term, 2026 is shaping up to be a year where Nvidia is firing on all cylinders. Vera Rubin was officially launched at CES and is an architecture that opens doors for the impending inference market. Jensen Huang calls the Rubin “extreme co-design" across six elements– CPUs, GPUs, NVLink, Ethernet, DPUs and NICs – with this generation more focused on memory movement and networking than the architectures in the past (which were centered around raw compute). 

The Rubin architecture delivers substantial performance gains at the system level, with up to ~5× improvements in inference and ~3.5× in training relative to prior generations (as always, this depends on workload and configuration).  

A key driver of these gains is a significant expansion in memory capacity and bandwidth, with Rubin designed to support up to roughly 288GB of HBM-class memory per GPU and materially higher memory bandwidth. This addresses one of the primary bottlenecks in inference: memory access and data movement. 

To support this shift, Nvidia is focusing on the context memory window, which refers to the memory used to store and access the model during inference. The key-value (KV) cache is a memory mechanism used in transformer models to store attention keys and values from prior tokens. This allows models to reuse previous computations and reduce latency from increased memory usage. 

In a Discovery tier article where we covered a major memory beneficiary of the KV cache increase, it was stated the KV cache has a substantial memory footprint, and during deployment it can consume 30% of GPU memory, making it a major bottleneck for large-context applications, such as coding, natural language processing, or handling simultaneous requests from many users on large models.    

In day-to-day use, the key-value cache is the memory that lets LLMs remember what’s already been said so it doesn’t have to rethink everything for each query. Each new response then builds on the stored context instead of recomputing the conversation again. When you use ChatGPT or Claude, the prior context is stored in the KV cache rather than relying on repeated compute.  

With the Rubin generation, by expanding the KV cache capacity, Nvidia greatly reduces the need for recomputation, and redirects resources to memory capacity, bandwidth and data movement to improve throughput and responsiveness. 

This becomes even more important in agentic AI systems, where models operate autonomously across multiple inference steps rather than responding to a single prompt. Agentic workflows require longer context windows and sustained access to KV cache as agents reason, plan, and act across extended sequences.  

As a result, memory and networking increasingly determine real-world inference performance and scalability as opposed to raw compute. This marks a shift for Nvidia – and one we argued years ago would open the door to more competition in AI accelerators as we exit the training-dominate phase and we approach the inference-driven monetization phase.  

Overall Revenue Growth: 

Nvidia’s Q3 rev grew by 62.5% YoY and 22% QoQ to $57B. Revenue growth accelerated by 6.9 percentage points from 55.6% YoY growth reported in Q2. Revenue beat estimates by 3.5% and is the strongest beat in the last four quarters.  

Management also provided a strong Q4 rev guide of $65B billion, YoY growth of 65.3% and up 14% QoQ. Beat the estimates by 5.1%. 

Nvidia’s total supply-related commitments, such as for CoWoS wafers, HBM memory, or other components, surged nearly 52% QoQ to $50.3 billion in Q3, with management noting that they are “ordering to secure long lead-time components, meet the demand for Blackwell, and support future architecture ramps.”   

Where the disconnect happens with analyst estimates is what will happen after next year as this is where analyst estimates show minimal growth through 2030 revenue with $437 billion whereas I am calling for double that by 2030. While Blackwell Ultra gets us to a new milestone of $50 billion to $75 billion quarterly revenue, quite a bit of my thesis depends on Vera Rubin, Rubin Ultra and the Feynman generations – not only execution on Nvidia’s side but also power availability is crucial. 

AI Segment Growth: 

Data center rev grew by 66% YoY and 25% QoQ to $51.2B. Rev growth accelerated 10 percentage points from 56% growth reported in Q2. Management sounded confident to achieve the $500B target in Blackwell and Rubin revenue set for FY2026/27 and hinted it could be more. Networking rev grew by 162% YoY and 13% QoQ to $8.19B. Rev growth accelerated by 84 percentage points from 78% in Q2. Largest QoQ growth in about two years (and done at scale). 

Nvidia’s guide pointed to this momentum continuing into the fourth quarter, implying that data center revenue could be on track to rise another $8 billion QoQ for 15% growth. Management repeated they “currently have visibility to $0.5 trillion in Blackwell and Rubin revenue from the start of this year through the end of calendar year 2026.”  

Q4’s guidance suggests that this $50 billion data center segment will quickly be in the rear view mirror, with the $65 billion guidance implying data center revenue of around $59 billion assuming similar mix shift as Q3. This represents another 15% QoQ growth on top of Q3’s 25%, or essentially the data center segment rising nearly 44% in just two quarters. 

Earnings: 

Q3 adjusted EPS grew by 60.5% YoY and 23.8% QoQ to $1.30, beating estimates by 3.5%.  

Looking forward, analysts expect FY2027 adjusted EPS to grow 49.5% YoY to $6.83 and 26.7% YoY to $8.65 in FY2028. 

Margins: 

Q3 GM was 73.4%, beat the guidance by 10 bps. Q4 GM guide is 74.8%, up 140 bps seq and up 180 bps YoY. Mgmt expects to maintain GMs in the mid-70s range for FY2027 despite the increase in the input costs. 

Cash: 

Q3 FCF grew by 31.6% YoY to $22.1B with a FCF of 38.7%, compared to 47.9% last year and 28.8% in Q2. The company has cash and marketable securities of $60.6 billion and debt of $8.47 billion. 

Valuation: 

Nvidia trades at a forward P/S ratio of 24.3. The company has traded at a minimum forward P/S ratio of 9.6 and a maximum of 45.8 in recent years. Nvidia is currently trading slightly lower than mid-range. On the bottom line, it trades at a forward P/E ratio of 39.4. Nvidia has traded at a minimum of 15.8 and the highest of 50.7. Nvidia is currently trading slightly higher than mid-range.   

Notable Risks: 

The risks to Nvidia are low – perhaps the lowest of any stock in the tech universe. With that said, I recently stated in a Seeking Alpha webinar that the predominant constraints are memory and energy now, which means Nvidia is losing its top place in terms of GPUs no longer being the top supply constraint in the AI market. 

Broadcom: Ethernet Wins at Scale-Out & Custom Silicon will Prevail with Inference 

It’s widely understood that Broadcom supplies Google with its custom TPUs. The incoming inference growth curve, that the I/O Fund detailed here, has led CEO Hock Tan to state Broadcom may witness an acceleration of XPU demand into the back half of 2026.  

Tan stated, “In fact, what we've seen recently is that they are doubling down on inference in order to monetize their platforms. And reflecting this, we may actually see an acceleration of XPU demand into the back half of 2026 to meet urgent demand for inference on top of the demand we have indicated from training.”    

Something similar was echoed in the FQ3 call, with Tan stating: “But also as for these guys, they got to be accountable to being able to create cash flows that can sustain their path. They [are] starting to also invest in inference in a massive way to monetize their models.”  

On that note, Google’s TPU business received a significant vote of confidence recently with Anthropic signing a deal for up to one million TPUs, including Ironwood, coming online in 2026. The deal is said to be worth tens of billions.   

For Broadcom, TPUs are expected to be the primary driver of AI revenue growth in fiscal 2026 – estimates from HSBC earlier this summer projected Google’s TPUs to represent ~58% of Broadcom’s ASICs shipments at 1.79 million, but account for ~78% of ASICs revenue at $22.1 billion. This is because Google’s TPUs were estimated to carry a significant price premium at $13,000 per chip versus Broadcom’s other projects at $5,000 per chip. However, this is still less than half the cost of Nvidia’s chips at $30,000 to $40,000 for a solo B200 ($60,000 to $70,000 for a GB200).   

Looking beyond fiscal 2026, projections for TPU shipments are surging. Morgan Stanley now expects 5 million TPUs to be shipped in 2027, a 67% rise from its prior estimate for 3 million; for 2028, the firm estimates shipments as high as 7 million, a 120% increase from its prior estimate. This would project YoY growth of 40% from 2027 to 2028, a substantial increase from 6% previously, and will represent more than 2X growth in two years. 

The I/O Fund first covered TPUs versus GPUs back in 2019 and revisited the topic in February 2024 in our analysis, Broadcom: Networking/ASICs Giant and the Second Largest by AI Revenue. Since then, we’ve provided quarterly coverage for two years. Broadcom: Networking/ASICs Giant and the Second Largest by AI Revenue. Since then, we’ve provided quarterly coverage for two years.   

The shift to Ethernet and away from Nvidia’s lock-in ecosystem of GPU + InfiniBand is benefiting Broadcom, with the industry pointing to rising Ethernet demand. Arista said that momentum for Ethernet “has really shifted in the last year” while Nvidia touted that its new Spectrum-X Ethernet is annualizing at $10 billion in revenue, or $2.5 billion quarterly.   

The company is committed to remaining on the leading edge of networking with its Tomahawk 6 switch, the industry’s first 102.4 Tbps Ethernet switch. The next-gen switch doubled the bandwidth of its predecessor, while offering flexible deployment ability with 1,024 100G or 512 200G SerDes options, reducing switch count.   

This raw performance upgrade paves the way for >100K to 1 million accelerator clusters by allowing larger leaf-spine fabrics to be constructed, while drawing less power and keeping latency low. Broadcom exec Ram Velaga said that the demand for the new switch is “unprecedented” with multiple >100K accelerator deployments “using Tomahawk 6 for both the scale-out and scale-up interconnect.”  

When discussing Tomahawk 6, management points toward the flattening of the AI cluster as an important catalyst for this product, stating: “[…] Tomahawk 6 enables clusters of more than 100,000 AI accelerators to be deployed in just two tiers instead of three … this flattening of the AI cluster is huge because it enables much better performance in training next-generation frontier models through a lower latency, higher bandwidth and lower power.” The two-tier topology also reduces complexity of cluster construction and reduces congestion choke points significantly, addressing another critical pain point of building larger and larger clusters.   

Additionally, in terms of the AI networking opportunity, scale up is 5-10X more than scale out – setting up a nice trajectory as AI clusters grow. Oppenheimer analyst Rick Schafer highlighted that they expect next-gen Tomahawk6 volumes to ramp up in the second half of next year, providing added growth and gross margin boost. 

Overall Revenue Growth: 

Second-highest in AI revenue among the semis. Broadcom’s FQ4 revenue grew by 28.2% YoY and 12.9% QoQ to $18.02 billion, beating estimates by 3.2%. Management also provided a strong FQ1 revenue guide of $19.1 billion, implying a YoY growth of 28.1% and 6% QoQ, beating estimates by 4.3%. The expected strong growth is primarily driven by AI revenue, which is expected to double YoY to $8.2 billion.  

FQ4 semiconductor solutions revenue grew by 35% YoY to $11.07 billion, primarily driven by strong AI revenue. Revenue growth accelerated by 9 percentage points from 26% growth reported in FQ3. Management expects semiconductor revenue growth to further accelerate 15 percentage points to 50% YoY, reaching $12.3 billion in FQ1, driven by a surge in AI revenue. For FY2025, semiconductor revenue grew by 22% YoY to a record $36.9 billion.  

Management expects renewals to be seasonal in Q1 and expects Infrastructure Software revenue to be $6.8 billion, down (2%) sequentially and up 1% YoY. 

AI Segment Growth: 

FQ4 AI revenue grew by 74% YoY and 25% QoQ to $6.5 billion and was higher than the management guide of $6.2 billion. For the FY2025, AI revenue grew by 65% YoY to $20 billion. Management expects AI revenue to accelerate in FY2026 and drive most of Broadcom’s growth in FY2026.  

During fiscal year 2025, AI revenue grew 65% year-over-year to $20 billion, leading to semiconductor revenue seeing an all-time high of $37 billion 

Next quarter, AI revenue is expected to double year-over-year to $8.2 billion.  

Broadcom’s total AI-related orders on hand exceed $73 billion, nearly half of the company’s consolidated $162 billion backlog. The $73B backlog is expected to ship over the next 18 months. This backlog includes not only XPUs but also networking components. Most of the earnings call was management explaining the $73 billion is a baseline for the next 18 months. 

Earnings: 

FQ4 adjusted EBITDA grew by 34.4% YoY to $12.2 billion with an adjusted EBITDA margin of 68% and was better than the management guide of 67%. For FQ1, management expects adjusted EBITDA margin to be down 100 basis points sequentially and YoY to 67%.  

Adjusted EPS grew by 37.3% YoY to $1.95, beating estimates by 4.3%.  

FQ4 GAAP EPS grew by 93.3% YoY to $1.74. While adjusted EPS grew by 37.3% YoY to $1.95, beating estimates by 4.3%. Analysts expect adjusted EPS to grow by 23.3% YoY to $1.97 in FQ1 and 28.7% YoY to $2.03 in FQ2. 

Margins: 

Adjusted gross margin was 77.9%, up 100 basis points YoY and down 50 basis points sequentially.  

Operating margin improved 8.8 percentage points YoY and 4.8 percentage points sequentially to 41.7%, primarily driven by operating leverage. This is up 2X from 1-2 years ago. 

The company will have to pass-through more third-party components such as memory, optics, and power infrastructure, which will lead to gross margins contracting. However, management was clear that gross profit dollars and operating income dollars will continue to rise due to scale and operating leverage. 

Cash: 

Broadcom’s cash flows are improving, driven by higher profits. FQ4 free cash flows grew by 36.2% YoY to $7.47 billion with a free cash flow margin of 41.4% compared to 39% in the same period last year.  

The company has debt of $65.1 billion and cash of $16.2 billion. The debt is high due to the past acquisitions. However, the company has a history of successfully reducing debt. Also, the company has strong cash flows. Cash has increased to $16.2 billion from $10.7 billion due to higher free cash flows. 

Valuation: 

Broadcom trades at a forward P/S ratio of 15.9. The company traded at a minimum forward P/S ratio of 6.7 and the maximum of 28.8 in recent years. Broadcom is trading slightly lower than mid-range. It is trading at a forward P/E ratio of 31.7. The company traded at a minimum forward P/E ratio of 17.3 and a maximum of 57.2. Broadcom is trading slightly lower than the mid-range on the forward P/E ratio as well.  

