Last quarter, I began providing a comprehensive ranking of the top AI stocks in our universe. Our portfolio benefited from this exercise, and we hope yours did too.  

Perhaps most beneficial is taking a moment on a quarterly basis to be as objective as possible. Information in the form of earnings reports, product announcements, news headlines, and even social media exuberance can cloud an investor’s decisions. It is not only the volume of information, but the speed at which information comes daily to where it can be hard to discern the true winners from those catching a fleeting headline. 

The report below is designed to be objective in a no holds barred approach. Because we manage our own money full-time, we are laser-focused on ways to improve on a quarterly basis. Similar to the Q3 2025 report, the analysis below tops out at over 16,000 words, totals 43 pages and took three weeks to write. Even if we’ve held a stock for years, we update the investment thesis and re-examine the fundamentals. There is no stock that is immune to being cut from our portfolio if it cannot prove it’s one of the best AI stocks at this moment. Vice versa, to be added for the first time, a stock must be able to prove it can hold its own among already-strong choices. 

After rigorous examination, the list below summarizes the strongest AI stocks of Q4 2025 and the top 3 Thematic Trends that I believe will help drive the AI market to new heights. 

AI Networking: The Exponential Investment Opportunity 

Each quarter, the I/O Fund scans hundreds of tech companies and their quarterly earnings reports for anomalies. To be precise, we track 188 data points on each portfolio company and top-ranking stocks we don't own totaling about 40 stocks a quarter for a total of over 7,000 data points. This quarter, one metric stands out as truly exceptional: Nvidia’s networking growth of 46% QoQ and 98% YoY to $7.25 billion. In the previous quarter, Nvidia’s networking grew 64% QoQ after declining (3%) in the previous quarter. Management stated the networking to compute attach rate is 75% in the fiscal Q4 earnings call. 

Here is why this data point is unique – if Nvidia’s networking segment were a standalone segment, it would place #3 in the world on AI revenue with Azure reporting about $36 billion per year in AI revenue (at $75 billion total with a statement that half is AI-related) versus Nvidia’s networking now on a $28 billion run rate. Broadcom would be fourth at $6.2 billion AI revenue guided for next quarter, or about a $25B run rate.  

There’s an argument to be made that Google Cloud or AWS would be in the running, however, they do not breakout AI revenue yet. An investor can also reasonably assume if it was at Azure’s level or higher, they would share their first place standing among the Big 3.  

Yet, if any of those leading AI companies grew their AI segment 46% QoQ, their stock would go wild. The metric was missed because it was buried in Nvidia’s Compute weakness/China noise, yet this number is truly the bellwether as we move into Q3.  

It also gives me a nostalgic pause as I was early to cover the importance of the Mellanox acquisition in both 2019 and 2020 stating: “Mellanox’s Ethernet switch systems are the most used internal system in the top 500 in a recent report released at ISC High Performance, with 247 systems, and InfiniBand is the second most-used, with 140 systems. However, InfiniBand, a computer-networking communications standard, connects the most high-powered computers where the presence of Ethernet is nearly non-existent […] This is a strategic acquisition for Nvidia and Mellanox to become the strongest combination for artificial-intelligence and machine-learning computations.” You will be hard-pressed to find an equity analyst covering the AI market that closely back in 2019-2020 at the height of Covid.  

Turning our attention to 2025 and beyond, we want to look at what stocks are downstream from the number one growth market at scale.at scale. As stated, it’s not only the largest QoQ growth across AI, but rather it’s that Nvidia was able to grow its networking segment 46% QoQ while operating as the world’s third-largest AI segment.  

My readthrough is that something important is ramping, and it will be back-half weighted. You’ve known for many quarters that the “something” is the new Blackwell GB200 system. What is newer information is that Blackwell Ultra GB300 is also ramping nearly simultaneously.  

Per Nvidia, the strong performance in Networking was driven by “growth of NVLink compute fabric for GB200 and GB300 systems, the ramp of XDR InfiniBand products, and adoption of Ethernet for AI solutions.” 

Double-Clicking on the Network Requirements of the GB200s and GB300s 

Nvidia’s networking segment is a proxy for the AI networking allocation we hold in our portfolio. The quarterly numbers that Nvidia provide relate to their proprietary InfiniBand networking and the NVSwitches that help to route NVlink connections for GPU-to-GPU communication.

However, as the 75% attach rate above describes, there is a roughly 25% opportunity for other networking vendors to participate. Additionally, there are times that Nvidia will outsource components or cabling rather than build every single component in-house. In some instances, sourcing raw earth materials such as indium phosphide (InP) would prove quite challenging. Therefore, the information below may relate to Nvidia’ proprietary networking, yet it has important readthroughs to carefully selected suppliers. 

Generally speaking, networking cables, ASICs and components required for GPU interconnects will increase 5X to 9X as we move from the HGX/DGX systems to the NVL72 systems. Given the GPU count will increase 9X from 8 GPUs to 72 GPUs, it makes sense that networking will increase similarly. 

This is inferred by a few key points in terms of how the architecture is shifting from the 8-GPU HGX/DGX systems in the Hopper generation. These older systems that shipped in 2023-2024 used four NVSwitch ASICs to connect eight TensorCore H100 GPUs.  

Source: Nvidia Technical BlogNvidia Technical Blog 

Networking is at the heart of the Blackwell architecture as the increased bandwidth is instrumental in driving the higher performance. The NVLink domain moves from supporting eight GPUs to 72 GPUs with a speed of 1.8 TB/s. Nvidia’s 5th generation NVLink interconnects will deliver a higher aggregate bandwidth of 9X to 18X compared to Ethernet and InfiniBand in previous generations. By increasing the bandwidth, the NVL72 systems will pool together compute and memory for up to 4X faster training and 30X faster inference. 

This is accomplished with 18 NVSwitch ASICs, up from four in the HGX/DGX systems. That’s a 4.5X increase in NVSwitch ASICs.  

Source: Nvidia Technical Blog, “Nvidia Contributes Nvidia GB200 NVL72 Designs to Open Compute Project”Nvidia Contributes Nvidia GB200 NVL72 Designs to Open Compute Project”

If you look below at the specs between the larger NVL systems and the two HGX SKUs, you can see it’s not only compute performance that increases 10X (petaFLOPs) but also aggregate memory bandwidth and speeds.  

The main takeaway is that scale-up architectures are as much about networking as they are about compute. The goal is to add more GPUs that can effectively distribute training and inference across a large cluster, which includes sharing data and memory, synchronizing, and exchanging model parameters.  

Therefore, the interconnect or “connectivity layer” must scale accordingly to prevent GPUs from idling and waiting for data across the increased number of GPUs that are operating in parallel and at scale. High-speed communication is central to Blackwell and future generations of GPUs because the goal of a GPU-cluster is to act as a single processor. As AI systems grow GPUs by 9X, but also clusters grow from 100s of thousands to millions, there will be significantly more components required such as retimers, switch fabrics, silicon photonics, transceivers, cables and controllers.

However, despite Nvidia and Broadcom dominating the AI networking space, there remains immense opportunity for smaller, lesser-known suppliers.  

Key opportunities mentioned above include PCIe components, as this is needed to connect the GPU to CPUs, memory and storage. Specialized copper cables are needed to connect AI servers to networking switches, and for linking many servers together. Retimers help to extend data integrity beyond short distances. The jump from 400G to 800G to eventually 1.6T on data rates will require an additional upgrade to nearly every component in the network layer, such as switch ASICs, optical transceiver modules, faster digital signal processors (DSPs) and retimers/redrivers. When it comes to products like EML lasers, the indium phosphide (InP) is hard to source. There is expected to be a new market for co-package optics (CPOs) with the new Rubin architecture – to name just a few of the AI networking opportunities currently in play and also squarely in front of us. 

AI Energy: How Much and How Fast? 

AI energy is all the rage today, yet it was rarely spoken about when we first covered the trend in the article AI Power Consumption Becoming Mission Critical. That article would later grow into lengthier premium thematic coverage, which would spark winning positions such as Bloom Energy, Oklo, NuScale and our first Bitcoin Miner/AI Data Center position Core Scientific in early 2025 (the first Bitcoin miner to retrofit for AI DCs). When we say we work hard to be early to new trends, we mean exactly that. 

While networking can get bogged down in jargon and specs, and suppliers can be dropped quite quickly in the ever-shifting landscape, energy presents an entirely different investing landscape.  

The problem that AI energy stocks seek to solve is simple to conceptualize compared to the intricacies of networking. Once a power solution is confirmed, the chances it is dropped from a qualified supplier list in the same manner as a networking stock is less likely. Think of energy to AI as water to humans – a daily essential that is less about the competitive landscape and more about the sheer necessity for survival.  

Hyperscalers are spending hundreds of billions of dollars annually on AI data center capex, from physical data center space, GPUs and servers, hardware and networking. With these substantial sums flowing towards GPUs that are now being refreshed on an annual cadence, the impetus for hyperscalers, neoclouds and other cloud providers turns to how quickly these GPUs can secure power and be deployed.  

There are three reasons the race is incredibly fierce to power these new systems:  

The first reason is the stock market, as the current capex numbers are significant, and investors will want to see a return on this investment. If a company like Microsoft buys tens of billions of Nvidia’s Blackwell GPUs, the longer the massive investment in GPUs waits for power, the more delayed that revenue and profits become.  

Secondly, for competitive reasons and to keep up with Nvidia’s product road map, the next generation of GPUs will arrive every 1-2 years and Big Tech will want to maximize returns before the next generation comes online. Competitors who can energize the newest generation of GPUs faster will have a critical head start over those that are waiting for power. But there’s a catch, every generation of GPUs requires more power – for example, there will be a 5X increase between the power requirements of the GB200 NVL72 and the Vera Rubin NVL576 “Kyber” rack over a two-year design timeframe. These figures do not include networking, interconnects, cooling and other hardware, which will further boost power draw per rack. 

Third, 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, 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. 

These three reasons are simple in concept, yet the lack of power having vast consequences cannot be overstated if you combine the sheer size of investments being made in AI alongside fierce, heightened competition.  

The push to 600kW racks over the next few years means this is not a transient problem, rather it is one that the industry will continue to face, meaning continuous new construction may be needed to handle surging power demand. For example, Vantage’s upcoming 1.4 GW campus in Texas for Oracle is designed for ultra-high density racks up to 250kW, yet this will not be enough power to able to host NVL576 racks in just two to three years’ time. Additionally, a former Microsoft Azure AI executive reportedly said he estimated that the “requirements in terms of power for the data center would probably at least double every three years and maybe exponentially so over a period of time,” further reinforcing this. 

AI Accelerators: 

AI accelerators such as GPUs and custom silicon need no introduction – raw compute and compute performance has driven the AI market up to this point. The analysis would be remiss to not acknowledge the trend that is so powerful, it is displacing the FAANGs of the last decade with Nvidia firmly the world’s most valuable company now and Broadcom within striking distance of passing up companies like Meta.  

As the market weighs the so-called AI bubble, there are many disparate facts thrown at investors: dot-com fears, China/tariff concerns, stock pullbacks when there are minor announcements, and things like circular investments from OpenAI. 

Forgive me if I sound repetitive, but what truly matters is Big Tech capex. This is the single, most important number as it far outweighs the importance of earnings reports, fiscal year guidance, Nvidia’s networking growth or their product roadmap, if AMD has a new deal from OpenAI, Oracle’s insane RPO, Broadcom’s networking chips and custom silicon announcements – all of the above is being single-handedly driven by Big Tech’s large capital expenditure (capex) budgets.  

Let me throw a few stats at you.  

• Analyst estimates cannot keep up with the capital expenditures being spent on AI infrastructure. This time last year, the expectations were for $250 billion in Big Tech capex. Morgan Stanley later forecast $300 billion in Big Tech capex for 2025. That number stands at $365 billion for 2025 with one quarter left to go.  
• It’s easy to tune out the words “big tech capex” at this point but zoom out for a minute and consider that Big Tech’s TTM capex was $24B at the start of 2015, or up 15X over ten years.  
• In terms of the opportunity looking forward, McKinsey is predicting 3.5X growth in gigawatts for AI data centers between 2025-2030. The costs associated with AI data centers range from $3 trillion to $8 trillion, or about $5 trillion at the midpoint. This correlates to about 3X growth if we assume the current run rate is $1.8 trillion at the current capex of $365 billion. 
• On a more near-term basis, Goldman Sachs sees hyperscaler capex increasing sharply through 2027 – capex is projected to be $1.15 trillion from 2025 through 2027, more than double the $477 billion spent from 2022 through 2024. 
• Going back to the first point, analysts thus far have missed the mark in their estimates. Every quarter, sell side analysts rush to update their models. We are penciling in that 3x is a baseline to work with over a 5-year time frame. 

To be objective, there are analysts calling for a stock market crash based on the risks around the consumer and a GDP that is propped up by capex spending.  

Stifel stated in August: 

“While the capex boom around AI temporarily supports GDP and asset prices, Stifel forecasts this bump will fade as corporate tech spending plateaus. Such a build-out, after all, occurs only once, while consumer spending power is entering a lull that could expose markets to abrupt correction.” 

There is weight to what Stifel is describing, which is why tariffs remained a risk on our last Top 15 report and remain a risk for this report, as well. You can read more here about how the consumer is fairly weak under the hood, and how capex spending is creating a false impression that GDP is stronger than it is. 

Where I disagree with Stifel is the idea that “such a build-out, after all, occurs only once” This is an antiquated opinion where perhaps the parallels between AI and electricity have gone too far, or perhaps it’s leftover from the cloud era where there were digestion periods every few years.  

Portions of the AI buildout will be occurring every 1-2 years with new generations of AI systems. Therefore, AI is less of a static end point that is readily achieved – and rather it is an evolving architecture that enables ambitions to expand each year. Although cloud was also architecture-driven, it reached its end goal rather quickly in terms of driving down costs and improving productivity, allowing companies to quickly scale, and providing pay-as-you-go compute and services rather than significant up-front costs from on-premise servers. The end goal for AI is far more ambitious, as it could take a decade or more before Big Tech accomplishes commercially viable AGI (general artificial intelligence).  

When listening to commentary on earnings calls, in sharp contrast to what macro analysts are saying, what we hear is that Big Tech management teams are nearly in a panic to add more capacity. This one is from AWS’ Andy Jassey: “The faster we grow, the more CapEx we end up spending because we have to procure data center and hardware and chips and networking gear ahead of when we're able to monetize it. We don't procure it unless we see significant signals of demand.”  

Although the tone on capex can temporarily shift from time to time, the risk that capex will dry up from a one-time buildout is low. In many ways, the greatest risk to AI isn’t within the AI economy itself, but in broader macro conditions. We are seeing many macroeconomists attempt to forecast an AI downturn, yet having followed tech cycles for years, I’ve learned that while the macro economy waxes and wanes, technology consistently resurfaces as the best way to participate in the market — and this is especially true for AI.  

For years, there’s been a debate on whether Big Tech’s AI spending will translate to revenue and profits. Meanwhile, during those years, the I/O Fund has been laser focused on where that AI capital is actually being allocated. Rather than thinking of our approach as the picks and shovels for those chasing a gold rush, we think of it as an “AI stack” strategy—investing in the lesser-known layers and components that are driving forward an ecosystem capable of massive GDP.  

With that, I present my current Top 15 AI Stocks List. 

#1: Nvidia is Simultaneously Shipping Two GPU Systems 

Fundamentals: 10/10
Thematic: 10/10
Valuation: 4/10 

Brief Overview: 

When looking at YTD performance, Nvidia has returned 35% compared to SMH at 42%. Admittedly, it’s not great to underperform your industry. Peers within AI have outperformed Nvidia YTD by a wide margin, such as AMD and MU outperforming 2-3X as they’re up 93% and 135% on a YTD basis. 

I was recently asked on Fox if AI chips are due for consolidation. My reply was that any dips should be bought. Although I talk a lot about Nvidia and AI, my process for covering this stock has not changed since the first day I spoke of the company. My process is that we stick close to the GPU releases rather than the earnings reports as we are dealing with the world’s best design company (fact). Therefore, revenue is a step-function of a new design being released. That may seem overly simplistic, but I can assure you, it’s worked quite well for this stock. 

You can find a Case Study on Nvidia and our library of research here.library of research here. 

While AI hardware players must contend with competitive forces and supply chain qualifications that can change their positioning very quickly, that is not the case with the King of Parallel Processing. Rather, Nvidia’s near-monopoly has been built carefully as each GPU release has virtually zero competitive pressure. 

Generally speaking, Blackwell and Blackwell Ultra are shipping simultaneously. There are nuances to this statement, such as the GB200 NVL72 systems began shipping in volume in Q2 2025 and GB300 NVL72 began shipping this quarter and will ship in volume next quarter in Q4 2025.  

I want to keep this really simple as the enormity of what Nvidia is shipping would be easy to miss. The market won’t be caught off guard like it was two years ago – we all know who the best slugger in the game is. However, there are two home runs lining up (dare I say, a grand slam) as Blackwell and Blackwell Ultra ship within 6 months of each other. In this analogy of a baseball game, Nvidia has been in the dugout for a few quarters now with the Blackwell delay. I am still in the front row of the stadium awaiting the number one slugger to return to the plate.  

Overall Revenue Growth: 

Nvidia reported $46.74 billion in revenue in Q2, slightly ahead of estimates for $46.13 billion. This corresponded to growth of 55.6% YoY, decelerating more than 13 points from 69.2% in Q1; on a QoQ basis, revenue increased just 6.1%, slowing from 12% in Q1. 

However, its Nvidia’s guide for this upcoming quarter that solidifies the stock as our top pick in AI chips for now. The guide for $54 billion in revenue corresponds to 53.8% YoY growth and a rebound to 15.5% QoQ growth. 

As you’ll see below, Broadcom is expected to report higher QoQ growth yet it’s the scale at which Nvidia is putting up these numbers that separates the company from its peers. 

AI Segment Growth: 

Nvidia’s data center revenue increased 56% YoY and 5% QoQ to $41.1 billion, a marginal miss versus estimates for ~$41.2 billion in the quarter. This is also the smallest sequential increase since Hopper’s breakout quarter at just ~$2 billion.  

While compute declined (1%) due to timing of shipments and loss of China revenue, networking was up a whopping 46% QoQ and 78% YoY to $7.25 billion. As stated above under the AI Networking thematic section, this spells good things for what is in the pipeline. 

Earnings: 

Nvidia reported just a 3.9% adjusted EPS beat in Q2, reporting $1.05 in earnings versus estimates for $1.01. This corresponds to growth of 54.4% YoY, rebounding substantially from Q1’s H20-affected 32.8% growth.  

Adjusted EPS growth is expected to remain strong through the rest of the fiscal year, at 47.2% and 53.2% in Q3 and Q4. 

Margins: 

GAAP operating margin was 60.8% in Q2, improving nearly 12 points QoQ and coming in 1.7 points ahead of guidance for 59.1%. Adjusted operating margin was 64.5%, also up nearly 12 points QoQ and 1.4 points ahead of guidance for 63.1%. 

Cash: 

Free cash flow was $13.45 billion, down from $26.14 billion in Q1. FCF margin was 28.8%, again down more than 30 points QoQ and more than 16 points lower YoY. 

Typically, Nvidia has very strong cash flows and this is not a concern, rather is a transient quarter for cash. Operating cash flow margin and free cash flow margin have been in the 60% range in recent quarters.  

Valuation: 

There is not a major takeaway based on valuation. 

Forward PS Ratio: 

Nvidia trades at 20 forward PS and is a strong buy in the 10 forward PS range yet is a strong sell in the 30 forward PS range. We are in the middle of the clear buy/sell indicators for sales valuation.  

Forward PE Ratio: 

Nvidia trades at 40 forward PE ratio and is a strong buy in the 20 forward PE range yet is a strong sell in the 50 forward PE ratio.  

Notable Risks: 

Nvidia has far fewer risks than other stocks in this report. China tariffs can affect peers in the supply chain especially since there are hundreds of components in each AI system.  