Notable Risks: 

Similar to Nvidia, the company is a high-quality stock with a relatively low risk profile. The primary risk is high debt, which as we have discussed above, is well controlled. However, Broadcom is in sharky waters on networking in terms of competition, and even on custom silicon with a rumor Google could be moving some orders for its next generation of TPUs (v7 and v8) over to MediaTek. 

AMD: The Element of Surprise 

Overview: 

AMD is a great example of the paradox of stock investing, which is that despite Nvidia and Broadcom posting higher growth on a much larger revenue base, AMD outperformed Nvidia and Broadcom last year by roughly 2X. 

Five years ago, I dubbed AMD the “Dark Horse” for my premium research members as the company had a mere 4% share in the CPU-data center and was up against the near-monopoly of Intel. AMD has proven there is an element of catching the market off guard that helps to compound returns. The opposite of this is known as a crowded trade – which leads us back to the chart pictured above.  

In more recent years, the I/O Fund has remained consistent in our conviction that AMD will eventually contend with Nvidia on GPUs while emphasizing that timing is key. About 18 months ago, I spelled out AMD could outpace Nvidia’s returns by 2030 stating in a Real Vision video interview that the company’s opportunity is closely tied to the inference market. At the time, AMD was in the doghouse: 

“Core to this thesis on AMD is giving time for the budding inference market to take off and mature – [I explained that]“where AMD is going to compete with Nvidia is a market that is very early, so we need time for that to mature, which is inference. Many people may get that confused, because we are fully in the AI market today because Nvidia is putting up those huge data center numbers. We are in the data center training market today; one day, we will be an AI market led by inference.”  

It’s important to note that my prediction that AMD can outpace Nvidia’s returns by 2030 hinges on AMD capturing 20% to 25% of the GPU-market. We all know that Nvidia is not Intel, and thus AMD faces a fiercer competitor on all accounts. However, the path that AMD took to overcome Intel is highly relevant. You can read more about that here and here. 

The overall thesis is that the data center GPU market desperately needs a second-place contender. Investors may appreciate Nvidia’s pricing power, but hyperscalers and companies like OpenAI do not; they’d like to see more competition and optionality including lower prices. That is why we are seeing Meta work alongside AMD to bring Helios to market and a recent 6GW deal from OpenAI. 

One key area where Helios stands out is memory — the platform offers roughly 50% more total memory capacity compared to Nvidia’s Vera Rubin rack architecture. AMD will offer 1.4 PB/s of memory bandwidth, slightly below Rubin’s 1.6 PB/s as Nvidia is said to be requiring pin speeds of 11 Gb/s, above the standard 8 Gb/s, driving the higher bandwidth despite lower HBM content. The HBM content and nearly comparable bandwidth will likely make AMD a compelling solution for inference workloads considering its price-advantage over Nvidia. 

That said, if you’ve followed AMD’s AI story as closely alongside the I/O Fund (and I know many of you have), then the most important leap in this generation of GPUs is not found in Helios specs or even this quarter’s commentary. Rather, it’s in the demand signals. For the first time, some of the most influential AI customers — including OpenAI, Oracle, and Meta — are preparing to deploy the MI400 Series in meaningful volume. That level of hyperscaler commitment is something AMD hasn’t enjoyed in prior GPU generations (MI300s), and it represents an important shift in the company’s competitive positioning. 

There are many investment opportunities in AI across AI networking, AI energy, AI software, AI data layer and more – but none compare to the sheer size and strategic importance of GPUs, particularly when there are so few players competing for that share. That scarcity dynamic is precisely why AMD remains a special case in our portfolio. 

Overall Revenue Growth: 

Q3 revenue grew by 35.6% YoY and 20.3% QoQ to a record $9.25 billion, beating estimates by 5.7%.  

Q4 revenue guide is $9.6 billion at the midpoint, representing a YoY growth of 25.4% and 3.8% sequentially. It beat the analyst's estimates by 4.3%. Similarly, to the last quarter, the revenue guidance does not include any MI308 chip sales to China. However, this time management indicated that MI308 chip sales could be coming soon. 

AI Segment Growth: 

Data Center revenue rebounded strongly in Q3 as it grew by 22% YoY and 34% QoQ to a record $4.3 billion. The strong growth was primarily driven by the ramp of the Instinct MI350 Series GPUs and server share gains. However, we aren’t quite there yet in terms of a strong inflection as it was stated data center would grow 4% QoQ with strong growth server (a nod toward CPUs instead of GPUs). Server CPU revenue reached an all-time high as adoption of 5th Gen EPYC Turin processors accelerated rapidly, accounting for nearly half of overall EPYC revenue in the quarter. The sales of prior generation EPYC processors also continued to be strong.  

The last guide that AMD provided on GPUs was $6.5 billion in revenue by the time we exit this year. Management is hinting they will see “tens of billions” in their AI business by 2027. If we assume this means a minimum of $20B (perhaps more) then it coincides with roughly 200% growth in AMD’s AI business over a two-year time span. 

Margins: 

The gross margin was 52%, up 200 basis points YoY primarily driven by a higher profitable product mix. Management has guided an adjusted gross margin of 54.5% for the fourth quarter.  

The operating margin improved by 300 basis points YoY to 14%. Adjusted operating margin was down by 100 basis points YoY to 24% and missed the management guidance of 25% as the adjusted operating expenses increased by 42% YoY to support the significant AI opportunities and go-to-market activities for revenue growth. Management has guided an adjusted operating margin of 25% for the fourth quarter. 

EPS: 

GAAP EPS grew by 59.6% YoY to $0.75, beating estimates by 10%. Adjusted EPS rose by 30.4% YoY to $1.20, beating estimates by 2.4%.  

Analysts expect adjusted EPS to grow by 22.3% YoY to $1.33 in Q4 and accelerate to 26.4% growth in Q1 and 183.2% YoY growth in Q2 to $1.36. Looking forward, they expect the adjusted EPS to grow by 61% YoY to $6.35 in 2026 and 45.7% YoY to $9.25 in 2027. 

Cash: 

Q3 free cash flows grew by 208% YoY to $1.53 billion or 17% of revenue, up 10 percentage points YoY.  

Cash of $7.24B and debt of $3.22B. 

Valuation: 

AMD is trading at a forward P/S ratio of 9.2. The company has traded at a minimum of 3.7 and a maximum of 13.3. AMD is trading at mid-range. On the bottom-line, it is trading at a forward P/E ratio of 38.6. The company has traded at a minimum of 19.7 and a maximum of 66.5 in recent years. AMD is trading at mid-range on a forward P/E ratio as well.  

Risks: 

AMD carries execution risk as taking Nvidia head-on is not for the faint of heart. Margins tend to be lower with AMD as one of their tactics is to offer much lower prices than their competitors.  

TSM: Multi-Year Visibility for AI Megatrend 

TSMC is one of the least sensational management teams in the AI space, yet management explicitly called AI a multi-year “megatrend” in their most recent earnings call, with demand now being pulled not just by chip designers, but directly by hyperscale cloud providers seeking to lock in capacity.  

Management stated: 

“Our customers’ customers, who are mainly the cloud service providers, are also providing strong signals and reaching out directly to request the capacity to support their business. Thus, our conviction in the multiyear AI megatrend remains strong, and we believe the demand for semiconductor will continue to be very fundamental.”Thus, our conviction in the multiyear AI megatrend remains strong, and we believe the demand for semiconductor will continue to be very fundamental.” 

When the world’s most advanced foundry says hyperscalers are coming to them directly for capacity, it signals that AI demand remains foundational. Perhaps most importantly, TSM is not a “flip the switch" business model to where demand can be turned on and turned off quickly. Wafer capacity must be planned years in advance, which makes these signals particularly meaningful. 

While 2nm defines the next phase of the roadmap, 3nm remains the node supporting most AI deployments today. The company’s advanced 3nm node offers roughly 15% better performance than 5nm at equal power and transistor density, with die sizes estimated to be ~42% smaller. TSMC also states the 3nm process can reduce power consumption by up to 30%, underscoring power efficiency as a key competitive advantage. 

This efficiency helps deepen TSMC’s moat. While Samsung introduced 3nm chips in 2022, it has lagged TSMC on yield and power efficiency by an estimated 10%–20%. This advantage is reflected in pricing power, with TSMC charging roughly 25% more for 3nm versus 5nm, as customers are willing to pay a premium to avoid Samsung. 

The company entered volume production of its most advanced node, N2, in 4Q 2025, marking a transition from FinFET to gate-all-around (GAA) transistor architecture. By wrapping the gate around all sides of the channel, GAA improves electrostatic control and reduces leakage versus FinFET designs. 

N2 introduces NanoFlex technology, enabling designers to mix cell types and optimize for performance or power by adjusting nanosheet dimensions. According to management on the Q2 2025 earnings call, N2 delivers 10%–15% speed improvement at the same power or 20%–30% power reduction at the same speed, along with more than 15% chip density gains versus N3E. 

As chips migrate to advanced nodes—such as Nvidia’s Rubin moving to 3nm and AMD building CPUs on 2nm—TSMC stands to continue to benefit from rising pricing power, as these nodes command significant wafer premiums in exchange for material performance and power efficiency gains.

Overall Revenue Growth: 

TSMC reported Q4 revenue of $33.73 billion, up 25.5% YoY and 1.9% QoQ and exceeding guidance range for $32.2 billion to $33.4 billion, and coming in $1 billion ahead of estimates.  

Full-year revenue was $122.42 billion, up 35.9% YoY; TSMC guided for Q1 revenue between $34.6-35.8 billion, up 37.9% YoY and 4.4% QoQ (also outpacing Q1 '25 growth of 35.3% YoY); 2026 revenue guided to be up close to 30% YoY in USD 

AI Segment Growth: 

HPC revenue rose 4% QoQ in NT$ and accounted for 55% of revenue in Q4. For FY25, HPC revenue in NT$ was up 48% YoY to 58% of revenue. Recent development in the AI market continue to be very positive. Revenue from AI accelerator accounted for high teens percent of the total revenue in 2025. 

Earnings: 

GAAP EPS up 40.2% YoY in Q4 to $3.14, beating estimates by 5.2%. FY25 EPS was $10.65, up 51.3% YoY; GAAP EPS is expected to be $3.28 in Q1, up 54.7% YoY while FY26 EPS is currently estimated to be $13.05, up 22.5% YoY (subject to revisions) 

Margins: 

GAAP gross margin in Q4 was 62.3%, well above guidance for 59-61%, and up 2.8 points QoQ and 3.3 points YoY on due to cost improvement efforts, favorable foreign exchange rate and high capacity utilization rate. For Q1, TSMC guided gross margin to be 63-65%, up 1.7 points QoQ and 5.2 points YoY at MP. GAAP operating margin was 54%, up 3.4 points QoQ and 5 points YoY; For Q1, TSMC guided operating margin to be 54-56%, up 1 point QoQ and 5.5 points YoY at MP; Net margin was 48.3%, up 2.6 points QoQ and 5.2 points YoY. 

Cash: 

Q4 operating cash flow was $23.4 billion for a 69.4% margin, down 2 points YoY, and FCF was $11.9 billion for a 35.2% margin, up 5.4 points YoY. Cash of $97.6 billion and debt of $31.6 billion. 

Valuation: 

TSM is trading at a forward P/E ratio of 22.9. The company has traded at a minimum of 13.5 and a maximum of 29.6 in recent years, placing the current valuation near the midpoint of that range. 

Notable Risks: 

TSM carries geopolitical risk that has been muted in recent quarters, yet could heat up again at anytime. 

Memory: The Leading Constraint in AI Systems 

Memory is typically a cyclical industry that is lower margin and lumpy, yet it is seeing a newfound resurgence from AI that is strong enough to transform commoditized hardware into a secular trend as the AI economy is built out. AI servers use more DRAM and NAND than traditional servers, relying heavily on high-bandwidth memory (HBM) for training and inference.   

The HBM market is projected to reach $35 billion this year, doubling YoY, with Micron’s September results confirming that the market was well on track to be over $30 billion as of Q3. Looking ahead, the shift to HBM4 with Nvidia’s Rubin architecture and AMD’s MI400 series will represent another important growth lever come 2026 as HBM content per GPU and per rack surges, paving the way for HBM to potentially triple again by as early as 2028.  

Not only is HBM a focal point due to its rising importance and thus increasing content per GPU, but other memory products are quickly coming to the forefront, notably low-power DDR5 memory (LPDDR5X) and data center solid state drives (SSDs).  

Through 2026 and 2027, the outlook for HBM remains fairly positive, with SK , SK Hynix, Samsung and Micron already selling out of HBM3e and HBM4 capacity through the end of 2026. This underscores the robust demand environment stemming from AI accelerators, with Micron seeing HBM bit shipments outpacing DRAM bit growth, but also may limit revenue upside as prices have been contracted over the next four quarters.   

On pricing, HBM4 is expected to carry a significant premium to HBM3e, currently used for Nvidia’s Grace Blackwell chips. Analysts from UBS had estimated that HBM4’s price premium could be as much as 30%, though reports of Samsung’s discussions over HBM4 supply with Nvidia dwarfed that – Samsung was said to be targeting price parity with SK Hynix on HBM4 around $500, up ~50% from the mid-$300s for HBM3e. These price increases will support strong growth as HBM4 volumes ramp.  

Looking forward, industry analysts project the HBM market to reach $98 billion to $100 billion by 2030, representing a 31.5% CAGR from 2024’s $18 billion, outpacing DRAM’s growth by 3X, which is expected to rise at an 11.7% CAGR to $194 billion. As a result, HBM’s share of DRAM revenue is expected to surpass 50%.   

However, in its Q1 report, Micron said it now expects the HBM TAM to reach $100 billion as early as 2028, two years sooner than its prior forecast. This would represent a ~42% CAGR from $35 billion, or more than 10 points faster than the base case forecasts.   

For information on HBM3e and the shift to HBM4, DDR5 prices surging and the rising demand for memory bandwidth, plus why NAND SSDs are surging, read our December report “The AI Memory Boom has Arrived.”The AI Memory Boom has Arrived.” 

Micron: Memory Market Takes the Crown from Compute on Growth 

You would be hard pressed to find another segment of the AI data center industry posting growth to this degree on a sequential basis. Data from TrendForce estimates that the global DRAM market recorded growth of 30.9% QoQ in calendar Q3 to $41.4 billion. In dollar terms, this represented QoQ growth of ~$9.7 billion, or nearly as large of a QoQ jump as Nvidia reported in its most recent quarter.  This growth was driven by “significant increases in conventional DRAM contract prices, higher bit shipments, and growing HBM volumes.”   

For a supplier breakdown, SK Hynix’s revenue grew 12.4% QoQ to $13.75 billion, fueled by seasonal price increases and significant bit shipment growth. Samsung also reported similar significant growth in bit shipments, with revenue up 30.4% QoQ to $13.50 billion. Micron followed with a substantial 53.2% QoQ increase to $10.65 billion, per TrendForce (note that this is for calendar Q3 which does not align with Micron’s fiscal year calendar).   

As of November, TrendForce estimates that DRAM contract prices will accelerate into Q4, predicting conventional DRAM contract prices will surge by another 45% to 50% QoQ, while total contract prices (which includes HBM) will increase by 50% to 55% QoQ – this is a substantial uplift from projections for 18-23% QoQ growth in Q4 at the end of October.   

Contributing to strong pricing is DDR5 DRAM, where prices rapidly skyrocketed – from late September to early November, prices have as much as quadrupled, with impacts felt most on consumer products. Samsung also reportedly just boosted DDR5 prices by 100%, citing no stock left.   

However, revenue growth in Q4 could potentially be lower than pricing growth as bit shipments are projected to decline sequentially due to rapid inventory depletion. DRAM supplier inventory levels are projected to range between two to four weeks, a major crunch from 5.5 weeks on average last quarter and more than 15 weeks at the start of the year.   

Overall Revenue Growth: 

Micron reported record Q1 revenue of $13.64 billion, beating estimates by 5.9% and accelerating 10.7 points to 56.7% YoY growth. Sequentially, growth was 20.6% QoQ, just one point slower than Q4’s 21.7% QoQ growth.  

Revenue accelerated from 46.1% YoY in FQ4 to 56.7% in FQ1 at $13.64B; FQ2 guidance points to sharp acceleration to 132.2% YoY, QoQ growth to accelerate from 20.6% to 37.1% QoQ. Some analysts are saying this is the biggest headliner beat they’ve seen since Nvidia’s 2023 moment.

AI Segment Growth: 

DRAM products (within that HBM and LPDDR5X) were the primary driver of Q1’s results, with revenue up 69% YoY and 20% QoQ to $10.8 billion, or 79% of revenue.

DRAM revenue up 69% YoY and 20% QoQ to $10.58B in Q1

Micron’s Cloud Memory Business Unit (CMBU), which consists of its HBM, high-capacity dual in-line memory modules (DIMMs), and low-power server DRAM solutions, saw Q1 revenue of $5.28 billion, up 99.5% YoY and 16.3% QoQ

HBM, high-capacity DIMMs and LP server DRAM revenue reached $10 billion as of Q4, up more than fivefold YoY

Earnings: 

In Q1, Micron reported GAAP EPS of $4.60, up 175% YoY; this also is a sharp uptick from $2.83 in Q4.  For Q2, Micron guided for GAAP EPS to be $8.19, +/- $0.20, nearly 74% ahead of estimates for $4.71 and corresponding to YoY growth of almost 481%, a 306 point acceleration. GAAP EPS growth is expected to remain >250% for both Q3 and Q4 to $9.37 and $10.04.  

For the full year, Micron is expected to deliver GAAP EPS of $31.17, more than quadrupling from $7.59 in fiscal 2025. Earnings estimates also moved more than 60% higher following Q1’s report and Q2’s blowout guide, moving from $19.42 to the now $31.17 estimate. 

Margins: 

GAAP gross margin in Q1 was 56%, up 17.6 points YoY, aided by the strong growth in CMBU which carried a 66% gross margin in the quarter. For Q2, GAAP gross margin was guided to be 67% at midpoint, an 11 point sequential expansion and up 31.2 points YoY.  

GAAP operating margin was 45%, up 12.7 points QoQ and 20 points YoY, again aided by CMBU which carried a 55% margin in the quarter. For Q2, Micron implied operating margin to be 58.7%, up 12.7 points QoQ and 36.7 points YoY, signaling strong tailwinds from surging DRAM prices.  

GAAP net margin was 38.4% in Q1, up 10.1 points QoQ and nearly 17 points YoY. 

Cash: 

Operating cash flow was $8.41 billion in Q1, up more than 159% YoY and nearly 47% QoQ. OCF margin was 61.7%, up 10 points QoQ and up 24.4 points YoY.  

Adjusted free cash flow was $3.91 billion in Q1, up sharply from $803 million in Q4 and $112 million in the year ago quarter. Adjusted FCF margin was 28.6%, up from 7.1% in the prior quarter and 1.3% in the year ago quarter.  

Micron reported total cash and equivalents of $12.0 billion and total debt of $11.76 billion. 

Valuation: 

Micron is now trading at peak multiples on the top line as shares continue to rally, currently valued 6.5x forward PS, in line with the highest level it achieved in the summer of 2024, and well above its 3.6x average multiple over the past five years. 

However, on the bottom line, Micron is trading at a much more reasonable 12.8x forward PE multiple due to the strong margin expansion and expected 300% earnings growth this fiscal year to $33+. This is notably below Micron’s 2025 peaks around 16x forward PE. 

Notable Risks: 

Sharply rising DRAM prices from tight supply could cut into demand for consumer electronics products, which is Micron’s second largest segment and growth driver (Mobile and Client) with nearly $4.3 billion in revenue and a 47% operating margin in Q1. Any demand softness from price hikes could be felt more acutely in 2026 with forecasts now pointing to smartphone and PC shipments declining YoY. 

SanDisk: Marketing-Leading Returns in 2026; Can the Stock Repeat? 

Overview: 

On a broader level, data center/enterprise SSDs are often overlooked but equally critical as HBM when it comes to AI training and inference. This is because data center SSDs offer higher read-write speeds critical for accessing and transferring data rapidly, along with higher performance and energy efficiency, vital factors for larger-scale AI training and inference workloads.    

Nvidia is positioning NVMe SSDs to become the backbone for the Inference Context Memory Storage architecture discussed at CES, there is the potential for SSD suppliers to see solid medium/long-term tailwinds from increased SSD capacity requirements in inference-optimized deployments over the next few years.  

For example, Bernstein estimates that Huang’s CES comments on SSDs and KV cache requirements suggest an additional 16TB per GPU, compared to 3-4TB per GPU today, or 4-5X growth. This will be more weighted towards year-end and into 2027 as ICMS rolls out with Rubin.   

Similar to DRAM, data center SSD shipments and prices were strong in Q3, driven by hyperscaler demand for AI infrastructure and general-purpose servers. Revenue from the top five companies – Samsung, SK, Micron, Kioxia and SanDisk – rose ~28% QoQ to a new record at $6.54 billion, per TrendForce. Notably, this was broad-based strength, with growth at the five firms all ranging between 26-30% QoQ.    

For Q4, there are a few dynamics in play that are likely to keep prices and thus revenue growth strong. For example, supplier inventories are expected to have fallen sharply, from 10-15 weeks in early Q3 to just 7-10 weeks at the start of Q4, which was said to be ‘below healthy levels’, with enterprise SSD supply growth substantially lagging demand. SanDisk says that its storage-focused SSD is “growing in demand with 2 hyperscaler qualifications underway and a third hyperscaler along with a major storage OEM planned for calendar year '26.”  

In November, TLC and QLC SSDs reportedly experienced strong price increases, with 1 TB TLC SSDs seeing sharp increases and the “most significant shortage due to persistent enterprise SSD demand.” 512 GB TLCs were estimated to see the most significant price hikes at ~65% MoM, while the QLC supply chain tightened and forced prices higher.   

Additionally, TrendForce points out that these inventory and demand dynamics mean “supply shortages in 2026 are becoming increasingly apparent,” providing an additional lever for SSD prices to rise through next year and support more revenue growth as long as inventories and bit shipments do not hinder that.   

Overall Revenue Growth: 

SanDisk reported a strong sequential revenue acceleration in its fiscal Q1, driven by NAND demand outpacing supply and increasing demand in its data center, edge and consumer end markets. Q1 revenue increased 22.6% YoY and 21.4% QoQ to $2.31 billion, accelerating from 8% YoY and 12.2% QoQ growth in fiscal Q4. Higher-than-expected bit growth drove the outperformance in the quarter relative to guidance of $2.1-2.2 billion, per management. Next quarter is expected to see 16.5% QoQ at the $2.69 billion consensus. 

AI Segment Growth: 

SanDisk’s data center revenue, as mentioned above, declined (10%) YoY but rose 26% QoQ to $269 million, driven by increasing demand for its ‘Stargate’ enterprise SSD product line. However, revenue contribution remains small, at less than 12% of revenue.   

Management also increased their forecast for data center exabyte growth, explaining that last quarter, exabyte growth expectations were in the mid-20% range, but now are in the mid-40% range. As a result, data center is expected to be the largest market in NAND on an exabyte basis in 2026, surpassing mobile.   

SanDisk’s Edge segment was the primary growth driver in Q1 with revenue up 30% YoY and 26% QoQ to $1.39 billion, driven by increasing NAND content in PCs and smartphones and a positive PC refresh cycle. Consumer revenue rose 27% YoY and 11% QoQ to $652 million, while data center revenue was down (10%) YoY but up 26% QoQ to $269 million.  

Earnings: 

SanDisk stands out for its strong expected earnings growth through fiscal 2026 and fiscal 2027, with adjusted EPS expected to reach more than $21 by then, or >7X higher than the $2.99 it earned in fiscal 2025.   

Q1 GAAP EPS was $0.75, a strong improvement from a ($0.16) loss in Q4, though this was down (49%) YoY from $1.46 in the year ago quarter as margins remained lower YoY. Adjusted EPS was $1.22, up 321% QoQ but down (33%) YoY.   

For Q2, SanDisk guided for adjusted EPS of $3.00 to $3.40, up more than 162% QoQ. Adjusted EPS is expected to further increase to $3.78 in fiscal Q3 and $4.82 in fiscal Q4.    

For fiscal 2026, SanDisk is expected to generate $13.29 in adjusted EPS, up 344.6% YoY, while GAAP EPS is projected to be $11.53, up from ($11.32) in FY25 due to the spin off. Fiscal 2027 is expected to see earnings power surpass $21, with GAAP EPS estimated to be up 86% to $21.47 and adjusted EPS up nearly 62% to $21.50. 

Margins: 

Margins are lower YoY compared to pre-spinoff margins, but Q1 saw strong sequential margin expansion that is expected to accelerate in Q2.    

Q1 GAAP gross margin was 29.8%, down 8.8 points YoY but up 3.6 points QoQ. Adjusted gross margin was 29.9%, down 9 points YoY but up 3.5 points QoQ.   

GAAP operating margin was 8.3%, down 8.3 points YoY but up 5.6 points QoQ. Adjusted operating margin was 10.6%, down 8.2 points YoY but up 5.3 points QoQ.   

For Q2, SanDisk guided adjusted gross margin to be 41-43%, or up just over 12 points QoQ at midpoint on higher pricing and cost reduction tailwinds, while adjusted operating margin is implied to be 24.2% at the midpoint of opex guidance, or up 13.6 points QoQ. Fab startup costs are expected to transition from headwinds to tailwinds during the quarter, potentially aiding more margin expansion into fiscal Q3 and Q4. 

Cash: 

Operating cash flow was $488 million in Q1 for a 21.1% margin, up from a (7%) margin in the year ago quarter and a 4.9% margin in Q4.   

Adjusted free cash flow was $438 million in Q1 for a 19% margin, up from a (10.5%) margin in the year ago quarter and 2.6% in Q4.   

SanDisk’s total gross capex to support the JV was $387 million in Q1, though its cash capex spend was only $40 million (1.7% of revenue) as the remainder was funded through external sources such as subsidies or tool depreciation recorded in COGS.  

Cash and equivalents totaled $1.44 billion while debt totaled $1.35 billion.  

Valuation: 

SanDisk’s valuation is somewhat hard to pin down given the company’s limited history on the public markets after its February spinoff, and its 1,000% rally in the past six months. On the top line, SanDisk is trading at 6.6x forward PS, having traded as low as 0.6x last summer and with an average multiple of 1.6x for its limited public history.  

On the bottom line, SanDisk is trading at 36.3x forward PE, having traded as low as 3x last August with an average around 13.3x. 

Notable Risks: 

The NAND flash market has historically been quite volatile, and is shifting from significant oversupply in 2023 to expectations for substantial supply shortages through 2026. However, if NAND capacity begins to come online quickly through next year, or if demand for PCs and smartphones falters due to rising memory prices, the NAND cycle could reverse and lead to pricing pressures cutting into revenue growth and margins. SanDisk also has limited AI data center exposure, contributing with <12% of revenue last quarter. 

AI Networking Stocks

Please refer to the section above entitled “Rubin Redefines AI Networking as a Bandwidth-First Constraint” for an update on the AI Networking trend, which is a Top 3 trend for the I/O Fund in 2026. 

Lumentum: EMLs Power 400G/800G Transceivers as Networking Scales 

EMLs are a critical component with Nvidia’s Blackwell generation, as the scale-up in GPU counts per rack from eight to 72 and subsequent increases in bandwidth and switch density will require low-power, efficient high-speed optics. The power advantages over SiPho also come to the forefront as power consumption becomes a central concern in scaling AI data centers, with Blackwell doubling power consumption versus Hopper at 140kW per rack. 