#2: Broadcom: Well-Deserved Second Place Contender 

Financials: 8/10
Thematic: 10/10
Valuation: 3/10 

The networking opportunity that Broadcom is positioned to capture has been evolving every quarter to where Hock Tan himself has not been able to correctly anticipate its size. Here is what Tan stated on a recent earnings call: 

“In fact, the increased density in scale up is 5 to 10x more than in scale out. And that's the part that kind of pleasantly surprised us and which is why this past quarter, Q2, the AI networking portion continues at about 40% from what we reported a quarter ago for Q1. And at that time, I said I expect it to drop. It hasn't.” 

This quote illustrates a few things – the strength of the networking market is surprising even to Broadcom, it helps to quantify scale-up versus scale-out in terms of networking components, and it shows that (likely) this is not priced in yet as it’s a relatively new inflection. In fiscal Q2, networking growth was 170% YoY although it’s not expressly broken down in earnings reports. 

Regarding custom silicon, the chances ASICs can keep pace with Nvidia on sheer compute is low (to nearly impossible). When it comes to raw compute density combined with the software ecosystem that Nvidia offers, Big Tech’s side projects may never catch up. The rapid product road map that Nvidia offers deepens the moat the company has firmly established with universal CUDA. Meanwhile, custom silicon programs can take years to fully develop and move into production.  

So then why are Big Tech companies turning to Broadcom for less flexible (yet highly optimized) AI chips with product road maps that are considerably longer than Nvidia’s? Decades ago Jeff Bezos stated “your margin is my opportunity,” referring to the fact that Amazon’s value proposition was to offer goods at a lower cost to consumers. Broadcom is similar, in that Nvidia’s margin is their opportunity to offer AI accelerators at a lower cost. The same can be true for AMD’s value proposition, as well, especially as we enter the inference stage of the AI market. Nvidia’s gross margin of 72% (and up to 78% about a year ago) is attractive to investors who seek a quality stock, yet the margin is also communicating gluttonous pricing power.   

According to a recent article by VentureBeat, industry conversations and analysis suggest that “Google may be obtaining its AI compute power at roughly 20% of the cost incurred by those purchasing high-end Nvidia GPUs. While the exact numbers are internal, the implication is a 4x-6x cost efficiency advantage per unit of compute for Google at the hardware level.” 

Where it becomes attractive to drive down costs is the inference market. Hundreds of millions of users interact daily with AI assistants, causing inference to become the focal point for providers such as OpenAI and Google. Meeting these levels of growing demand, without significant response delays or downtime, requires more and more accelerators, networking and interconnect products.   

Broadcom’s edge goes beyond the fact that custom accelerators are often multiples cheaper than Nvidia’s GPUs for inference tasks – it's that custom silicon is increasingly performant with each generation. By optimizing algorithms (software), Big Tech can drive higher performance from large language models (LLMs) while continuing to use Nvidia’s compute power excellence for training (and also some inference tasks). 

Lastly, the VMWare acquisition has been particularly fun to watch as it’s one of the best execution M&A moments recently. A few quarters back, Tan stated VMWare was the “star of the show” as it’s been reporting accelerating bookings and backlog. Here is why it’s done well post-acquisition: “This allows customers to deploy their AI models on-prem. And wherever they do business without having to compromise on privacy and data — in control of their data. And we are seeing this capability drive strong demand for VCF, from enterprises seeking to run their growing AI workloads on-prem.” 

Overall Revenue Growth: 

Q3’25 revenue was $15.95 billion, beating estimates for $15.82 billion, and reflecting top line growth of 22.0% YoY and 6.3% QoQ. Looking ahead, management provided Q4’25 guidance of $17.4 billion of revenue, implying 24% YoY growth and a slight uptick to 9% QoQ growth. 

AI Segment Growth: 

Semiconductor Solutions accelerated nine points to 26% YoY growth due to a rebound in AI accelerators (+63% YoY). Within this, AI Semiconductor revenue surged 63% YoY to $5.2B, showing re-acceleration after a slower Q2 (+46% YoY). AI now represents 57% of Semiconductor revenue and 32% of total company revenue.   

Management guided Q4 AI revenue to $6.2 billion, which would represent ~19% sequential growth and eleven consecutive quarters of YoY growth. 

Earnings: 

Non-GAAP EPS growth of 38% outpaced revenue growth of 22%. EBITDA margin was 67%. 

Margins: 

GAAP Gross Margin of 67.1%, down 90 bps QoQ from 68.0% in Q2’25, essentially flat from 66.8% in Q3’24. 

GAAP operating margin of 36.9%, down 190 bps QoQ from 38.8% in Q2’25, but up significantly from 30.3% in Q3’24.  

Non-GAAP operating margin of 65.5%, slightly up QoQ from 65.3% in Q2’25 and up 180 bps from 63.7% in Q3’24. 

Cash: 

Free Cash Flow of $7.0B represents a free cash flow margin of 44.0%, compared to 42.7% in Q2’25 and 35.6% in Q3’24. 

Valuation: 

Broadcom’s valuation is in unchartered territory.  

Broadcom trades at a 52 forward PE ratio and has traded as low as 13 forward PE two years ago and as low as 21 forward PE in the April rout.  

Broadcom trades at a 26 forward PS ratio and has traded as low as 10 forward PS in April and was at 6.5 two years ago.  

Notable Risks: 

Valuation is the predominant risk as Broadcom has never traded at these levels in its multi-decade listing history. 

#3: AMD: The Dark Horse is Leaving the Stable 

Financials: 5/10
Thematic: 10/10
Valuation: 5/10 

We are finally seeing evidence that the Dark Horse is leaving the stable.  

The company secured a long-term deal with OpenAI to supply 6GW of GPUs with the first GW to be delivered in H2 2026. We saw a flurry of sell-side activity with one analyst raising their price target from $185 to $310 stating the Open AI deal could generate $80 billion in chip revenue for AMD over the next few years.  

As stated above under the Broadcom section, this is not a matter of AMD offering the best end-to-end performance. That will remain Nvidia for the foreseeable future. Rather, this is about driving down costs for Big Tech (and Open AI) while focusing less on raw compute power for training and more on memory and throughput for inference. 

If we zoom out (as I like to do), you may recall the MI400s are expected to be the moment that AMD tightens the competition with Nvidia. The MI400 series will be the start of rack-scale systems for AMD, starting with Helios, which will connect up to 72 GPUs similar to Nvidia’s NVL72 systems.  

According to AMD, Helios will “deliver up to a 10x generational performance increase for the most advanced Frontier models, and we believe it will be the highest-performance AI system in the world when it launches.” The last part is doubtful, yet the effort to close the gap with Nvidia will likely go a long way when coupled with lower pricing.   

AMD stated in their last earnings report they have an ambitious goal of reaching tens of billions in MI400 sales. Investors should take note that management is specifically calling out the MI400 for this, arriving in H2 2026. The readthrough is that OpenAI is an early validator that the MI400s have serious chops, and where OpenAI goes, the rest of the AI market tends to follow. 

Overall Revenue Growth: 

AMD reported a slight beat on the top line at $7.685B in revenue compared to estimates of $7.43B. This represents growth of 31.6% compared to growth of 27.4% expected. 

AI Segment Growth: 

Last quarter, management had stated, “we expect data center segment to decrease due to the exclusion of MI308 revenue.” Therefore, it was not a surprise when data center was down (11.8%) QoQ yet was up 14% YoY for revenue of $3.24B. 

Earnings: 

EPS was in line with expectations at $0.48 yet was down (30%) from $0.69 in the year ago quarter. The company is expected to rebound quickly with EPS of $1.15 next quarter. 

Margins: 

Operating margin of (2%) for operating profits of ($134M) also included the $800M in inventory changes. The adjusted operating margin of 12% was guided correctly and was in line with expectations. 

Cash: 

AMD’s cash flow margins sustained well at 20% operating cash flow margin compared to 13% last quarter and 10% OCF margin last year. Free cash flow margin of 15% also expanded from a year ago at 8% margin and up from 10% FCF margin last quarter. 

Valuation: 

Forward PS Ratio: 

AMD is trading at 13 current PS with the stock failing to hold 14 current PS two times in the past (precedes a larger selloff). The forward PS ratio of 11.7 is at the stock’s peak forward PS ratio of 12. 

Forward PE Ratio: 

AMD is trading at 61 forward PE ratio, the highest the stock has traded since the AI boom began in early 2023.  

Notable Risks: 

As the contender to the world’s most valuable company, AMD has execution risk. The company’s lead in Data Center CPUs are often at risk due to companies like Nvidia wanting to cut down costs on the instructions side of AI systems.  

The valuation remains the most notable risk. 

#4: Micron Quietly up 120% YTD 

Financials: 8/10
Thematic: 10/10
Valuation: 7/10 

Micron deserves a second look as the company is no longer tied to consumer device cycles. Instead, high bandwidth memory (HBM) had led to higher margins and multi-year supplier agreements, resulting in a leveraged approach to participate in the AI infrastructure buildout.  

As pointed out in our free analysis last week, HBM is seeing a 3.5X increase in per-GPU capacity across the last three years and AI systems are commanding an increase of 34X as the number of GPUs rises and is further compounded by each GPU system requiring more HBM per package. 

• The B200 features 180GB of HBM3e content, more than double the H100 and a 28% increase versus the H200. In an 8-GPU server configuration, the B200 boasted 1.44TB of HBM content.   
• The B300 boasts 288GB of HBM3e content, a 60% increase versus the B200 and over 3.5x more than the H100. In an 8-server configuration, the B300 has 2.3TB of HBM content. This chip is beginning to ship now in Q3-Q4 2025.  
• Putting in context Nvidia’s rack-scale solutions, the GB200 and GB300 NVL72, shows just how rapidly HBM content is increasing. The GB200 supports up to 13.4TB of HBM content, while the GB300 supports up to 21.7TB of HBM, nearly 34X higher than the 640GB of HBM content in the 8-GPU DGX H100 servers. 

The line in the sand for AI hardware companies is the margins and Micron has performed beautifully in that regard. I have to admit, when I saw their last earnings report at end of September, I had to look twice to make sure it was really Micron. 

Overall Revenue Growth: 

Micron reported record FQ4 revenue of $11.32 billion. Revenue growth accelerated 9.4 percentage points sequentially to 46% YoY, and on a sequential basis, growth was 21.7% QoQ, a solid 6.2-point acceleration. 