EMLs are the main driver for Lumentum’s growth as these are good for short-to-medium reach inside data centers up to 2km and a strong choice for 400G and 800G optical transceivers, with the company having begun its 100G EML ramp for these data rates in early 2024. EML laser shipments reached a fresh record in fiscal Q1 2026, driven once again by 100G speeds and an increase in 200G shipments. 

Lumentum’s Q1 provided more confirmation that EML laser shipments are ramping in full force, with another record quarter driven by 100G speeds and an increase in 200G shipments. EMLs have been the primary driver of growth so far for Lumentum, though the supply-demand imbalance is widening due to tight indium-phosphide (InP) capacity. Looking ahead to 2026, InP capacity will be a key factor to focus on as Lumentum is targeting 40% capacity growth over the next few quarters, with the potential for this to drive even stronger revenue growth. 

One important discussion on EMLs is that the supply-demand imbalance continues to widen, meaning that substantial growth in capacity through 2026 should quickly convert to revenue. CEO Michael Hurlston explained that “last quarter, I think we characterized it as roughly a 20% shortfall relative to total customer demand. Even with the add in supply, I would say that number has increased to 25% to 30%. We are quite a bit short right now relative to the customer demand.” 

On the positive side, Lumentum shared that while its indium phosphide fab is fully allocated due to high demand, it has made “better-than-expected progress on yields and throughput and now see a line of sight to add approximately 40% more unit capacity over the next few quarters.” CEO Michael Hurlston clarified at UBS’ tech conference that “we gave in the last earnings call a new benchmark saying, over the next 3 quarters, meaning our December, March and June quarters, we expected to add that 40%. So that’s a forward-looking statement where we’d expect an increase in capacity of 40% on what already is a doubled number.” 

Management expects to be well positioned for both EML and CW lasers ramping for 1.6T transceivers, as its capacity is interchangeable between the two components, despite management noting a difficulty in forecasting how the two will ramp – the primary takeaway here is that even if faster data rates such as 1.6T are less dependent on EMLs, management believes there is more than enough content for them to do well. 

Overall Revenue Growth: 

Lumentum fulfilled its guidance for a >$500 million revenue quarter in calendar 2025, reporting a record $533.8 million in revenue in fiscal Q1, beating estimates by just 1.4%. Revenue growth accelerated 2.5 points to 58.4% YoY through QoQ growth slowed to 11%. Lumentum guided for $630 to $670 million in revenue in Q2, accelerating to 61.6% YoY and 21.8% QoQ, whereas consensus estimates were pegged at almost 40% growth to $561.5 million.  

On the financials side, the number one item was Q2’s impressive 22% QoQ revenue growth guide to $650 million at midpoint. This is significant as Lumentum is reaching its $600 million quarterly revenue target two quarters ahead of schedule, with this also marking its highest revenue in company history. The 22% QoQ guide would also reflect Lumentum’s fastest sequential growth since the September 2020 quarter.  

To put in perspective how strong Lumentum’s growth curve is, current estimates for the June 2026 quarter sit at $740.3 million, more than 23% ahead of the company’s target revenue. This is also up from $689.9 million on November 7, a 7.3% revision higher in less than one week.  

As discussed previously, Lumentum guided for $630 to $670 million in revenue in Q2, accelerating to 61.6% YoY and 21.8% QoQ, whereas consensus estimates were pegged at almost 40% growth to $561.5 million. 

AI Segment Growth: 

Components revenue rose 18.4% QoQ and 63.9% YoY to $379.2 million, fueled by “robust demand inside the data center”, strong momentum for DCI products, and record EML shipments. Looking through 2026, Lumentum expects another breakout year for laser chip shipments, supported by “better-than-expected progress on yields and throughput” providing “line of sight to add approximately 40% more unit capacity over the next few quarters.” Management added that they also “expect a significant increase in shipment volumes in the second half of calendar 2026” for ultra-high power laser assemblies, which are currently in the initial ramp phase.  

Management provided a deeper discussion on margins moving through 2026, with product pricing from supply-demand imbalances serving as a strong lever for margin expansion:  

“I think we're moving the margin line up. Pricing, obviously, is a lever. And when you look at that very, very carefully, I think what you see in the guide is some pricing, very targeted price increases happening. I think as you look out next year in 2026, our agreements with customers will include more pricing, more broad-based price increases, just given the supply-demand imbalance.” 

Earnings: 

Lumentum reported a razor thin $0.05 in GAAP EPS, while adjusted EPS of $1.10, up 511% YoY, beat estimates by 6.8%. 

For Q2, Lumentum guided for adjusted EPS in a wider range of $1.30 to $1.50, up 233% YoY, coming in well ahead of the $1.16 estimate at the midpoint. 

Lumentum did not provide a full year adjusted EPS guide, though consensus now sits at $5.35, up from $4.90 and pointing to growth of 160% YoY. 

Margins: 

GAAP gross margin was 34.0%, in Q1, up nearly 11 points YoY and 0.7 points QoQ. Adjusted gross margin was 39.4%, up 6.6 points YoY and 1.6 points QoQ. 

GAAP operating margin was 1.3%, up nearly 26 points YoY and 3 points QoQ. Adjusted operating margin was 18.7%, up 15.7 points YoY and 3.7 points QoQ, ahead of guidance for 16-17.5%. For Q2, management guided for continued expansion to 20-22%. 

GAAP net margin was 0.8%, up 25.3 points YoY and not comparable QoQ due to an income tax benefit in Q4. Adjusted net margin was 16.2%, up 12.6 points YoY and 3 points QoQ. 

Cash: 

Operating cash flow was $57.9 million in Q1 for a 10.8% margin, down from 11.8% a year ago and 13.3% in Q4. Free cash flow was ($18.3 million) for a (3.4%) margin, up from (10.2%) a year ago but down from 2.1% in Q4. 

Cash and equivalents were $1.12 billion while debt was $3.24 billion. 

Valuation: 

Lumentum is valued at peak multiples, with shares now trading above a 10x forward PS multiple, up from the 4x range in September and October. This is also 3x its five-year average forward PS of 3.3x and at peak levels.  

On the bottom line, Lumentum is trading at 64.6x forward PE, above its prior peaks around 50x and above its 40.6x average over the past five years. Similar to the top-line, shares have seen a pretty rapid expansion from the 25-30x range in October.  

Notable Risks: 

Lumentum has many competitors in the optical transceiver space, while navigating rather severe InP and EML capacity shortages may pose a near-term challenge as the supply-demand imbalance continues to widen. Cash flows are also thin with FCF negative, and debt is around 3x of cash.  

Coherent: InP Capacity to Double 

Overview: 

Coherent is not nearly as flashy as Lumentum when it comes to revenue growth or even data center growth, yet the company is sitting in a prime position moving through 2026 as the industry navigates extremely tight indium phosphide (InP) capacity coupled with elevated demand for InP-based EML lasers. This is because Coherent is preparing to double indium-phosphide capacity via a multi-faceted expansion plan with multiple facilities ramping output in unison, while shifting to a larger wafer size that can deliver 4X output per wafer at half the cost.  

This dynamic is expected to help drive a reacceleration in Coherent’s data center segment to 10% QoQ growth next quarter, a notable uplift from 4% this quarter, along with margin expansion driving solid adjusted EPS leverage. Management also stated they expect “strong sequential growth through the balance of this fiscal year given very strong demand and improving supply.”  

On the product side, Coherent sees strong demand for both its 800G and 1.6T transceivers, with 1.6T expected to drive a significant portion of the guided sequential growth. This first wave of 1.6T growth is expected to be split between both EML-based and CW laser-based silicon photonics transceivers, with Coherent able to benefit from both as it can quickly shift capacity for whichever customers prefer.  

For Coherent’s AI-related revenue exposure, Datacenter and Communications account for ~69% of total revenue. This also includes some contribution from telecom so is not an exact figure yet provides a rough idea as to Coherent’s AI exposure. 

Revenue: 

Coherent delivered 17.3% YoY and 3.4% QoQ revenue growth in fiscal Q1 to $1.58 billion, beating estimates by nearly 3%. On a pro-forma basis excluding the $33 million in Q1 revenue from the now-divested Aerospace & Defense unit, revenue growth was 19% YoY and 6% QoQ.  

For Q2, Coherent guided for revenue between $1.56 billion to $1.70 billion, which on the headline figure would be decelerating to 13.6% YoY and 3.2% QoQ at midpoint, before reaccelerating to 15.9% by Q4. 

However, our internal pro-forma estimate shows a better trajectory for revenue through fiscal 2026 – pro-forma growth may decelerate slightly to the 17.4% YoY and ~5.7% QoQ in Q2, before reaccelerating to nearly 21% by Q4, the highest growth rate in the past five quarters. 

AI Revenue: 

Coherent’s Datacenter and Communications revenue rose 26.2% YoY and 7% QoQ to $1.09 billion, accounting for ~69% of revenue. Growth has decelerated rather steadily since Q1 FY2025’s 68% YoY print. 

Datacenter revenue rose 4% QoQ and 23% YoY. As mentioned previously, Datacenter growth was constrained by InP laser supply, with management expecting QoQ growth to accelerate to 10% in Q2 and remain strong through the end of the fiscal year. 

Communications revenue, which includes telecom and data center interconnect (DCI) rose 11% QoQ and 55% YoY, driven primarily by DCI products. Management said they witnessed strong growth in demand for ZR/ZR+ DCI products, with 100G, 400G and 800G products expected to continue ramping through fiscal 2026. 

Earnings: 

Fueled by margin improvements, Coherent reported a solid adjusted earnings beat in Q1, with adjusted EPS rising 73% YoY and 16% QoQ to $1.16, beating estimates by 11.3%. 

For Q2, Coherent guided for adjusted EPS between $1.10 to $1.30, decelerating sharply to 26.3% YoY at the $1.20 midpoint, and only showing a small sequential improvement. 

Margins: 

Coherent made solid progress on the margin front and expects gross margins to strengthen towards 42% with the ramp of its 6-inch InP wafers and higher margin 1.6T transceivers, and continued cost cutting measures. 

GAAP gross margin was 36.6%, expanding 2.5 points YoY and 0.9 points sequentially. Adjusted gross margin came in at 38.7%, above the midpoint of guidance for 37.5-39.5%, expanding two points YoY and 0.6 points sequentially. Management said the gross margin expansion was driven by “cost reductions and product input costs as well as yield improvements,” while pricing optimization was also a meaningful contributor.  

GAAP operating margin was 16.4%, up nearly 11 points YoY and 16 points QoQ, though this was impacted by a $115 million gain from the Aerospace divestment. Adjusted operating margin was 19.5%, up 3.4 points YoY and 1.5 points QoQ.  

GAAP net margin was 14.3%, up 12.4 points YoY and more than 21 points QoQ; adjusted net margin was 14%, up 3.8 points YoY and 1.4 points QoQ. 

Cash: 

Cash flows were also thin with OCF margin down nearly 10 points YoY, and FCF widened deeper into negative territory due to capex for the upcoming capacity expansion. 

Operating cash flow was $46 million in Q1, down from $130.3 million in Q4 and the first time falling below $100 million in the past seven quarters. OCF margin was 2.9%, down from 11.4% a year ago and 8.5% in the prior quarter. 

Free cash flow was ($57.9 million), widening from ($1 million) in Q4 and a stark contrast to $61 million in the year ago quarter, driven by capex of $103.9 million. FCF margin was (3.7%), widening from (0.1%) in the prior quarter and down from 4.5% a year ago. 

Cash and equivalent totaled $852.8 million, while debt was $3.31 billion, down from $3.69 billion in the prior quarter  

As a result, Coherent has made substantial progress on its debt leverage ratio, paying down $400 million in debt in Q1. On that note, Coherent’s debt has declined approximately $1 billion over the last two years, from $4.29 billion in Q1 FY24 to $3.31 billion this quarter – a nearly 23% reduction.  

Coherent’s debt leverage ratio has now improved to 1.7x, down from 2x in the prior quarter and 2.4x a year ago 

Valuation: 

Similar to Lumentum, Coherent is trading at peak multiples on the top and bottom line. Shares are valued at 5x forward PS, more than double its 2.2x average over the last five years and a significant discount to Lumentum’s 10x multiple likely due to Coherent’s lower growth.  

On the bottom line, Coherent is trading at 42.5x forward PE, just slightly below its June 2024 peak at 45x, though this is also well above its five-year average forward PE of 26.3x. 

Risks: 

Coherent’s data center revenue growth was soft in Q1 at 4% QoQ, though management expects this to return to 10% QoQ in Q2 and remain strong, thus the company needs to execute on this given the multi-faceted tailwinds from 1.6T transceiver demand and InP capacity expansion.   

Astera Labs: Scorpio-X Set to Provide a Boost Amid Tough Comps 

Overview: 

In an effort to identify a catalyst that can sustain Astera’s exceptional growth, it would be this product that does so. The X-series is used to interconnect GPUs for higher GPU utilization, resulting in higher ASPs. 

Regarding the X-Series: “And this one, like Mike noted, it's a greenfield use case, meaning if you keep Nvidia and NV Switch aside, everyone else is starting to build configurations that are obviously going to need some kind of a switching functionality, which is what we are addressing with our X Series device.”  

And on that basis, the X-Series will always be a much more valuable, much more higher ASP product than a P-Series.” 

Notably, Astera maintains their largest opportunity for the X-Series is on the custom silicon side although they foresee hyperscalers wanting to customize their racks in a way that prevents vendor lock-in from both Nvidia and Broadcom.  

Regarding Ethernet Scale-up Networking (ESUN), ESUN is attempting to make Ethernet work for scale-up whereas UALink was built from scratch for scale-up. The primary benefit ESUN offers is to move quicker than UALink (in the most recent earnings report, ALAB stated it’ll be 2027 for UALink to be fully deployed).  

However, in the meantime, Astera’s PCIe solutions are in high demand and deployable now. Even if ESUN moves faster commercially, there is a performance gap that helps to ensure that Astera’s positioning with PCIe/CXL remains intact. That performance gap is best described as the low latency required for what are the most in-demand AI workloads today – those that require memory pooling and GPU-to-GPU communication.   