AI Segment Growth: 

FQ4 DRAM revenue grew by 69% YoY and 27% QoQ to $8.98 billion, a second consecutive quarter of strong sequential growth. 

Earnings: 

In Q4, Micron reported adjusted EPS of $3.03, up 157% YoY and beating estimates by 5.9%. 

For Q1, Micron guided for adjusted EPS to be $3.75, +/- $0.15, more than 23% ahead of consensus and corresponding to YoY growth of 110%. Earnings growth is expected to reaccelerate to 155% in Q2 but then decelerate to 126% in Q3 (but still growing handsomely). 

Margins: 

Micron’s margin turnaround story has been impressive, with gross margin up more than 55 points over the last two years and operating margin up more than 66 points.  

Adjusted gross margin in Q4 was 45.7%, up 6.7 points QoQ and 9.2 points YoY, aided by strong growth in CMBU which carried a 59% gross margin in the quarter, DRAM pricing, favorable product mix, and cost controls. For Q1, adjusted gross margin was guided to be 51.5% at midpoint, a 5.8 points sequential expansion and up by a solid 12 points YoY.   

FQ4 adjusted operating margin was 35%, up 8.2 points QoQ and 12.5 points YoY, driven by operating leverage. 

Cash: 

FQ4 adjusted free cash flow grew by 149% YoY to $803 million or 7.1% of revenue, an improvement of 2.9 percentage points YoY. Management expects adjusted free cash flow to strengthen in FQ1 and to be significantly higher in FY2026. 

Valuation: 

Micron trades at an attractive valuation of 4.2 forward PS. The stock has traded as low as 2 forward PS and as high as 7 forward PS. 

Micron trades at 12 forward PE ratio. The stock has traded as low as 7 yet as high as 124 due to the lumpy bottom line from the previous cyclical low in 2023 timeframe. 

This goes back to the debate on if MU is a cyclical stock that deserves a lower valuation or is it emerging as a major, secular AI player. Should it be the latter, there is quite a bit of room in the valuation.  

Notable Risks: 

If Micron announces set pricing like they did in 2024, the stock could plateau. There are fierce competitors in the space, such as SK Hynix and Samsung. If pricing proves cyclical, the current valuation will not hold. If the pricing proves more of a secular trend, there is ample room in valuation – how the market will view the stock 2026-2027 is not clear although with technicals, some of the risk can be mitigated.  

#5: TSM Report Provides 5-Quarter Runway 

Financials: 5/10 (HPC declining QoQ)
Thematic: 10/10
Valuation: 4/10 

TSM is typically off cycle from AI semiconductors, meaning, we will see a boom in the high-performance computing segment about 5-6 quarters before we see a boom in the AI chip market.  

It would be easy to assume TSM is a quarter or two ahead of shipment times when, in fact, it’s more like three quarters when factoring in HBM and CoWoS packaging. From there, it takes an additional two quarters for system integration and assembly from companies like Dell, Foxconn, and Supermicro before the servers are shipped. 

Perhaps second to capex, TSM can be used as a strong proxy for the health of the AI market as the common denominator across AI chips. The HPC segment is communicating that we have a few quarters of strong growth in the pipeline. From there, we will need to monitor how long the QoQ decline in HPC lasts as a couple of quarters is typical; anything longer would be a concern to monitor further.  

On the topic of timing, the chart below helps to illustrate that even though AI is a secular trend, due to shipping cycles, TSM and even MU can still see cyclical results. When a new generation is ramping, TSM will naturally see the results first as will Micron before the systems are shipped. If we draw a similar parallel to the last QoQ decline in TSM’s shipping cycle, we can see that putting our money in the AI Chips companies makes more sense right now. 

Pictured Above: When TSM’s HPC segment declined for three quarters in the past, the stock underperformed compared to a stock like Nvidia, which was reaping the benefits of the new generation finally shipping in volume (Hopper). 

Overall Revenue Growth: 

Q3 revenue grew by 40.8% YoY and 10.1% QoQ to $33.10 billion, beating the guidance midpoint by 2.2%. TSMC boosted the full year revenue guidance by 5 percentage points for the second consecutive quarter to mid-30% on continued strong AI demand. This is up from the 30% growth provided in Q2.   

For Q4, TSMC guided revenue of $32.2 billion to $33.4 billion. At midpoint of $32.8 billion, this represents a YoY growth of 22% and down (0.9%) sequentially.   

AI Segment Growth: 

TSMC stated that HPC revenue was flat QoQ in NT$ in Q3, though there was more pronounced increase on a US$ basis due to FX. TSMC’s revenue is recognized in US$, so every 1% appreciation of the NT$ adversely impacts NT$ reported revenue by ~1%. Management sounded very optimistic during the Q3 earnings call about long-term AI growth opportunities. 

Earnings: 

The company’s Q3 EPS grew by 50.5% YoY to $2.92, beating estimates by an impressive 12.3% with the strongest beat in the last two years. Analysts expect Q4 EPS to grow 26.8% YoY to $2.84 in Q4 and grow 25.5% in Q1. Looking forward, they expect EPS to grow 19.8% YoY to $12.34 in 2026 and 24.6% YoY to $15.48 in 2027. 

Margins: 

Margins continue to expand due to cost controls, higher capacity utilization rates, economies of scale, and better price negotiation with customers and suppliers.   

The company’s gross margins improved 170 basis points YoY and 90 basis points sequentially to 59.5%. Cost improvements, better capacity utilization, and better price negotiation with customers and suppliers primarily drove the strong margins. 

Q3 operating profits grew by 50% YoY to $16.74 billion, with an operating margin of 50.6%, an improvement of 310 basis points YoY and 100 basis points sequentially, primarily driven by higher gross profits and operating leverage. 

Cash: 

Despite higher capex, cash was also strong with operating cash flow at 43.9% compared to 51.6% in the same period last year. Q3 free cash flows were down (16.1%) YoY to $4.8 billion or 14.6% of revenue compared to 24.4% in the same period last year. The free cash flows were down due to higher capex which grew by 51.6% YoY to $9.7 billion to support strong further growth.   

Valuation: 

Forward PS Ratio: 

TSM is trading at 13 forward PS ratio compared to 6 at the April low and 6 at the start of the year. Regarding the current PS ratio of 17 – I cannot find a higher valuation going back 10 years.  

Forward PE Ratio: 

TSM is trading at 31 forward PE ratio, the highest it’s been going back to early 2023 during the AI boom.  The current PE ratio of 35 is the highest its been going back to ten years except on rare exception briefly at the 2021 top.  

Notable Risks: 

Similar to Broadcom, the predominant risk is valuation.  

AI Networking Stocks: 

My number one ranked trend is AI networking. As discussed in the AI Chips section, stocks such as Nvidia and Broadcom are also the networking leaders. However, the below stocks are networking pureplays, with many on a tear since the April lows.  

As you have likely noticed recently, the networking ecosystem is nuanced as recent announcements from Oracle/AMD and Nvidia/Intel have caused some networking stocks to plunge. To provide the bigger picture, I am revisiting the thesis and rankings for our networking stocks. 

Tied for #1: Astera Labs: Increased ASPs from Scorpio 

Financials: 10/10
Thematic: 10/10
Valuation: 4/10 

When Astera is asked why they stand apart within a crowded networking market, management responds that the drive for low latency PCIe is the primary contributor to the beat/raise across both the Aries and Scorpio products.  

Last March, Astera announced further collaboration with Nvidia by offering NVLink solutions for PCIe/CXL within servers (scale up): “Most recently, Astera Labs demonstrated the industry’s first end-to-end PCIe 6 interoperability with Scorpio P-Series Fabric Switches, Aries 6 Retimers and a NVIDIA Blackwell GPU at NVIDIA GTC 2025. Scorpio P-Series Fabric Switches have also been integrated with the NVIDIA MGX platform for PCIe 6-ready modular designs.”  

Launched only this year, Scorpio now exceeds 10% of total revenue “making it the fastest-ramping product line in Astera Labs’ history.” Keep an eye specifically on Scorpio as a GPU-to-GPU scale-out and scale-up product line. 

On the custom silicon side, it’s widely understood that Astera supplies Amazon as there were disclosures around Amazon having a warrant to buy shares in exchange for guaranteeing $650 million in orders in the SEC filing. As of mid-2025, GuruFocus confirmed Amazon still holds shares in Astera Labs.  

Astera’s agile ability to compete head-on with Broadcom goes beyond only Amazon and Nvidia (though certainly, those two are enough). Management stated they had design wins with ten customers including merchant GPUs and ASICs, plus a line of sight to further sales growth from their many product lines in H2 2025 and 2026, and the upcoming UALink consortium in 2027.  

It takes some time for a company like Astera to become a qualified supplier. I would need at a minimum an earnings report to state otherwise or a QoQ decline of some kind to be convinced this has changed. Instead, what I’m seeing is quite the opposite.  

Revenue: 

Astera Labs reported revenue of $191.9 million, beating consensus of $172.5 million for growth of 150% YoY and 20% QoQ. About eight months ago in November, analyst consensus for the June quarter was for 85% growth — thus the company has nearly doubled these expectations in less than a year. 

AI Segment: 

Same as revenue growth (pureplay) – up 150% YoY and 20% QoQ. 

Earnings: 

Adjusted EPS of $0.44 beat by 36%. Consensus is $0.39 for 69% growth. GAAP EPS was $0.29 

Margins: 

Astera delivered in that regard with a GAAP operating margin of 20.7% compared to 7.9% expected. This operating margin is a major win for ALAB investors as the company is now comfortably GAAP profitable despite stock-based compensation being around 20% of revenue. 

Cash: 

Astera’s cash from operations increased significantly with an operating cash flow margin of 70.5%, up from 38.7% last year. This totaled operating cash flow of $135.4M with $1.07B in cash on the balance sheet and no debt. 

Valuation: 

Astera’s valuation of 35 forward PS is in the mid-range as the company has seen as low as 11 forward PS and as high as 60 forward PS.  

The stock’s forward PE ratio of 103 is steep, yet the company is recently GAAP profitable, thus it’s hard to go by this ratio. 