For more information on how the relevancy of PCIe will persist, read more information on this topic under the Top 3 trends section under AI Networking. 

Revenue: 

Revenue grew by 103.9% YoY and 20.1% QoQ to $230.6M, beating estimates by 11.7%. This maintained Q2’s sequential growth rate of 20%, though YoY decelerated by ~46 points as the company begins to lap tougher comps on a dollar basis.  

For Q4, Astera guided for $245 million to $253 million in revenue, coming in well ahead of estimates for $216.5 million and pointing to YoY growth of 77% and QoQ growth of 8%, driven by continued PCIe 6 momentum and robust growth from Taurus Ethernet SCMs. This would technically mark the company’s first <100% growth quarter since the end of 2024. 

AI Revenue: 

Scorpio P-Series represents 10% of revenue now, yet management stated it will quickly double to exit the year at 20% of revenue. From there, management has implied Scorpio X will exceed Scorpio P’s revenue percentage. Net-net, that means Scorpio will reach 50% of revenue sometime in H1 2026 up from effectively 0% of revenue in H1 2025.  

The longer refresher on Scorpio P-Series and Scorpio X-Series is necessary because the primary catalyst we identified earlier this year has not even ramped yet. Scorpio P-Series only began shipping this quarter and Scorpio X-Series will begin to ship next year. 

Earnings: 

Adjusted EPS grew by 113% YoY to $0.49, beating estimates by 25.6%. GAAP Operating Margin Expands ~32 Points YoY to 24%.  

With the strong expansion in GAAP net margin, Astera delivered a 92.3% beat on GAAP EPS, reporting $0.50 in Q3 versus the $0.26 estimate. Adjusted EPS was $0.49, up 113% YoY and solidly ahead of the $0.39 estimate.  

For Q4, Astera guided for $0.20 in GAAP EPS, below the $0.26 estimate due to a 45% income tax rate. Adjusted EPS was guided at $0.51, up 38% YoY. This guidance would bring FY25 GAAP EPS to $1.17 (versus estimates for $0.96) and adjusted EPS to $1.77 (versus estimates for $1.58). 

Margins: 

Operating margin improved 31.9 percentage points YoY to 24% and adjusted operating margin improved by 9.3 percentage points YoY to 41.7% driven by strong operating leverage.  

GAAP gross margin was 76.2%, ahead of guidance for 75%. This marked a marginal 0.4 point sequential improvement but a 1.5 point YoY contraction. Adjusted gross margin was 76.4%.  

GAAP operating margin was 24.0%, well ahead of guidance for 17.9%, and expanding 3.3 points QoQ and nearly 32 points YoY. This YoY expansion from (7.9%) in Q3 ’24 is quite impressive considering the company was reporting triple-digit revenue growth in each quarter; this also reinforces that the company is comfortably GAAP profitable. Adjusted operating margin was 41.7%, up 2.5 points QoQ and 9.3 points YoY.  

For Q4, Astera guided for slight sequential moderation in margins down the line, with gross margin guidance at 75%, in line with prior quarter guidance. GAAP operating margin was guided to be 22.2% at midpoint, down 1.8 points QoQ but still up more than 22 points YoY. Adjusted operating margin was guided at 39.9%, down 1.8 points QoQ but up 5.6 points YoY. 

Cash: 

The company free cash flows grew by 41% YoY to $65.8M. Cash of $1.13B and debt is Nil.  

Cash flow margins contracted sharply Q3, though this was primarily driven by a large QoQ increase in accounts receivable, providing an extra layer of confidence in the upcoming revenue acceleration in the next couple of quarters.  

Operating cash flow was $49.1 million for a 6.4% margin, though OCF margin had been >22% for the past five quarters. The sharp contraction was primarily due to a $161 million sequential increase in accounts receivable.  

Free cash flow was $2.4 million for a 0.3% margin, down from 20.2% in the prior quarter due to the jump in AR. 

Valuation: 

Unlike many of the other networking stocks on this Top 15 list, Astera is below its average multiples, with shares nearly one-third off the highs. On the top line, Astera trades at 23.8x forward PS, around 10% below its five-year average of 26.6x and far below its 50x peak at the highs of $250 in September.  

On the bottom line, Astera trades at 71.5x forward PE, nearly 13% below its 82x five-year average multiple, with shares having traded as high as 140x in September and as low as 30x last April. 

Risks: 

Astera has faced some fears in the past that ESUN will become a viable third option due to the familiarity of Ethernet, though PCIe solutions remain in high demand and likely will remain relevant in next-gen GPU systems. On the financials, Astera does have to face decelerating growth rates from tougher comps, with current consensus pointing to 77% growth next quarter to 36% by the end of this year.  

SiPho Stock Could See 8X increase in Orders 

Our Advanced Market Signals members received an analysis and real-time trade alerts for a supplier of optical modules that has outlined plans to expand capacity for 800G and 1.6T products by 8.5× by year-end. Management reiterated on both Q1 and Q2 earnings calls that the expansion remains on schedule. Equipment ordered earlier this year has begun arriving, and production is expected to scale through the second half. 

This expansion stands out in the context of an industry that is growing materially, but not at that rate. Industry demand for 800G and 1.6T optics is generally expected to grow at a multiple closer to 3× this year. A capacity ramp that exceeds industry growth implies a strategic effort to capture incremental share as volumes move higher in 2025 and into 2026. 

To learn more—including how this company is rapidly expanding its facilities, why that capacity supports a faster ramp than the broader networking trend, and the resulting revenue implications—join Advanced Market Signals today.Advanced Market Signals today. Members receive real-time trade alerts, access to the I/O Fund’s momentum stock list (including this silicon photonics name), and weekly webinars every Thursday at 4:30 p.m. ET. To Join Advanced with 30% off, please click here to email usplease click here to email us

Key Supplier to the Next Ethernet Upgrade Cycle 

On the I/O Fund’s Discovery tier, we recently covered a Broadcom networking supplier that sits at the center of Broadcom’s Ethernet roadmap, supplying customized systems to two major hyperscalers plus a major deal with OpenAI for 2027. 

Growth opportunities are primarily centered around its high-bandwidth Ethernet switch portfolio focused on back-end networking, with the company being the leading supplier with 41% share of the >200G switch market through Q2, and with 55% share of the custom switch market (up from 40% in 2024).   

The back-end networking positioning is important for this Key Supplier stock as it means the company is exposed to the faster-growing segment of Ethernet switching – the back-end TAM is forecast to grow at a 56% CAGR through 2029 on scale-out, and potentially soon, scale-up demand, whereas front-end (user-facing) is forecast to grow at a 20% CAGR. 

The next leg of growth is expected with the transition to 1.6T switches, which will introduce higher system complexity, new cooling architectures, and expanded content per rack. Initial customer ramps are expected to begin in late 2026, with broader adoption unfolding through 2027. 

To learn more about this company’s positioning within high-bandwidth AI networking and how it fits into the upcoming 800G and 1.6T upgrade cycle, access the full write-up in the Discovery tier. To subscribe to Discovery with 40% off, click here to email usclick here to email us or email premium@io-fund.com and mention code DISCOVERY40DISCOVERY40

AI Energy Stocks 

Please reference above under our Top 3 trends for thematic commentary on Tech’s biggest bottleneck: AI Energy. 

Bloom Energy: 

Overview: 

Bloom Energy needs no introduction to our Research Members as it was one of our biggest winners last year with a return of 376%. Knox carefully layered in at the lows, outperforming Bloom’s 2025 returns of 291%. Even when our trims proved to be too conservative, we gladly bought back near the levels we sold. 

Grid interconnection timelines are now misaligned with AI deployment timelines. Utilities often project power delivery in 2028–2029, while hyperscalers need capacity in 2025–2027. Bloom Energy is bridging an important power gap in data center expansion as grid access and delays is becoming a limiting factor. What they offer is onsite power generation through solid oxide fuel cells that are behind the meter to reduce dependency on the grid. 

The thesis can be summarized in three words: time to power. Here is what management described as to the competitive advantages regarding time to power for solid oxide fuel cells: “A big shift in our business today is time to power. We are providing solutions to meet the urgent needs of our customers who cannot fulfill their power needs from the grid. In these cases, we rapidly book, build, ship, install and power sites for our customers in a matter of months, a much faster timeline than a grid connection.” 

For example, over the past quarter, Bloom stood power up for Oracle in 55 days – lightning fast compared to other power solutions. The company counts one massive energy partner Brookfield, two hyperscalers and one neocloud as customers (ORCL, AWS via AEP and CRWV) plus they hinted of a fourth large customer in the previous earnings call via a gas company partnership. Additionally, Brookfield is a capital partner that can bring Bloom Energy from the MW-level to the GW-level. 

Higher utilization rather than relying only on capacity growth can also help drive higher revenue for Bloom. For example, there was a hint on the last earnings call that despite doubling capacity, Bloom may be able to expand revenue by 4X over the coming quarters, stating: “As we have previously announced, we are doubling our capacity to 2 gigawatts by December 2026, which will support about 4x our 2025 revenue. That expansion is all systems go. Bloom's capacity will not be a bottleneck for our customers.” 

Overall Revenue Growth: 

Bloom smashed analysts' revenue estimates by 21.3%. The company reported record revenue of $519.05 million, versus estimates of $428.07 million. Revenue grew by a solid 57.1% YoY and 29.4% sequential growth, accelerating 37.6 percentage points from the previous quarter’s YoY growth of 19.5%. 

AI Segment Growth: 

Products revenue grew by 64% YoY to $384.3 million, accelerating from the 31% growth in Q2.  

Installation revenue growth spiked 105% YoY to $65.78 million, accelerating from a (13%) decline in Q2 

Earnings: 

GAAP EPS came at ($0.10) in Q3 compared to ($0.06) in the same period last year. GAAP EPS was negatively impacted by a one-time loss related to unconsolidated affiliates of ($19.6 million) or a ($0.08) per share.  

The company reported adjusted EPS of $0.15, beating estimates by 47%, and was up from ($0.01) in the same period last year and $0.10 in the previous quarter. Bloom reported strong profits growth driven by operational efficiency, product cost improvements, and operating leverage.  

Analysts expect adjusted EPS of $0.31 in Q4 and $0.04 in Q1. Looking forward, adjusted EPS is expected to grow strongly by 84.7% YoY to $0.93 in 2026 and 122.4% to $2.07 in 2027. 

Margins: 

Q3 gross profits grew by 92.7% YoY to $151.68 million or a gross margin of 29.2%, up 5.4 percentage points YoY and 2.5 percentage points sequentially. Similarly, adjusted gross margins showed strong YoY and sequential improvement, primarily driven by product cost improvements and manufacturing efficiencies.  

Operating margins improved 4.4 percentage points YoY and 2.4 percentage points sequentially to 1.5%, primarily driven by strong operational efficiencies. Adjusted operating profits grew by 470% YoY to $46.2 million or an adjusted operating margin of 8.9% compared to 2.5% in the same period last year and 7.1% in the previous quarter. 

Cash: 

Q3 operating cash flows were $19.67 million or 3.8% of revenue compared to ($69.5M) or (21%) of revenue in the same period last year. Operating cash flow improvement was primarily driven by higher profits and working capital improvements.  

Strong operating cash flows also led to higher free cash flows. Q3 free cash flow was $7.4 million or 1.4% of revenue compared to ($83.8 million) or (25.4%) in the same period last year. 

Valuation: 

Bloom Energy is trading at a forward P/S ratio of 13.8. The company has traded at a minimum of 1.4 and a maximum of 17.8. Bloom Energy is trading at premium valuation as the company is a key player in solving the AI data center power bottleneck. 

Notable Risks: 

The valuation is a risk, yet we are less concerned as Bloom Energy is a key beneficiary of the AI-driven energy demand. 

GEV: Nat Gas Behemoth – Boring but Steady 

GE Vernova is the world’s largest gas turbine supplier at 25% ahead of Schneider at 24%. Even still, GEV nearly tripled its gas turbine equipment this past quarter – a statement that has us sitting up in our seats. Per the earnings call: “Power orders grew 44%, led by Gas Power equipment nearly tripling year-over-year.”  

Also, consider that we have been covering Bitcoin miners and other energy sources that can quickly help hyperscalers secure powered shells in the 1GW to 3GW range – yet GEV has 50 GW in backlog for gas equipment contracts with expectations the backlog will reach 60 GW by the end of this year. In other words, the chances that GEV is not a significant player in supplying energy to data centers for many years to come is nil.   

In a bid to supply options quickly to alleviate bottlenecks, GEV is also shipping aeroderivative gas turbine packages and doing extensive R&D on a small modular reactor (SMR) design. As detailed below, how exactly GEV evolves to solve the crucial bottleneck around AI power consumption is not set in stone, rather the company is experimenting rapidly with how to leverage their deep experience in natural gas, electrification and renewables like wind to meet global demand.

Overall Revenue Growth: 

GE Vernova Q4 revenue grew by 3.8% YoY to $10.96 billion, beating estimates by 7.1%. Organic revenue grew by 2% YoY to $10.8 billion. The company is a major beneficiary of the increasing energy requirements from the global AI infrastructure build-out, positioning the company as a key beneficiary of this secular trend. The continued slowdown in the Wind segment was offset by the growth in power and electrification segments that are benefitting from rising electricity consumption driven by data centers and artificial intelligence demand.  

The company’s revenue growth is expected to accelerate to 9.8% YoY growth to $8.8 billion in Q1 and is expected to grow 7.8% YoY to $9.82 billion in Q2 2026. 

AI Segment Growth: 

Q4 power orders increased 77% YoY to $11.7 billion, driven primarily by a sharp acceleration in gas power equipment orders, which more than tripled on higher volumes and favorable pricing. Gas turbine orders rose 71% YoY to 41 units, while power services orders grew 15%, reflecting continued customer investment in existing fleets. 

Q4 power segment revenue grew organically by 5% YoY to $5.7 billion. Management expects high single digit organic growth in Q1. 

Electrification orders were 2.5x revenue and were up 50% YoY to $7.4 billion primarily due to growing grid equipment demand, particularly for synchronous condensers, substations partially to support data center growth and switchgear. The company also witnessed strong equipment orders growth in the Middle East, which increased over $1 billion and in North America, which more than doubled YoY. 

Q4 organic electrification revenue grew by 32% YoY to $2.9 billion primarily driven by strong growth in switchgear and HVDC equipment. Management expects a similar revenue as Q4 in the next quarter which will also include Prolec GE. 

Due to a sudden surge in AI-related electricity demand, the company’s turbine orders are vastly outpacing demand, and the company’s order book is sold out through 2028. 

Earnings: 

Q4 GAAP EPS was $13.39, up from $1.73 in the prior-year period, reflecting a one-time tax benefit of $10.58. Excluding this benefit, GAAP EPS would have been $2.81, below the consensus estimate of $3.13, primarily due to losses in the Wind segment. 

Analysts expect strong EPS growth in the coming quarter with Q1 EPS expected to grow 127.7% YoY to $2.07 and Q2 EPS to grow 65.1% YoY to $3.07. 

Margins: 

The company’s adjusted EBITDA grew by 7.4% YoY to $1.16 billion with an adjusted EBITDA margin of 10.6%, an improvement of 250 basis points sequentially and 40 basis points YoY. Organic adjusted EBITDA margin improved 10 basis points YoY to 10.7%. 

2025 adjusted EBITDA margin improved 260 basis points YoY to 8.4% and was in-line with the management mid-point guidance of 8.5%. Management expects 2026 adjusted EBITDA margin to improve to 12% in 2026 driven by growing backlog, favorable pricing, and improved operational efficiency. Management also expects adjusted EBITDA to be more second half weighted with highest revenue and adjusted EBITDA in Q4 2026. 

Q4 net income was $3.7 billion or 33.5% of revenue compared to $484 million or 4.6% of revenue in the same period last year. The Q4 net income included a one-time tax benefit of $2.9 billion. 

Cash: 

Q4 operating cash flows grew by 169% YoY to $2.48 billion with an operating cash flow margin of 22.6% compared to 8.7% in the same period last year. The company benefitted from down payments on higher orders and slot reservations at Power as well as higher orders at Electrification. 

Q4 free cash flow grew by 214.7% YoY to $1.8 billion with a free cash flow margin of 16.5% compared to 5.4% in the same period last year.  

The company had cash of $8.85 billion and no debt at the end of Q4. 

In early February, the company expects to issue roughly $2.6 billion of debt in order to complete the previously announced acquisition of the remaining 50% ownership stake of Prolec GE. 

Valuation: 

GEV trades at a forward P/S ratio of 4.3. The company has traded at a minimum forward P/S ratio of 1.0 and a maximum of 5.3. Similar to Bloom Energy, the company is trading above the mid-range as it is a key beneficiary of rising energy demand from the global AI infrastructure build-out. 

Notable Risks: 

Valuation remains a key risk to monitor, alongside the ongoing weakness in the Wind segment. That said, management expects a meaningful recovery in the wind business to materialize in the second half of 2026. 

Please note, GEV reported earlier today with updated earnings report hitting inboxes Thursday.  

PJM Auction Stock: A Stock that Benefits from Grid Stress 

For our Discovery tier, we covered a stock that is grid dependent with up to 13GW of power, of this 1.9GW is contracted with a hyperscaler. The remaining capacity includes gas plants that are grid dependent, which means it does not solve time-to-power, but rather is a leveraged bet on auction pricing and wholesale pricing. These gas plants offer sizable capacity as they generate electricity for the grid rather than being load specific. 

Although this company does not solve the urgency around the AI data center expansion as transmission and grid allocation remain hurdles, it materially benefits as PJM pricing tightens. The investment thesis for the 13GW is that grid stress would cause the loads to run more often and clear at higher energy and higher capacity pricing. 

In our Discovery analysis, it’s pointed out that due to the rapid tightening in power supply, clearing prices for the PJM auction have surged to the tune of 11X over the past two years. Much of this arose in the 2025/26 auction, where clearing prices jumped 833% from $28.92/MW-day to $269.17/MW-day, reaching the annual cap. The 2026/27 auction saw prices once again hit the FERC-approved cap at $329.17/MW-day, a 22% YoY increase. 

As a merchant generator, this stock benefits from grid stress. Although hyperscalers must solve the issue of transmission, such as building data centers near the power assets, this can be hard to produce at scale. To help alleviate this, the operator is targeting regions popular for data centers for current capacity and newer acquisitions, such as Pennsylvania, Ohio and Maryland. 

To learn more about stocks in our new idea generation (NIG) pipeline, including a Top 10 list of NIG stocks, join Discovery todayjoin Discovery today. To subscribe to Discovery with 40% off, click here to email usclick here to email us or email premium@io-fund.com and mention code DISCOVERY40DISCOVERY40

AI Software: Tougher Trade than Previous Cycles

The I/O Fund has approached the AI software trade with caution as it’s been our contention that cloud software will go through a period of consolidation. In fact, we expressed this quite clearly in December of 2022, stating in the article “Slowing Growth in Cloud Stocks: When Will We Hit a Bottom?” 

“In some ways, the Q4 guides – assuming most come in at or near those guides – marks a historic slowdown for cloud as it’s always been a resilient category.” 

I emphasized this again in March of 2023: 

“There are a lot of cloud software bulls and for good reason, this category has treated investors well with predictable revenue growth. Cloud software is resilient because it drives down costs and increases productivity. We know this scenario well as we wrote about it many times in the past few years to defend cloud. Often, cloud selloffs were welcomed to position for a 6-month bounce back after the category sold off (40%) or more. I pointed this out in the past on the free side and here on MarketWatch (behind paywall) in 2019 (i.e., when we weren’t facing a brick wall on growth). 

The issue with this assumption is that Cloud growth is actually slowing down —- that is the reality of things —- and this wasn’t true in 2019 and hasn’t been true in the last decade. Couple this with weak bottom lines that require cash injections, and what get is a sector that is largely out of favor.” 

Around that time, I was on Real Vision and was asked for a long-only pick (I chose NVDA) and which stock(s) I would short (I chose GTLB and Bill.com). Here are the results after three years: 

I share this perspective because the opportunity cost in technology is immense—staying invested in the wrong areas can be just as costly as missing the right ones. While we spend significant time on AI semiconductors today, many of our Research Members originally found us through cloud software ahead of COVID. At that time, we made a deliberate—and unpopular—call that it was time to move on from cloud and reposition for what was next.  

How does this apply to AI software? 

First, we continue to see downward pressure on cloud stocks. Anthropic was the cause of a selloff recently after the company announced Claude Cowork, a new autonomous agent tool that builds spreadsheets, draft reports, browses the web and executes multi-step tasks. This marks an early attempt for an R&D firm to go after recurring, high-margin enterprise software budgets. Frankly, it makes a ton of sense that R&D firms will go for this low-hanging fruit, which is enterprise software that is not AI native. 

Overall, I predict that we will see immense disruption in the software layer to where the I/O Fund is considering very few software stocks for our portfolio at the moment. For every Palantir, there will be dozens that do not survive the incoming AI inference cycle. 

This is distinct from hardware, where in many instances, legacy players offer the most competitive solution given the hardware cycle requires many iterations, requires deeply entrenched supply chains and large, upfront capital investments compared to software. The barrier to entry on hardware is quite high, and that ultimately has played out well for public investors since the biggest winners across compute, networking, memory and power components are on the public markets as opposed to being smaller teams incubated in the private markets.  

Conservative tech investors looking for a small slice of AI exposure will gravitate toward software and may capture a winner or two, but that approach does not represent the full-fledged AI participation our portfolio seeks. Those who chase high-margin, recurring revenue only are not seeing the full picture, which is that software is the easiest path to compete and ultimately disrupt.  

Finally, because the Street tends to stick to what is familiar, software valuations will skyrocket leaving most investors exposed to buying high from the exuberance and selling low when early participants take gains. Avoiding this trap is critical.  

I took the long way to say that AI hardware remains the easier trade for public investors, while private investors will gun for the software market. CrunchBase says North American AI startups raised ~$168B in 2025, led by OpenAI’s $40B round and Anthropic’s $13B round, with funding soaring 46% in 2025. Eventually, venture capitalists will cash-in by putting leading AI software companies on the public markets, but it benefits them to wait a few years. Meanwhile, the I/O Fund is hard at work to make sure our Members can participate early in the cycle, and strategically too – I have no interest in waiting for AI software IPOs with bloated valutions when this report contains stocks supplying that private capital today. 

With that said, we have our eyes on the following stocks: 

Reddit: Contextual, High-Intent Data 

Reddit represents monetization momentum in the AI era as its data is highly valuable for training LLMs. There is something far more important that Reddit provides in the AI era than simply a forum; rather Reddit offers a continuous supply of human-generated conversations. What was once a forum is now a wealth of opinions and loads of sentiment that AI models desperately need to produce more natural and sentient-sounding responses. A few months back, Reddit announced they are suing companies like Perplexity and Anthropic for scraping their site.  

In exchange for data, Reddit ranks high on Google Search and in AI search results from Open AI, as well. This has helped Reddit move from #85 ranked site to #2 and #3 in 2025. In the last earnings call, management stated they are currently ranked #3: “Today, Reddit is the #3 most visited site in the U.S. for Semrush October 2025. That puts us in a rare company. YouTube is #2 and Amazon is #4.”  

The increased search ranking helped Reddit grow both their daily active users (DAUq) and weekly active users (WAUq) at a rate of 20% YoY.  

From the IOF’s internal checks, as of January 15th, Reddit has continued to have the 3rd place among the most visible site with YouTube taking the 2nd place spot. 

For user engagement, our internal checks show that Reddit notched 3.972 billion visits in October, up 4.5% MoM. For November, it was down (0.70%) MoM to 3.945 billion, better than Facebook’s decline of (4%) MoM to 11.27 billion. For December, Reddit’s monthly visits grew by 6.8% MoM to 4.2 billion, while Facebook’s MoM visits grew by 5.1% to 11.85 billion. 

With that said, Reddit’s report is not a slam dunk. First, the logged-out user growth is outpacing the logged-in user growth, which will take some getting used to for Street analysts as they often imply in the Q&A that logged-out users don’t monetize as well. Reddit may ultimately prove this wrong, but as an analyst team, we like to note where nearly-perfect fundamentals face headwinds. In this case, the concerns are not rooted in results, as the company has reported strong financials since the Google-sparked inflection. 

The company is rumored to be seeking a dynamic pricing model “where pay would be determined by how useful or important content is to the answers generated by AI tools.” This could provide more upside to Reddit’s data licensing side, which currently accounts for 6% of revenue in Q3, considering how frequently it is cited in AI Overviews and on ChatGPT. 

During the last earnings call, an analyst noted that roughly half of Reddit’s traffic is direct, while half comes from Google. Management confirmed the 50/50 split is “approximate, but pretty close.” This means Reddit is receiving an additional benefit from Google that isn’t fully visible within the data licensing revenue line item – rather, it’s mainly visible in the strong advertising growth from the traffic Google is sending to Reddit. Overall, the true impact of Reddit’s partnership with Google is hard to quantify. 

Overall Revenue Growth: 

Reddit once again reported stellar revenue growth of 67.9% YoY and 17.1% QoQ to $584.9 million. Revenue growth was more than 60% for the fifth consecutive quarter. The company’s Q3 revenue beat the analyst’s estimates by 6.4%. The strong growth was primarily driven by 74% YoY growth in the advertising revenue to $549 million. The total active advertising customers grew by over a solid 75% YoY as the company added new accounts across businesses, including large mid-market and SMB businesses. 

While its other revenues, which include licensing deals with Google and OpenAI, rose by a modest 7% YoY to $36 million. Regionally, revenue grew 67% and 74% YoY in the US and internationally, respectively 

AI Segment Growth: 

The company’s Q3 Average revenue per user (ARPU) grew by 41% YoY to $5.04. Management believes that this is still low on an absolute basis and remains an opportunity for the company. Though growth has decelerated from 47% reported in Q2 due to tough comps, it was up 11% on a sequential basis.  

The US ARPU grew by 54% YoY to $9.04, a 5-point deceleration from a strong 59% YoY growth in Q2. However, it grew by 15% sequentially.  

The International ARPU grew by 39% YoY to $1.84, a slight deceleration from the 40% growth reported in Q2 and was up 6% sequentially. 

The company’s Daily Active Uniques (DAUq) are witnessing strong international growth. The Daily Active Uniques (DAUq) global grew by 19% YoY to 116 million. While US growth is stabilizing as it grew by 7% YoY to 51.6 million, it showed a sequential growth of 3%, while it was flat in Q2. The international DAUq growth was solid as it was up 31% YoY to 64.4 million.  

The company’s Weekly Active Uniques (WAUq) grew by 21% YoY to 443.8 million. International growth outpaced US growth as it grew by 37% YoY to 256 million, while the US grew by 6% YoY to 187.8 million.  

Earnings: 

Analysts expect strong EPS CAGR of 49% during the period 2025 to 2027. EPS is expected to grow from $2.32 in 2025 to $12.74 for the year 2030, growing at a CAGR of 41%.  

The company’s Q3 GAAP EPS grew by 400% YoY and 78% sequentially to $0.80, beating analyst estimates by a solid 53.8%. Analysts expect EPS to grow 119.6% YoY to $0.79 in Q4 and 226.7% YoY growth to $0.42 in Q1 2026. Looking forward, they expect EPS to grow 76.3% YoY to $3.35 in 2026 and 39.9% YoY to $4.69 in 2027.  

Q3 adjusted EBITDA grew by 151% YoY to $236 million. Adjusted EBITDA margin improved by 13.3 percentage points YoY and 6.9 percentage points sequentially to 40.3%, beating the management guidance by 5.1 percentage points. 