Typically we use technicals in a situation where a hypergrowth stock is surging on revenue and has strong bottom line results, causing the valuation to send mixed signals. On one hand, stocks like this should be highly valued. On the other hand, how rich of a valuation are buyers willing to pay? Rather than get too stuck in the weeds, technicals can help us participate in the upside while protecting the downside. 

Notable Risks: 

Every stock has risks yet Astera less so than others on my Top 15 list. I can see a scenario where the market softens on AI valuations (temporarily) and a scenario where there are strong earnings and valuations march onward.   

Tied for #1 Credo: Active Electric Cables (AECs) for Miles  

Financials: 10/10
Thematic: 10/10
Valuation: 4/10 

The saying “I can see for miles” implies visibility into the future. The saying makes me think of Credo as there is over two miles of copper cabling for each NVL72 system. We can literally see Credo’s AECs for miles, and figuratively, there is more visibility than usual for this particular stock given the inflection of 274% YoY and 31% QoQ growth last quarter helps to confirm its positioning in Nvidia’s Blackwell systems.  

Credo’s new 800G HiWire ZeroFlap AECs are designed to reach 7 meters with full host-to-switch connectivity, and are especially designed for liquid cooled servers. The over 7 meter distance helps to enable large AI clusters sized into hundreds of thousands of GPUs.  

Credo competes with 800G OSFPs AOCs, yet these are particularly troublesome due to physical constraints that cause the connectors to break. There is also link lapse with AOCs, which are “momentary disruptions in network links.” Credo’s AECs aim to solve these issues, and the results speak for themselves. 

For distances between two meters and seven meters (or about six to 24 feet), active electric cables (AECs) are also seeing heightened demand as servers scale up to eight GPUs to now 36 GPU to 72 GPU per rack-scale AI system.  

In a nutshell, this is why Credo is reporting surging growth in a highly competitive market: “Reliability and power efficiency [leads] to choosing AECs over optical solutions as they are up to 1,000x more reliable and consume half the power. AECs virtually eliminate link fabs, which are intermittent losses of connection, boosting cluster reliability and productivity while reducing power consumption.” 

Regarding “consume half the power” … Credo’s proprietary serializer/deserialzer (Ser/Des) technology, active electric cables and digital signal processing (DSPs) give the company a significant competitive advantage as it enables power-efficient connectivity that is reasonably priced. 

Revenue: 

The company reported growth of 274% YoY and 31% QoQ for revenue of $223.1 million. This beat estimates on the top line by 17% with management raising full-year revenue growth outlook by 35 points, from 85% YoY to 120% YoY. 

AI Segment: 

Product Revenue came in at $217.1 million, up 279% YoY and 31% QoQ. 

Earnings: 

The bottom line also shined with adjusted EPS beating estimates by 44.4%. This represents growth of 1,200% YoY from a thin $0.04 in the prior-year quarter. Triple-digit growth of 425% on the bottom line is expected to follow although flat QoQ. 

Margins: 

GAAP Operating Margin was 27.2%, up from 19.9% in the last quarter and up from (24.2%) in prior-year quarter.  

Adjusted Operating Margin was 43.1%, up from 36.8% last quarter and up from 3.7% in prior-year quarter.  

This is the standout – massive operating leverage as opex grew only ~11% QoQ vs. the 31% pick up in revenue. 

Cash: 

FCF Margin of 23.8%, down from 31.9% last quarter but up more than 45 points from (21.9%) in the prior-year quarter. Debt free. 

Valuation: 

The forward PS ratio is 24 and on the higher range of where Credo trades with the upper region being 33 earlier this year yet has traded as low as 8 at the April low. 

Credo trades at a 65 forward PE Ratio yet similar to Astera Labs, this is hard to put much weight into as the company is newly profitable.  

Notable Risks: 

Every stock has risks yet Credo less so than others on my Top 15 list. I can see a scenario where the market softens on AI valuations (temporarily) and a scenario where there are strong earnings and valuations march onward.   

#3: Small Cap Networking Stock with Strong QoQ 400G Growth and Incoming 800G/1.6T Growth 

This past quarter, the I/O Fund was on the hunt for networking stocks that reported an inflection. Given Nvidia is reporting 46% QoQ growth at scale on their networking segment, we figured there would be some breadcrumbs to follow in the supply chain as to companies that are beneficiaries of the incoming AI networking boom. 

The stock we identified for our Discovery members reported 40% QoQ growth last quarter and is forecasting 17% QoQ growth in the upcoming quarter – some of the highest we’ve seen in the supply chain landscape. Management discussed the ability to grow capacity 8.5X this year before doubling this by mid-2026. 

To learn more, sign up for our Discovery tier here. 

#4 Lumentum: EML Lasers in High Demand 

Financials: 6/10 (not a pureplay)
Thematic: 7/10
Valuation: 7/10 

Lumentum has been on our radar for more than one year, as the company supplies components for datacom transceivers and optical interconnects with tech that has caught the attention of heavyweight Nvidia. We’ve been closely monitoring Lumentum and waiting patiently for their EML lasers for 200G to ship, enabling 800G and 1.6T bandwidths.   

As discussed in the past, optical interconnects help data centers accelerate data throughput between data centers and inside the data center between servers or racks, while reducing latency and power consumption. AI is driving cloud demand higher from the hyperscalers, leading to more data being created and processed, thus helping drive a need for these interconnects to meet demand for high-speed, low power data transmission in data centers.  

Specifically, Q4’s report provided confirmation of the EML laser ramp, as EMLs achieved an all-time high for shipments with revenue more than doubling versus its June 2024 baseline.  Management also cited a “substantial” 200G EML order to be fulfilled in the December quarter, although offered little additional clarity on the size of the order.   

Revenue: 

Lumentum reported Q4 revenue of $480.7 million, beating analyst estimates by a modest 2.29%. This was a notable uptick on the top-line, with growth of 13.1% QoQ, accelerating 7 points, and 55.9% YoY, accelerating 42 points. 

AI Segment: 

Cloud & Networking Q4 revenue came in at $424.1 million, representing 66.5% YoY growth. Additionally, the segment’s QoQ growth of 16.1% accelerated sharply from 7.6% QoQ in Q3, coming in much stronger than Coherent, where growth decelerated from 10% QoQ to 5%. 

Earnings: 

Lumentum reported adjusted EPS of $0.88 in Q4, beating estimates of $0.81 and improving against the $0.57 reported in Q3 and $0.06 reported in the year ago quarter.  

Q4’s adjusted net income margin of 13.1% reflects continued operational improvements and the fourth consecutive quarter of sequential improvement (from 9.6% in Q3, 7.5% in Q2, and 3.6% in Q1). 

Margins: 

Q4’s GAAP operating loss of ($8.4 million) represents a (1.7%) operating margin, compared to (8.9%) in Q3 FY25 and (43.3%) in Q4 FY24.  

Non-GAAP operating margin of 15.0% expanded nicely compared to prior quarter of 10.8% and prior year quarter of (5.1%). 

Cash: 

For FY25, operating cash flow was up ~5x to $126.4 million, for a 7.7% margin, improving from a 1.8% margin in FY24. 

Valuation: 

Lumentum’s Forward PS ratio is above what it typically trades at 5 forward PS. It’s previous top this year was 4 forward PS and it failed to hold that valuation twice. 

The company’s forward PE Ratio is 33 and is within a reasonable range as the company has traded as high as 50-60 and as low as 18 over the past 12 months. 

Notable Risks: 

Lumentum has many competitors and its supplier agreements are not as clear to our firm as Credo and Astera Labs. However, the fundamentals are communicating that Lumentum’s management commentary is reliable, which is that their EML lasers are in high demand.  

AI Software and AI Data Layer 

As it stands today, there are a handful of AI software companies and AI data layer companies that have set themselves apart in fundamental strength. As the introduction pointed out, the noise in stock investing is often quite loud and the rate at which stock tickers are exuberantly discussed is detached from reality. We do a fairly comprehensive scan and find the list of software stocks that offer concrete proof of participating in the AI trend is far fewer than one would originally imagine. 

I’ve stated below that AI hardware companies will also be some of the best AI software companies (incoming tangent here): 

To illustrate this, Broadcom is preparing to displace a few of the FAANGs at a $1.6T valuation compared to Meta at $1.8T valuation and Amazon at $2.3T valuation. This is on the cusp of Broadcom stating their serviceable addressable market will be $60B to $90B in revenue by fiscal 2027, which is in two years from today as their fiscal year ends in October (roughly 200% at the midpoint given the current $25B run rate).  

One has to wonder, why do the AI software juggernauts not discuss their revenue as openly as the AI hardware companies, especially given software tends to be recurring and easier to model? On the other end of the spectrum, we see up to 100X forward sales valuations in this category.  

Yet, if we look at Nvidia AI systems as an example, it’s easy to see that software is what will accelerate AI over the next few years. There are four layers to Nvidia’s full-stack accelerated computing: hardware, system software, platform software, and applications. When you consider Blackwell, for example, there are transformer engine libraries, integrated with CUDA and cuDNN that determine when to use FP8 versus FP4 during training or inference to maximize speed yet also maintain accuracy. Software determines which precision to use for every layer and every tensor, which helps to deliver a 2X to 4X improvement in training and inference. There are additional revenue segments, such as automotive and robotics, where Nvidia will be able to license its software and gradually add this high-margin revenue to its already profound stock market performance. 

If you look below, you’ll also notice that I’ve put Oracle in the software/data layer category for two quarters now. Although it's Oracle Cloud Infrastructure (OCI) driving growth of 55% and expected to accelerate to 77%+ growth and also strong RPO of 359%, it’s the software and data layer that separates Oracle from other cloud infrastructure providers. For example, software-defined remote direct memory access (RDMA) helps to lower latency and increase bandwidth by bypassing the CPU during training and inference tasks. Oracle says it “consistently charges less than Amazon Web Services (AWS) for the equivalent compute capacity.” But we have to look closely at why that is. 

Vectorized data is another area where Oracle sets itself apart using the data layer as it allows both structured and unstructured data to be understood by AI models, which helps to increase the use of private data alongside public data. 

As we approach the AI software/AI data list now and into the future, the following should be kept in mind: 

• Flexibility in what defines an AI software winner (wait until you see my #2 below) 
• It can’t rank until we see real, tangible evidence of a stock participating in the AI trend. Why pre-emptively invest in a company that is not reporting strong AI revenue when there are so many choices on the market? (Looking at you Tesla Optimus fans).  

I believe the stocks listed below fit this criteria. 

#1: Reddit is an AI Data Layer Frontrunner 

Financials: 10/10
Thematic: 7/10
Valuation: 4/10 

If you had asked me a few years back to describe Reddit’s value proposition, descriptions like “front page of the internet,” “world’s largest forum,” “best place to query the crowd” may have come to mind. Yet, there is something far more valuable that Reddit provides in the AI era, which is a continuous supply of human-generated conversations. To some extent, Reddit is transforming into a human data farm for AI models as it provides continuous, conversational data. 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.  

Therefore, Reddit ranks higher than you might originally guess going into an AI software discussion. The reason for the high-ranking is that Reddit’s forum-based, real-world discussions are at the top of the list of data sets that can help advance AI models. Therefore, this is less about forum users and more about the licensing of data (and what Reddit gets in exchange). 

Google clearly agrees as the company is licensing data from Reddit, and what’s interesting about this partnership is that it’s easy to see Google lacks highly contextual, human sentiment type data that social platforms provide. Meta, for example, has something similar to Reddit – whereas Google does not have this social aspect from search. In exchange, Google is boosting Reddit in search results. 

Although the world’s leading forum site has only 416 million weekly active users compared to Facebook’s 2 billion, it ranks fifth behind Facebook as the most visited site in the United States. In addition, due to a few changes in how Google surfaces content with AI overviews, Reddit is now the second most visible site in the United States – ranking above Facebook for example – and the top line results show the company is reaping the rewards of being in the search giant’s good favor.   

Over the last two years, Reddit has seen an explosion in SEO visibility on Google, with data from Sistrix placing growth from July 2023 to April 2024 at a whopping 1,328%. This moved Reddit from 85th most visible site to the 7th most visible.   

Now, as of October 2025, Reddit has moved to the 3rd most visible site in the US, per Sistrix, behind Wikipedia and YouTube, and ahead of popular sites such as Facebook in 7th and Amazon in 4th. This major improvement in SEO ranking may be a potential contributor to Reddit’s accelerating growth over the past five quarters – yet as stated, the surge in growth is out of Reddit’s control and relies on Google SEO placement, which could change at anytime.  

In terms of user engagement, Reddit notched 3.8 billion visits as of September 2025, according to data from Similarweb, compared to 11.4 billion for Facebook, 6.5 billion for Instagram, and 4.3 billion for X. Users visited an average of 4.77 pages per visit with an average visit duration of nearly 6 minutes, compared to 12.18 pages per visit and an average duration of 10 minutes for Facebook. Similarweb places Reddit as the fifth-most visited site in the US, behind Facebook in fourth place. 

This rise in search ranking has created a fundamentals profile that is hard to ignore. 

Revenue: 

Reddit delivered a rather impressive Q2 on July 31st with revenue beating estimates by more than 17%.  

Q2 marked Reddit’s fastest growth since the start of 2022, and a significant improvement over the past two years from just 12% growth at the start of 2023. What’s even more impressive is that Reddit delivered this 77.7% growth on top of a rather difficult 53.6% comp, yet this may shape up to be the peak growth quarter for the year as comps get tougher. 

AI Segment: 

Behind the substantial revenue beat in Q2 was 84% growth in advertising revenue to $465 million. This marked a sharp 23 point acceleration from 61% growth in Q1. Sequentially, advertising revenue grew almost 30%, with growth of more $106 million QoQ, outpacing Q4 2024’s $79 million sequential increase.  

There was 50% YoY growth in active advertisers as it continued to acquire new advertisers. Additionally, performance ads and brand ads both increased more than 80% YoY, reflecting strong engagement from advertisers on the platform. 

This past quarter marked the highest sequential growth in ARPU in more than three years at 25%, outpacing even Q4 24’s 18% growth. Management believes global ARPU is “still low on an absolute basis and remains an opportunity” for long-term improvement – for example, Meta’s global ARPU is around 3x of Reddit’s at $13.65 as of Q2, and though Meta hasn’t updated regional metrics since the end of 2023, it’s possible that US ARPU is 10x that of Reddit’s. 

Earnings: 

Over the past month, consensus EPS estimates through Q4 2026, or the next six quarters, have been revised 23% to 66% higher; over the past six months, estimates have moved 20% to 57% higher, as margins strengthen. For example, Q2 2026 has seen its estimate move from $0.42 to $0.70 over the past month, and Q3 2026 from $0.55 to $0.83. This now projects three consecutive quarters of triple-digit YoY growth followed by three consecutive quarters of >50% growth. 

Margins: 

Gross margin was 90.8% in Q2, up 1.3 points YoY and marginally higher QoQ.  

Operating margin was 13.6%, up nearly 25 points YoY and 12.6 points QoQ. Notably, this also exceeded Q4 2024’s operating margin of 12.4%. 

Cash: 

Operating cash flow was $111.3 million in Q2, up 292% YoY. OCF margin was 22.3%, down 10 points QoQ but up more than 12 points YoY. Free cash flow and operating cash flow are correlated nearly 1:1. Free cash flow was $110.8 million in Q2 for a 22.2% margin. 

Valuation: 

Reddit has some room at 19 forward PS but not a ton of room before it will test its previous top at 24. After about a 25% move from here, Reddit will be testing a level it has not held two times in the past year. The stock has traded as low as 8.3. 

The bottom line valuation is harder to identify a trend as the company was not GAAP profitable when it went public, although is firmly GAAP profitable now. The stock trades at 52 forward PE yet has traded as high as 70 forward PE and as low as 25 in recent quarters. 

Risks: 

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

#2: CoreWeave: Beating the Big 3 on Utilization Rates  

Financials: 1/10
Thematic: 11/10 (higher than our highest rank)
Valuation: 3/10 

CoreWeave is perhaps the first stock in the history of the market to have short sellers before it was a public listing. While it’s risky and novel to collaterize GPUs, we want to remain open-minded as CoreWeave’s software-defined infrastructure is one of the most advanced in the industry for AI workloads. 

Please note, CoreWeave carries outsized risk as there is high debt leverage, negative free cash flow and is not profitable. The stock’s thematic ranking is quite high, yet we would only approach this stock using technicals for risk management. 

The company’s software stack is specifically designed to maximize GPU utilization, elasticity, and cost efficiency to the point of booking out GPUs like it’s a leading hyperscaler. Remarkably, it’s the first company in almost twenty years to meaningfully disrupt the dominance of the Big Three in cloud infrastructure. 

By focusing only on GPUs and software optimizations, CoreWeave offers bare metal servers at a cost that is up to 20% to 50% cheaper than hyperscalers. Its value proposition is best summarized in its utilization rates for GPUs. CoreWeave has published its MFU rate of 35% to 45%, stating it is 20% higher than competitors, which means other AI data centers have MFU rates more in the 30% range. Due to FLOPs performing an astronomical number of calculations, small percentages translate to an important advantage.   

The company is able to scale quickly with new GPUs due to the Mission Control automation layer that provides automated deployments of systems like the GB300 NVL72s. The company stated: “Mission Control continues to be the cornerstone of CoreWeave's ability to scale at breakneck speed, building a fully automated and rigorous process for cluster life cycle management with unmatched visibility for our customers.” 

CoreWeave also offers a Virtual Private Cloud for a private network space. By combining an isolated virtual private cloud with Nvidia’s Quantum InfiniBand, customers get ultra-low latency with enhanced security. 

The company's Kubernetes Service is an AI-optimized Kubernetes environment for scheduling AI workloads and scaling up/down for the right mix of CPU, GPU, memory and storage (known as elasticity). SUNK, known as Slurm on Kubernetes, combines container orchestration with a job scheduler to manage large batch jobs. AI labs use this service to combine scheduling for high performance computing with a cloud-native environment.   

Local Object Transport Accelerator (LOTA) for AI object storage is another feature that is optimized for AI workloads by focusing on performance and cost efficiency. The company recently added archive tier object storage, which allows data to move between hot and cold storage based on access patterns, which optimizes costs. In the recent earnings call, the company stated they are seeing customers “shifting petabytes of their core storage to CoreWeave in the form of multiyear contracts.”  

CoreWeave recently completed its acquisition of Weights & Biases in May to add observability, which means engineers can quickly diagnose a failure or inefficiency in the software layer and infrastructure layer. For example, if a model is training slowly, the observability platform will help an AI engineer identify and resolve this quickly.   

More recently, CoreWeave integrated W&B for a joint launch of its Inference-as-a-service feature, which allows developers to use APIs to tap into AI models from OpenAI, Meta, DeepSeek, etcetera. Inference is key for CoreWeave to fully monetize its investments in capex-heavy infrastructure. For example, these popular LLMs combined with chain of reasoning inference, which means generating step-by-step reasoning, will become compute-intensive especially at scale. This will lead to CoreWeave monetizing every chatbot response, API calls and applications to easily payback their initial investments plus some (in time).  

The paragraph above is quite important and worth summarizing – essentially, CoreWeave’s path to monetization will become clearer as the inference market takes off in the coming quarters/years. What you see is not what you get, rather what you see could surge should CoreWeave execute well – and find financing.  

This stock comes with outsized risk. Please note the cash/debt section below. Despite being in the Pro tier’s Top 15 due to a strong thematic, this is one we will hold only with close risk management. 

Revenue: 

CoreWeave reached a new milestone of over $1.0 billion in revenue in Q2 2025, growing 206.7% YoY to $1.21 billion. On a sequential basis, the Q2 revenue grew by 23.6%. The company beat analyst consensus estimates by 12.2%, driven by strong demand for the company’s AI cloud infrastructure services. 

Looking forward, revenue is expected to grow by 174% YoY to $5.26 billion in the year 2025 and 129.6% YoY to $12.08 billion in 2026 and 48.3% growth in 2027. Most importantly, management has increased the full-year revenue guidance for the second quarter in a row due to the strong customer demand. Management increased guidance by $250 million at the midpoint to a new range of $5.15 billion to $5.35 billion for the year 2025. 

AI Segment: 

The company’s backlog was $30.1 billion at the end of Q2, up 86% YoY driven by the company’s strategic deal with OpenAI in March 2025 and the signing of subsequent expansion deals with the company. 