Margins: 

Q3 gross profits grew by 69.7% YoY to $532.4 million with a gross margin of 91%. The gross margin is up 90 basis points YoY and up 20 basis points sequentially. The company reported its fifth consecutive quarter of above 90% gross margins.  

Operating income was $138.5 million compared to a mere $6.9 million in the same period last year. Operating margin improved by 21.7 percentage points YoY and 10.1 percentage points sequentially to 23.7%, primarily driven by operating leverage.  

Cash: 

The company reported strong cash flows primarily driven by record profits.  

Q3 operating cash flows grew by 158.6% YoY to $185.16 million with an operating cash flow margin of 31.7%, up 11.1 percentage points YoY.  

Q3 free cash flows grew by 160.5% YoY to $183.1 million, with a free cash flow margin of 31.3%, up 11.1 percentage points YoY. The company generated $510 million in free cash flows in the last twelve months. 

Valuation: 

Reddit is trading at a forward P/S ratio of 13. The company has traded at a low of 4.2 and a high of 24.4 since the company’s listing in March 2024. Reddit is currently trading at the mid-range. On the bottom-line, the company is trading at a forward P/E ratio of 35.3 with a low of 18.3 and a high of 95.8. It is trading lower than the mid-range of 57. It is also important to note that the company only achieved GAAP profitability in Q4 2024, which limits the usefulness of earlier P/E comparisons. Looking ahead, earnings growth remains strong, with EPS expected to increase from $2.33 in 2025 to $12.73 by 2030, representing a 40.4% CAGR and suggesting meaningful upside as profitability scales. 

Notable Risks: 

Reddit’s primary risk is the surge in traffic relies on a third-party relationship with Google that could be terminated at any time. It may not be terminated given the emphasis on contextual data for models, yet the recent success hinges on this data licensing deal. 

AppLovin: Sentiment Doesn’t Match Fundamentals 

AppLovin is a stock that needs a strong technical analysis overlay. Despite fundamentals that rank the company as one of the strongest FA stocks in the tech sector, the market struggles with AppLovin following short seller reports and other sentiment-driven concerns.  

From a 10,000-foot view, AppLovin is in the crosshairs of Big Tech, as it’s one of the only grassroots companies to emerge as a formidable data-driven advertising player since the walled gardens of Facebook and Google solidified in the early 2010s. It’s unfortunate that healthy competition to Big Tech often has a target on its back, as I’ve seen many times throughout the years (Zoom’s so-called security and encryption issues come to mind when they offered similar settings as Microsoft Teams).  

Point being, it’s hard to find fault in AppLovin’s exceptional fundamentals, yet technicals suggest there will be continued volatility that must be closely navigated.  

Regarding potential catalysts, although very early and based on small numbers, management stated AXON’s self-serve feature is seeing strong traction with advertiser spend growing 50% week-over-week since the launch October 1st. This is invite-only, referral-based demand in the e-commerce vertical with the platform expected to open up more broadly in early 2026.  

“While it takes a while for new customers to get going, to integrate, to learn how to use our system and to ramp spend, we're already seeing spend from these self-service advertisers grow around roughly 50% week-over-week. It's too soon to be significant, but this type of early growth gives us even more confidence that our platform will excel at being an open platform to any type of advertiser.” 

According to management, their AI models continually learn for better behavior targeting and ad personalization. Generative-AI based creatives are also a feature being built out to generate more effective ads (also leading to higher conversion rates). An area where Applovin sets themselves apart is the 35 second ad creatives compared to 7 seconds on social, which could (presumably) also lead to higher conversions.  

According to management, improving conversion rates is a path to sustained growth: “We believe that giving our powerful recommendation engine, a more diverse set of advertisers to recommend will dramatically improve conversion rates, paving the way for elevated growth rates for years to come.”  

Overall, it’s important to remember that Applovin is demand constrained rather than supply constrained as they reach over 1 billion users. Therefore, opening up the AXON ad manager to more demand is the primary catalyst for the next few quarters. 

Overall Revenue Growth: 

AppLovin reported strong revenue of $1.405 billion, beating analysts' estimates by a solid 4.7%. The company’s revenue grew by 68.2% YoY and 11.6% QoQ. 

However, investors should be aware that App’s long-term target is much lower at 20% to 30% – yet management has openly discussed their path to > 30% growth. At Goldman Sachs’ Communacopia conference, executives dove deeper into the long-term growth framework provided in Q2, calling for a baseline 20% to 30% annual growth. Management explained that this hinges on two primary factors: reinforcement learning and continuous improvement on the ad engine, and opening the recommendation engine up to e-commerce and exposing it to a wealth of new demand.  

The update regarding 20% to 30% growth is the self-service platform could help exceed this baseline: “We're still believing very confidently in this 20% to 30% long-term growth rate in our core category. But even in the core, we're beating that. And then now you're layering on, on top of that, all this opportunity with the self-service platform” 

AI Segment Growth: 

The company’s Q3 advertising revenue grew by 68.3% YoY to $1.405 billion. The ad revenue exceeded the management guidance by a solid 5.6%, primarily driven by strong gaming advertising revenue.  

Management guided advertising revenue of $1.57 billion to $1.60 billion, representing a YoY growth of 58.6% at the midpoint. Management stated that the guidance incorporates optimism around the e-commerce referral program, continued model enhancements, and the normal holiday seasonality. 

Earnings: 

Gross profits grew by a solid 72.2% YoY to $1.23 billion, with a gross profit margin of 87.6%. The gross profit margin was up 210 basis points YoY and down 10 basis points sequentially. 

Operating profits grew by 102% YoY to $1.08 billion, driven by solid operating leverage. The operating margin improved by 12.8 percentage points YoY to 76.8%. 

Margins: 

Gross profits grew by a solid 72.2% YoY to $1.23 billion, with a gross profit margin of 87.6%. The gross profit margin was up 210 basis points YoY and down 10 basis points sequentially. 

Operating profits grew by 102% YoY to $1.08 billion, driven by solid operating leverage. The operating margin improved by 12.8 percentage points YoY to 76.8%. 

Cash: 

Q3 operating cash flows grew by 91.3% YoY to $1.05 billion with a margin of 75%, up 9.1 percentage points YoY.  

Q3 free cash flows grew by 92.4% YoY to $1.049 billion with a free cash flow margin of 74.7%, up 9.4 percentage points YoY.  

The company’s cash improved to $1.67 billion, up from $1.19 billion at the end of the previous quarter. While debt remained the same at $3.51 billion. 

Valuation: 

APP is trading at a forward P/S ratio of 22.7. The company has traded at a minimum of 1.1 and a maximum of 43.2. On the bottom line, the company is trading at a forward P/E ratio of 37.6. APP has traded at a minimum of 3.1 and a maximum of 73.1 in recent years. Currently, it is trading at mid-range. 

Notable Risks: 

APP is the subject of short reports, and the company has been under an SEC probe over its data collection practices. In addition, the stock’s strong outperformance over the past three years raises the bar for future execution, as market expectations are elevated. However, we think the AI-powered ads business model, which has driven strong revenue and profit growth and a strong market presence, is worth a shot, especially when using technicals to guide our entries and exits. 

Cloudflare: Early but the Positioning is One of a Kind 

As pointed out in our analysis: “Cloudflare Entering Act 3 to Become a Leader in AI Inference at the Edge,” the company has a few distinct advantages as the platform of choice for AI developers. Here’s a summary:  

• Does not rely on Big 3 infrastructure and can drive down costs  
• Is faster on performance because of its position at the edge; this lowers costs and latency for AI inference and keeps data as close to the user as possible  
• Geographically equipped to handle compliance issues that will inevitably result from using training data for inference.   
• The company has moved diligently into compute, storage and application services. Combined with its global network, this positions the company for AI inference as-a-service. There is no other company doing both edge network plus compute and storage except the hyperscalers. However, in some cases such as serverless, Cloudflare exceeds the performance of the hyperscalers.  
• CDN as a core product and security as a seamless upgrade shows the importance of being a middleman, helping to position Cloudflare to innovate around Serverless in ways that outperform even AWS.     
• Training models is prohibitively expensive by requiring upfront costs, Nvidia GPUs are hard to obtain, and AI development is not democratized for developers with proprietary, blackbox APIs that run counter to an open-source movement (GPT-4 versus Llama). Cloudflare aims to solve these problems by allowing popular models to run closer to the user, which is the next logical step for AI. 

Ultimately, the bigger and the faster a network is, the more it’s capable of providing “as a service.” AI can create a fortuitous moment for Cloudflare because the company is both positioned to offer AI inference-as-a-service yet also solves important pain points for developers.  

Overall Revenue Growth: 

Cloudflare reported its largest beat since Q1 2022, reporting revenue of $562.0 million in Q3, 3.1% ahead of estimates as growth accelerated nearly three points to 30.7%. This also marked Cloudflare’s first >30% growth quarter in the past five and its fastest revenue growth in the last seven quarters. This is the first step in confirming a sustained revenue acceleration aided by AI, yet the more important piece is showing that >30% growth can actually be sustained. 

For Q4, Cloudflare guided for revenue of $588.5 to $589.5 million, a slight deceleration to 28% on the topline. This was ahead of estimates for $580.8 million 

AI Segment Growth: 

Cloudflare has not broken out specific AI revenue or contribution to growth, although other key metrics strengthen significantly in Q3. 

RPO was $2.14 billion, accelerating four points to 43% YoY, while current RPO accounted for 64% of RPO, or ~$1.37 billion. Current RPO rose ~30% YoY, a three point deceleration from 33% in Q2. Billings growth accelerated sharply, from 33% in Q2 to 40% in Q3, rising to $624.4 million. 

Paying customer growth accelerated six points sequentially to 33% YoY, impressive at this scale considering paying customers now total 295,552. Growth was 10% QoQ, the highest on record since at least 2022. Additionally, DBNRR ticked five points higher sequentially to 119%, the highest since Q4 2022, driven by accelerating spending at its largest customers 

Earnings: 

Cloudflare reported a solid adjusted EPS beat in Q3, reporting 35% YoY growth to $0.27 versus the $0.23 estimate. GAAP EPS was on the brink of shifting to positive territory at ($0.00), versus the ($0.07) estimate. 

For Q4, Cloudflare guided for adjusted EPS to be flat QoQ at $0.27, up 42% YoY. For fiscal 2025, Cloudflare raised its adjusted EPS forecast to $0.91, up from $0.85 to $0.86 previously. However, GAAP profitability is not expected on an annual basis until 2027. 

Margins: 

GAAP gross margin was 74.0% in Q3, down 3.7 points YoY and 0.9 points QoQ. Adjusted gross margin was 75.3%, down 3.5 points YoY and 1 point QoQ, again impacted by increases in allocated costs from higher network traffic from paying customers. 

GAAP operating margin was (6.7%), up 0.5 points YoY and 6.4 points QoQ. Adjusted operating margin was 15.3%, up 0.5 points YoY and 1.2 points QoQ; for Q4, adjusted operating margin was guided to be 14%. GAAP net margin was (0.2%), up 3.4 points YoY and 9.6 points QoQ. Adjusted net margin was 18.3%, up 1.4 points YoY and 3.6 points QoQ. 

Cash: 

Operating cash flow was $167.1 million for a 30% margin, up from a 24% margin in the year ago quarter and a 19% margin in Q2. Free cash flow was $75 million for a 13% margin, up from 11% in the year ago quarter and 6% in Q2. Network capex was 14% of revenue. 

Cash, equivalents and available-for-sale securities totaled $4.04 billion, while convertible notes outstanding totaled $3.26 billion. 

Valuation: 

Cloudflare is trading at a forward P/S ratio of 22.2. The company has traded a minimum of 10 and a maximum of 41.4 in the last few years. Cloudflare is trading slightly lower than the mid-range after the recent weakness in its share price. 

Notable Risks: 

The company is not yet GAAP profitable even after 16 years of the company’s operations.  

Palantir: The Trade-Off Between Discipline and Conviction 

Since 2023, Palantir’s stock has defied gravity, delivering steady performance that no other AI software stock has come close to matching (yet). The thesis is two-fold: the company must continue to scale its Commercial segment after posting multiple quarters of over 50% growth, while also sustaining a high valuation. Both matters and the bar is undeniably high.  

What separates Palantir, however, is not simply growth, but capability. The differences matter as unlike traditional AI-enabled database or business intelligence competitors, Palantir can operate effectively even when data sets are incomplete or fragmented—situations where most models struggle. In that regard, traditional business intelligence companies require a complete data set, whereas Palantir can handle situations where one isn't available. You can think of the competitive advantage as actionable depth, as Palantir has described it: “the reasoning that goes into decision-making, not just data.”     

Palantir’s Artificial Intelligence Platform (AIP) integrates generative AI with operational data and workflows, and, when combined with Palantir’s other platforms, Foundry and Apollo, it provides an AI service mesh that can run hundreds of microservices, scale compute via its Rubix engine, and orchestrate updates through Apollo.    

Additionally, Palantir’s knowledge graph, referred to as Ontology, is a distinct advantage. The graph offers better context than a large language model would on its own – or as Palantir states, it’s “the reasoning that goes into decision-making.” Palantir made key upgrades to AIP with the introduction of AI-forward-deployed engineers (FDEs) and the AI Hivemind, and brought Ontology to the edge, enabling deployment on mobile devices.  

Palantir Stock leads the AI software pack, delivering one of the best reports across tech in Q3. Revenue accelerated nearly 15 points sequentially to almost 63%, with strong growth in key metrics and a 28-point acceleration in US Commercial revenue to 121% YoY. The Artificial Intelligence Platform (AIP) is driving most of the Commercial growth, as there was a clear revenue inflection when AIP launched in mid-2023.  

The company reported strong key metrics, with net retention rate (NRR) expanding six points sequentially to 134%. Over the past two years, NRR has risen an impressive 27 points, and Palantir noted that AIP is continuing to drive existing expansions and new customer conversions. On the other hand, Palantir’s forward P/S ratio trades at an outstanding 64.4 multiple and has been as high as 112 forward P/S. 