The company derived 77% of 2024 revenue from its two largest customers, i.e., Microsoft and Nvidia. While in the recent quarter, Microsoft accounted for 71% of the total revenue. Goldman Sachs estimates that Microsoft’s share is expected to drop to 38% in 2026, followed by OpenAI at 21%, Nvidia at 6%, and the remaining 35% to be attributed to other customers.   

Earnings: 

The company reported GAAP loss per share of (-$0.60) in Q2 compared to the analyst consensus estimate of (-$0.49), missing estimates by –21.7% due to the higher operating expenses, particularly the technology and infrastructure expenses.  

Analysts expect GAAP loss per share of (-$2.67) for this year, followed by (-$0.90) for 2026. They expect a positive GAAP EPS of $1.59 in 2027.

Margins: 

Q2 gross margin was 74%, up 200 basis points YoY and up 100 basis points QoQ. The company is investing heavily in data center and server infrastructure to meet very strong AI demand from its customers. The operating margin was 2%, compared to 20% in the same period last year and (3%) in Q1. Operating expenses increased by 276% YoY, driven by high technology and infrastructure expenses. The management tried to explain in the Q2 earnings call that expenses are front-loaded and have a short-term impact on the margins. 

Cash: 

CoreWeave’s business model is based on aggressive capacity expansion, currently fueled primarily by debt. As a result, cash is rather thin and gets spent quickly, and free cash flow is widely negative.   

CoreWeave reported $1.15 billion in cash and equivalents (excluding $0.56 billion in restricted cash and equivalents), though CoreWeave updated in an 8-K related to its now upsized $1.75 billion raise that total cash will be closer to $5 billion.   

Operating cash flow was ($251.3 million) for a (21%) margin in Q2, widening from ($117.8 million) in the year ago quarter. Free cash flow was ($2.7 billion) for a (223%) margin, widening slightly from ($2.36 billion) in the year ago quarter.  

Debt was reported at $11.05 billion in Q2, with $3.62 billion being current. Current debt is likely closer to $12 billion now, with a majority (~$6.7 billion) tied to its two existing delayed draw term loan facilities; if the new DDTL is drawn upon, debt could rise to $14.6 billion. On the other hand, the company is trying to reduce its cost of capital by raising cash through secured debt, using its highly valuable GPUs as collateral, which is positive.  

As discussed above, capex for the second half of the year is expected to be >$15 billion, with cash on hand only covering one-third of that at maximum. This will place the emphasis on finding alternative funding to finance this spending. 

Valuation: 

CoreWeave’s valuation of 12.4 forward PS is at the higher range for where the stock has traded this year. The stock traded as high as 15 forward PS and as low as 3.4 during April.  

The company is not profitable for a bottom line valuation and is the only stock on our list that is not profitable.  

Risks: 

Cash/Debt is a very high risk with debt at $12 billion and growing with only $1.15 billion in cash and negative free cash flow.

#3: Oracle Inflecting with 8% QoQ Growth, RPO of 359% Soars and The Future is Bright 

Financials: 7/10 (other segments weigh on AI segment)
Thematic: 7/10 (not a pureplay)
Valuation: 2/10 

As stated in the intro, even though bare metal servers are a large part of the story for both CoreWeave and Oracle, it’s also the software-defined optimizations that set these cloud infrastructure players apart from the Big 3. For Oracle, it’s also RDMA, which is a networking fabric that increases performance and density, plus the data layer (of course) as Oracle is able to leverage its deep roots in relational databases. Inference will need a mix of private data sets to augment public data sets to fine-tune reasoning tasks in an effective way for enterprises to use the models internally, and Oracle is positioned to capture this opportunity.  

Oracle offers the widest range of bare metal GPU instances among major cloud providers, and scalability at any size up to 65,536 Hopper GPU clusters and 131,072 B200 GPU clusters, which are expected to come online in 2025. Oracle also offers very flexible VM instances, letting customers pay for only the capacity they need as they need it for any size workload, rather than offering fixed instance sizes. 

With less overhead and fewer CPU cycles, RDMA helps Oracle offer its AI clusters at a lower cost: Oracle says it “consistently charges less than Amazon Web Services (AWS) for the equivalent compute capacity.”   

Oracle says that it can offer less than 10 microseconds of latency between nodes, improving efficiency. In the most recent earnings call, Oracle emphasized how cheap they are compared to the Big 3, stating: “We have gotten the entire Oracle Cloud, the whole thing, every feature, every function of the Oracle Cloud down to something we can put into a handful of racks, 3 racks, we call it Butterfly that costs $6 million. So we can give you a private version of the Oracle Cloud with every feature, every security feature, every function, everything we do for $6 million. I think the cost for the other hyperscalers is more than 100x that.” 

Oracle’s AI vector capabilities also stand out given Oracle’s database roots, offering native AI vector search capabilities with seamless integration to leading AI models from OpenAI, xAI, Meta, Cohere and more. AI vector search lets enterprises search both structured and unstructured data in a variety of manners, enabling intelligent, relevant and accurate AI responses utilizing their data. 

The announcement of Oracle’s AI database is particularly interesting in terms of ways the stock can extend its run. As explained in the earnings call, the combination of vectorizing data to where it can be understood by AI models with the ability to connect private databases to AI reasoning models will result in enterprises unlocking higher value from AI. 

Oracle is teasing a more beefed-up AI database, which management stated will officially launch at Oracle World Cloud, describing a combination of private enterprise data, large reasoning models and automated agents: “Who's offering that to customers? We'll be the first when we deliver it and demonstrate it at AI World next month.”  

Oracle has already made major headway with AI embedded databases with 23ai, which converts vector data into contextual information. By connecting a database to Chat-GPT, there is more reasoning layered into the results.   

The inference market will define by size and quality of data for reasoning purposes, and Oracle sits on arguably the world’s largest enterprise data sets. Although we have grown used to compute driving the AI training market, there will be an important shift toward the data layer driving the inference market.   

With Oracle embedding the AI database, inference will happen inside the database where the data resides. This is distinct from pulling data out of the database into the large language model, which is inefficient. Oracle’s move to embed the database supports a sustained, upward trajectory in the stock price. 

Revenue: 

Oracle delivered Q1 revenue of $14.9 billion, growing 12% YoY but slipping 6% sequentially, coming in just shy of the Street’s $15.0 billion estimate. 

AI Segment: 

Remaining performance obligations (RPO) grew 359% YoY with cloud RPO growing “nearly 500%” on top of 83% growth last year. This compares to RPO growth of 41% YoY last quarter and cloud RPO growth of 83% last year.  

Oracle Cloud Infrastructure (OCI) was forecast to “grow 77% to $18 billion this fiscal year and then increase to $32 billion, $73 billion, $114 billion and $144 billion over the following 4 years.”  

You can think of this as an acceleration from roughly 50% growth on IaaS in recent quarters to up to 128% growth in future years, specifically from the $32B to $73B in the medium-term of two years out. 

OCI (IaaS) revenue grew 55% YoY to $3.3 billion, faster than hyperscaler peers.  

This led to multi-cloud database revenue with Amazon, Google, and Microsoft surging 1,529% YoY. 

Earnings: 

GAAP EPS of $1.01, down (15%) QoQ from $1.19 in Q4’25 and flat YoY vs. $1.03 in Q1’FY25. This figure was also lower than the analyst estimates of $1.04. 

Non-GAAP EPS of $1.37, up 6% YoY from $1.39 in Q1’25 but down (14%) QoQ from $1.70 in Q4’25. 

Margins: 

Q1 figures represent a 29% operating margin, down from 32% in the prior quarter and 30% in the prior year quarter. 

Cash: 

(39%) FCF and 55% OCF. Highly leveraged cash to debt ratio 

Valuation: 

Oracle’s valuation is at the top range of where the stock has traded this year at 12 forward PS. The stock briefly touched 13 before selling off, and otherwise, has not traded higher than 12 this year (or for decades really, but is transforming into an AI stock) 

The forward PE Ratio of 42 is the highest it’s ever traded this year or in previous years.  

Risks: 

Highly leveraged cash to debt ratio is the predominant risk.  

#4: AppLovin Has the Best Operating Margins Sector-Wide 

Financials: 10/10
Thematic: 7/10
Valuation: 3/10 

APP has many aspects to focus on for the bull story, yet the margins are truly one-of-a-kind. It is hard to take the title of “most impressive margins” in a software category as AppLovin is up against the most operationally efficient, cash loaded companies worldwide.  

Regarding its segments, management has repeatedly stated that gaming ads alone can sustain growth of 20% to 30% YoY. Therefore, the catalyst for the next few years is securing additional supply, such as e-commerce, as well as opening up the AXON ad platform to more advertisers.   

The AXON ads manager recently became self-service, which means it can scale at levels not previously seen by offering self-service interface for AppLovin’s 1 billion reach. As of now, AppLovin is limited in the number of advertisers it can manually onboard. According to the opening remarks: “With the rollout going smoothly, we are ready to widen access. On October 1, 2025, we plan to open the AXON ads manager on a referral basis, perfectly timed for the holiday season. Feedback from these partners will guide our global public launch in the first half of 2026. To date, web advertising campaigns have been limited to the United States. On October 1, we plan to open our platform to most major international markets.”  

Management also hinted these improvements will lead to “a lot of upside in the numbers we’re able to report.”  

Here is the full quote: 

“We expect that will increase the advertiser count quite quickly and also allow us to go through live examples of advertisers coming in self-service all the way to scale on our product. Assuming all that goes well, then we talked about opening up the platform entirely to the world in first half of next year. We think as advertiser count grows on our business, especially in categories outside of gaming, you're going to see a lot of upside in the numbers that we're able to report”. 

The last earnings report was stellar with all fundamental boxes ticked, a team that has proven to execute, and incoming catalysts that are quite well-timed to where we maybe have two quarters (or less) to wait for an inflection. 

And check out those margins! 

Revenue: 

AppLovin reported revenue of $1.26 billion compared to consensus of $1.28 billion according to some sources yet others reflect the consensus we had of $1.22 billion, thereby it’s debatable if the top line beat. Overall revenue last quarter was $1.48B versus this quarter at $1.26B. As we covered in the past, this is due to AppLovin divesting its mobile gaming “Apps” business, with the sale completed on June 30th. Therefore, if you adjust for this sale, revenue for the ads business in Q1 was $1.15B for QoQ growth of 8.7%. 