I don’t recall another stock the I/O Fund has followed this closely without taking action. That caution was intentional, driven by valuation and our focus on risk management. Ultimately, Palantir is an extreme outlier, to where for those ignoring discipline, it worked out. Often times, it does not work out to buy a stock that trades at up to a 100 forward PS, and that must be weighed carefully for each investor. 

Overall Revenue Growth: 

Palantir reported $1.18 billion in revenue in Q3, up an impressive 18% QoQ and beating estimates by 8.4%, driven by unwavering momentum in US Commercial. On a YoY basis, revenue growth accelerated 14.8 points to 62.8% YoY, the largest sequential acceleration to date and marking Palantir’s highest growth rate since going public. Over the last nine quarters, topline growth has accelerated ~50 points, from just 12.7% in Q2 2023, a rare feat to accomplish. 

AI Segment Growth: 

Fueled once again by AIP, Palantir delivered one of the best reports across tech in the third quarter, with revenue accelerating nearly 15 points sequentially to almost 63%, with strong growth in key metrics and a 50 point acceleration in US Commercial revenue since the start of the year.  

US Commercial revenue grew 29% QoQ and 121% YoY to $397 million in Q3, accelerating from 93% YoY growth in Q2. Since the start of the year, US Commercial growth has accelerated 50 points, and since the start of 2024, growth has accelerated 81 points. 

Earnings: 

Palantir reported $0.18 in GAAP EPS in the quarter, up 200% YoY, while adjusted EPS was $0.21, beating estimates by 25.5% and rising 110% YoY. Palantir did not provide a specific guide for EPS for Q4, though current estimates are pegged at $0.12 in GAAP EPS and $0.22 in adjusted EPS, up 300% YoY and 57% YoY, respectively.  

For FY25, Palantir is expected to earn $0.72 in adjusted EPS, up nearly 76% YoY, before slowing to 39% growth to $1.01 in FY26. 

Margins: 

Margins strengthened considerably in the quarter, with adjusted operating margin surpassing 50% with more expansion guided for Q4. Palantir’s Rule of 40 score (revenue growth + adj operating margin) expanded to a wild 114%, up from 94% last quarter and 68% last Q3.  

Gross margin was 82% in Q3, up one point QoQ and two points YoY, while adjusted gross margin was 84%, up two points YoY and QoQ. 

GAAP operating margin was 33%, an impressive 6 point QoQ and 17 point YoY expansion. Adjusted operating margin was 51%, breaking past 50% for the first time and up 5 points QoQ and 13 points YoY. For Q4, Palantir guided for adjusted operating margin to be 52%, showcasing its ability to drive strong margin expansion alongside swift revenue acceleration. Full year adjusted operating margin guidance was raised from 46% to 49%. 

Cash: 

Cash flows were strong, though cash flow margins dipped on a YoY and QoQ basis. Operating cash flow was $507.7 million for a 43% margin, shrinking from a 54% margin in Q2 and 58% in the year ago quarter.  

Adjusted free cash flow was $539.9 million for a 46% margin, down from 57% in Q2 and 60% in the year ago quarter. Palantir raised its adjusted FCF guidance for the year to $1.9 to $2.1 billion, or a 45.5% margin, up from a 42.8% margin previously.  

Cash and equivalents totaled $6.4 billion and debt remained zero. 

Valuation: 

Palantir is trading at a forward P/S ratio of 64.4. The company has traded at a minimum of 6 and a maximum of 112 in the last few years. The company is trading at a significant premium to the other best of breed cloud companies like CrowdStrike that is currently trading at a forward P/S ratio of 23.7 and Cloudflare at 22.2. 

On the bottom line, the company is trading at a forward P/E ratio of 167.6. The company has traded at a minimum of 25.6 and a maximum of 285.9 in the past few years. 

Notable Risks: 

The company’s primary risk is its high valuation. 

CoreWeave: Legacy Cloud IaaS Wasn’t Built for AI 

CoreWeave breaks all of the rules, including not cooperating with our portfolio criteria. I’ll get right to the point by saying CoreWeave’s cash to debt is frightening. The company reported FCF of ($1.6 billion) with $14 billion in debt and a mere $2.5B in cash on the balance sheet, leaving a cash to debt ratio of 0.18, or said differently; debt is 5.6X cash at the end of Q3. Notably, this excludes the $2.25 billion convertible senior notes issued in December and on a pro-forma basis, debt is 3.4X cash, for a deep net-debt position and very limited balance-sheet flexibility. 

Perhaps most concerning, the debt issues are about to worsen as CoreWeave is expected to spend $6.75 billion in Q4 on capex and over $26 billion in 2026, as management expects capex more than double next year. My best estimate is that 2026 will see 12X debt to cash with what I know today. The only way we would touch this stock is with heavy technical analysis and risk management.  

There are lower risk ways to participate in AI, yet the positioning CoreWeave offers is second to none. The company is in the “build” phase but will eventually be in the “yield” phase. 

The company also entered the US federal market, which should further help to diversify its customer base. CoreWeave will provide secure, compliant, high-performance AI cloud services to US government agencies and their key partners, including the Defense Industrial Base. NASA already uses its services to advance scientific exploration at its Jet Propulsion Lab. 

Altogether, CoreWeave sits on the front lines of the shift from legacy cloud infrastructure to AI-optimized workloads. While the full importance of this transition from cloud to AI is difficult to quantify today, its impact is likely to be transformative for how compute is built and consumed. CoreWeave is positioned at the center of a shift too great to fully envision today. 

The closest historical parallel is AWS in the mid-to-late 2000s—before the economics of the build-out were fully visible to investors. The key distinction is that CoreWeave represents a pure-play on AI infrastructure. It is now widely understood that AWS went on to generate the majority of Amazon’s profits, providing investors with a clear blueprint of what the yield phase of infrastructure-as-a-service can look like. 

For more information on how CoreWeave is unique compared to the Big 3, including why the model FLOPs utilization (MFU) gap matters quite a bit, reference our article “CoreWeave Stock Soars 200% since IPO – Can it Defy the Odds?” 

Overall Revenue Growth: 

CoreWeave’s Q3 revenue grew by 133.7% YoY and 12.5% QoQ to $1.37 billion. The company beat analyst consensus estimates by a solid 6.6%, driven by continued strong demand for the company’s AI cloud infrastructure services.  

Looking ahead, analysts expect 2026 revenue to grow 132% YoY to $12.23 billion, and these estimates will be increased due to the push-out caused by the delay in Q4 revenue recognition to Q1. For 2027, revenue is expected to grow 49.4% YoY to $18.27 billion. 

AI Segment Growth: 

The backlog of $55B represents nearly double Q2 and is approaching 4X YTD yet the debt is also up 2X YTD. The company stated the backlog grew by $25 billion to $55.6 billion, up from $30.1 billion for growth of 85% QoQ.  

Management also highlighted that they reached $50 billion in RPO, faster than any cloud in history.  

Active power footprint grew by 120MW sequentially to approximately 590MW with contracted power capacity growing over 600MW to 2.9GW. That represents 25.5% QoQ growth. Management expects to end the year with over 850 megawatts of active power. 

Earnings: 

Q3 GAAP EPS was ($0.22) compared to the analysts' estimates of ($0.51). However, the strong beat was due to a one-time noncash tax benefit of $0.25. Excluding the one-time benefit, the company would beat estimates by $0.04.  

Looking forward, analysts expect GAAP EPS of ($0.84) in 2026 and to be GAAP profitable in 2027 with an EPS of $1.63. 

Margins: 

The margins are strong yet the cash remains troublesome. For example, CoreWeave is a recent IPO that is already GAAP positive on operating margin at 4% and reported an adjusted EBITDA margin of 61%. 

Q3 gross profits grew by 126% YoY to $995.85 million with a gross profit margin of 73%, down 200 basis points YoY and 100 basis points sequentially.  

Q3 operating margin was 4%, down from 20% in the same period last year and up 200 basis points sequentially. The operating expenses increased 181% YoY to support strong growth. The adjusted operating margin was 16%, compared to 21% in the same period last year. 

Cash: 

The company reported negative free cash flow of ($1.6 billion) with $14 billion in debt and $2.5B in cash on the balance sheet at the end of Q3. This leaves net debt of $11.5 billion – yet this is mild given what the company plans to spend in capex next year (expect the debt to go up rapidly).  

Free cash flow was ($1.6 billion) compared to ($573.9 million) in the same period last year and ($2.7 billion) in the previous quarter. 

Valuation: 

CoreWeave is trading at a forward P/S ratio of 3.8. The company has traded at a low of 3 and a high of 17.2 in the past year.  

The company is not profitable for a bottom-line valuation and is expected to be profitable on a non-GAAP basis in Q4 2026. 

Notable Risks: 

The company has negative free cash flow due to high capex for infrastructure, and it also has high debt. 

Honorable Mention: Meta 

In a recent analysis entitled “The AI Revenue Leader Nobody is Talking About,” our firm was early to point out that Meta’s AI revenue places it as number two, second only to Nvidia. Although Google has many supportive points as to why the stock outperformed compared to other Big Tech names, the I/O Fund is a growth stock portfolio. Margins matter, cash matters, but what matters more is the 3X growth Meta has seen in its Advantage+ segment in less than a year, as the company had reported $20 billion about three quarters ago, with the recent update from last quarter at $60 billion. If this runaway growth continues, then Meta will easily be outpacing Search and Google Cloud combined on AI revenue.

On the other hand, Meta is witnessing a deceleration in margins due to rising expenses supporting its AI infrastructure. Reality Labs also continues to incur losses, recording a $4.43 billion loss from operations in Q3 2025, and its cumulative losses now total $73.04 billion. Due to continued investments in AI infrastructure, the company’s capex is expected to be significantly higher in 2026.  

Meta has the weakest balance sheet among the Big Tech companies, with a net cash position of $15.7 billion. Meta has also entered a joint venture with Blue Owl Capital to fund its development at the Hyperion data center in Louisiana. Thereby, helping it to keep about $27 billion in debt off-balance sheet, where it would sit in a special-purpose vehicle tied to Blue Owl. While this approach may improve reported leverage and financial ratios, it carries inherent risks as the company is indirectly responsible for the off-balance sheet debt.  

Despite Meta being in the quality bucket for the most part, its high capex spending necessitates technical analysis and a risk management overlay. 

For more information on Big Tech with I/O Fund takeaways, please read our free article “The $530 Billion AI Question: Which Big Tech Stock is Winning?”The $530 Billion AI Question: Which Big Tech Stock is Winning?”

What’s Next for Our Discovery and Advanced Tiers … 

Miners have been attracting significant deal activity from neoclouds, with a handful notching hyperscaler deals and growing interest. Miners continue to benefit from their ability to offer hundreds of MW for AI data centers in relatively quick fashion, bypassing interconnection queues for greenfield builds, while also offering lower electricity costs through long-term power contracts. A handful of miners have disclosed power costs around $0.046–0.047/kWh, representing a meaningful discount to PJM’s grid and average commercial electricity pricing. 

In an upcoming analysis for our Discovery tier, we will recap three Bitcoin miners leading the push to power AI data centers. One has secured a second multi-billion-dollar AI data center deal with a hyperscaler and is pursuing a multi-GW development pipeline that could represent 7–8X growth from current contracted capacity. Another has signed a nearly $10 billion deal for phased deployment through 2026, while the third is targeting several hundred MWs online in 2026 with additional capacity in pre-development for 2027. 

While AI-related revenue contributions remain modest today, growth is expected to accelerate through 2026 and 2027 as capacity comes online. However, the risk with miners is that capex requirements to retrofit facilities often exceed current balance sheet capacity, forcing increased leverage to transition assets from mining to AI-ready infrastructure. 

Outside of miners, we are also revisiting nuclear power for Discovery members, including an SMR developer with a multi-GW pipeline. Unlike miners or Bloom Energy, SMRs represent a long-term solution, with commercial operations not expected until closer to the end of the decade. 

Conclusion: 

The I/O Fund team is ready for the upcoming earnings season armed with a list of stocks we will be watching very closely and many honorable mentions prepared to step-in should one of our chosen stocks not perform as expected.

Following a report of this size, it’s worth pausing to acknowledge a reality that often gets overlooked: AI investing remains difficult for many tech portfolios, despite the growing list of winners and the market’s potential to meaningfully reshape GDP.

This raises a fair question—why do so many hedge funds and ETFs remain underexposed to AI beyond a narrow set of Big Tech names, and why is that exposure so concentrated? The AI trade is actually quite complex and unforgiving, demanding deep product-level analysis, precise timing, and disciplined risk management that many portfolios are not built to execute.

Our goal is to solve that problem for our Members—building on our history from prior cycles, striving to be early to market trends in the near-term, and positioning thoughtfully for the second half of this AI-driven decade.

Damien Robbins and Royston Roche, Equity Analysts at I/O Fund contributed to this analysis.

I/O Fund Members Get 40% off Discovery

Discovery members recently received an analysis on a stock poised to benefit as Nvidia tackles the context memory bottleneck and extends KV cache memory with its new Inference Context Memory Storage platform. To subscribe to Discovery with 40% off, click here to email usclick here to email us or email premium@io-fund.com and mention code DISCOVERY40DISCOVERY40

I/O Fund Pro Members Get 30% Off Advanced Market Signals

Unlock real-time trade alerts, access to the I/O Fund’s momentum stock list, and weekly webinars every Thursday at 4:30 p.m. ET. To join Advanced Market Signals with 30% off, click here to email usclick here to email us or email premium@io-fund.com and mention code ADVANCED30.ADVANCED30.

Please note: The I/O Fund conducts research and draws conclusions for the Fund’s positions. We then share that information with our readers. This is not a guarantee of a stock’s performance. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis.

Recommended Reading:

• The AI Memory Boom Has Arrived
• The I/O Fund’s Top 15 AI Stocks for Q4 2025
• The I/O Fund’s Top 15 Stocks for Q3 2025
• Meta: Growth is Quietly Benefitting from AI, Though Margin Risks Weigh Heavy