AI Segment: 

Same as revenue (AXON ad engine is powered by AI). 

Earnings: 

On the bottom line, the company had a large beat with EPS of $2.39 compared to $1.99 EPS expected, representing growth of 169%. This was a 45 point beat on growth rate for the bottom line. Adjusted EBITDA doubled to $1.02 billion, up from $943 million last quarter. This represents an adjusted EBITDA margin of 81%. 

Margins: 

Operating margin of 76% expanded from 44.7% last quarter and more than doubled from the year ago quarter at 36.2%. Wow! 

Cash: 

As management alluded to on the earnings call, the company “prints cash” with a 61.3% operating cash flow margin and a 61% free cash flow margin. 

Valuation: 

The company is trading at a PS ratio of 40.3 and a forward PS ratio of 37.6. The company’s forward PS ratio peaked at 46.7 on September 30 and is currently trading about 20% below its peak. While the forward PS ratio above 30 is considered high, the market is giving the company a premium valuation due to a remarkable turnaround in margins. 

Another key catalyst is that the company completed the divestment of its low margin mobile gaming business in Q2 2025. The company’s adjusted EBITDA margin has swiftly moved higher from 68% in Q1 to 80.9% in the recent quarter. It would also make more sense to look at the EV/EBITDA ratio for AppLovin. It currently trades at an EV/EBITDA ratio of 63.4 and a forward EV/EBITDA ratio of 49.1. The company has traded at a peak EV/EBITDA ratio of 110 during June 2021 and around 88 during the tech market bubble in November 2021. With profitability improving post-divestiture, there could be further room for valuation expansion.  

Risks: 

APP is the subject of short reports and has a business model the SEC, short sellers and others find suspicious. However, we think the management team, AI-powered ads business model and strong market presence is worth a shot especially when using technicals alongside any entries or exits. 

#5: Cloudflare is Locked and Loaded for AI Inference Market 

The Workers Platform, known as Act 3, is positioned to take advantage of the massive inference trend. The I/O Fund team recently dug up a stat that inference is expected to account for 60% to 70% of AI workloads by 2030. In particular, Cloudflare emphasizes their position is what will help the company win this market: “The fact that we sit in front of so much of the web and that more than half of our dynamic traffic is already between APIs means that we are strategically positioned to deliver the agentic web of the future.” 

Revenue: 

Cloudflare reported its largest revenue beat in the last six quarters at 2.1% above consensus, with Q2 revenue up 27.8% YoY to $512.3 million. This also marked a slight 1.3 point acceleration on the top-line from 26.5% growth in Q1.   

For Q3, Cloudflare guided for revenue of $543.5 to $544.5 million, ahead of estimates at the time for $538.9 million. This corresponds to a slight deceleration to the mid-to-high 26% YoY growth range, where Cloudflare is expected to remain through Q4. This provides no clear indication yet that the company is able to drive a sustained revenue acceleration aided by AI.   

Current RPO accounted for 66% of total RPO, or ~$1.30 billion, increasing 33% YoY in Q2, a four point acceleration from 29% growth in Q1. This is also a notable uplift from 26% growth in the year ago quarter.   

AI Segment: 

There is no official AI segment yet. 

Earnings: 

Cloudflare topped estimates in Q2 driven by the revenue beat and stronger adjusted margins, and boosted its FY25 adjusted EPS outlook as a result.  

GAAP EPS was ($0.15), missing estimates for ($0.08) as GAAP margins drifted lower.  

Adjusted EPS was $0.21, beating estimates for $0.18, fueled the outperformance in adjusted operating margin in the quarter.   

Margins: 

GAAP gross margin was 74.9% in Q2, down nearly 3 points YoY and 1 point QoQ. Adjusted gross margin was 76.3%, down 2.7 points YoY and 0.8 points QoQ.  

GAAP operating margin was (13.1%), down 4.4 points YoY and 2 points QoQ. Adjusted operating margin was 14.1%, approximately flat YoY and up 2.4 points QoQ; this was also ahead of guidance for 12.6%.  

For Q3, Cloudflare guided for adjusted operating income of $75-76 million, pointing to adjusted operating margin of 13.9%, down nearly 1 point YoY and moderating slightly QoQ. 

Cash: 

Operating cash flow was $99.8 million for a 19% margin, flat YoY but down from a 30% margin in Q1.   

Free cash flow was $33.3 million for a 6% margin, down 4 points YoY and 5 points QoQ.   

Network capex was 11% of revenue in Q2, down from 17% of revenue in Q2. Cloudflare stuck to its guidance for network capex to be 12-13% of revenue for the year, suggesting slight moderation in 2H.   

Valuation: 

Cloudflare is officially trading at its highest forward since the tech bubble popped in late 2021-early 2022. The current PS of 40 and forward PS of 36 does not offer much support in terms of the stock holding well at these levels. 

Cloudflare is trading at a wild forward PE ratio of 255, which is far above any forward PE ratio over the past year (the previous highest forward earnings ratio was 205 where it sharply reversed twice). 

Risks: 

Valuation is the predominant risk coupled with little evidence of real AI revenue right now (more of a future winner that we want a placeholder for).

AI Energy 

#1: Right Place, Right Time for Bloom Energy 

Bloom Energy provides on-site 24/7 power generation using their proprietary solid oxide fuel cells (SOFCs). The SOFCs are stacked up by the hundreds to thousands in Bloom Energy Servers (BES), which enable the conversion of fuels like natural gas, biogas and hydrogen to electricity. 

Bloom Energy is securing data center deals due to fast deployment of about three months. Here is what management described as the competitive advantages regarding time to power for 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. Such rapid drill activities will necessarily come with timeline variances, both pull-ins and delays, and will affect our quarterly revenue line. You are seeing this in our Q3 numbers.” 

Although we expect Bloom to be a very volatile stock, the fact is that very few alternative energy companies can move as quickly as BE in what our firm has dubbed an energy crisis in getting power to data centers.   

As the CEO stated on the call, to wait 5-7 years is “untenable.”  To compare, Bloom will power Oracle with on-site power solutions in as soon as 90 days. Additional key customers for BE include American Electric Power (AEP), Quanta and Equinix. Notably, Amazon and Cologix are customers of Bloom through AEP in Ohio.   

Perhaps the most important statement on the earnings call was when the CEO stated: “We expect new orders from other AI hardware ecosystem players soon, complementing demand we see from our more traditional commercial and industrial customers.” 

Revenue: 

Bloom reported a nearly 6% beat to estimates in Q2, reporting $401.2 million in revenue versus estimates for $378.9 million. Revenue grew 19.5% YoY, slowing from 38.6% growth in Q1. 

AI Segment: 

Product revenue increased 31.1% YoY to $296.6 million, slowing from 38% growth in Q1. As the CEO stated on the call, to wait 5-7 years is “untenable.” To compare, Bloom will power Oracle with on-site power solutions in as soon as 90 days. 

Earnings: 

Adjusted EPS of $0.10 beat estimates for $0.02, and represented a notable $0.16 improvement YoY.  

GAAP EPS was ($0.18), missing estimates for ($0.10) as GAAP net margin declined sequentially. 

Margins: 

GAAP operating margin is approaching positive territory at (0.9%) in Q2, up nearly 5 points QoQ and 6 points YoY.  

Adjusted operating margin was 7.1%, up more than 3 points QoQ and 8 points YoY. Adjusted EBITDA was $41.2 million 

Cash: 

Free cash flow was ($220.4 million) in Q2 for a (54.9%) margin. For the first half, FCF was ($345.3 million), just over a 1% improvement YoY.  

Unrestricted cash and equivalents totaled $574.8 million, down from $794.8 million in Q1. This raises the risk that Bloom will turn to financing methods as Bloom likely awaits cash flows meaningfully improving in Q4. 

Valuation: 

If you want to see what AI can do for a valuation, look no further than BE. The stock used to trade in the range of 1.5 forward PS for many years and is now trading at a 14.5 forward PS. I can’t offer much in terms of where the valuation trends as it’s far beyond anything BE has traded at historically.  

However, the forward PE ratio offers a bit more data as the forward PE of 219 is an area the stock has traded at consistently over the past year. With that said, the stock has traded as low as 33 forward PE, as well. 

Risks: 

The valuation is a risk yet we are less concerned with what BE does as an individual stock and would see any selloff being more of a broader catalyst. 

#2: GEV: All Roads Lead to GEV 

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 in the second 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 62 GW in backlog for gas equipment contracts, already surpassing expectations of reaching 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.   

This year, the company is expected to report $37 billion in revenue with strong earnings growth of 45.3%. The stock is not a hypergrowth profile, rather, it is a quality, defensible position that could do well during any periods of doubt in the broader AI trend.  

The defensibility is particularly attractive when you consider that gas turbines is the crux of the issue for expanding gas power plants. According to Bloomberg’s calculations, more than “$400 billion worth of gas-fired power plants through the end of the decade are in jeopardy of delay or cancellation because of the lack of capacity to meet future turbine orders.” The same article points toward GE Vernova filling orders as far out as 2030. 

Revenue: 

The company’s Q3 revenue grew by 12% to $9.97 billion, beating estimates by 8.8%. Organically, revenue grew by 10% YoY to $9.83 billion, driven by strong electrification and power. Analysts expect revenue to decline (2.3%) YoY in Q4 but rise 3% QoQ, before rebounding to 8.7% in Q1 2026.  

On the back of strong demand for power and equipment, management reiterated full-year revenue guidance and expects 2025 organic revenue to come at the high end of the guidance of $36 billion to $37 billion. The power segment organic revenue guide was maintained at 6-7%, and electrification segment was raised to ~25%, up from 20% previously in Q2. On a side note, the wind segment is expected to be down high-single digits, lowered from down mid-single digits due to the more challenging market conditions.  

Revenue growth is set to accelerate over the next three years. Analysts expect a 6.6% increase in 2025, bringing the total revenue to $37.2 billion. Momentum is projected to build further, with revenue climbing to $41.0 billion in 2026, up 10.2% and to $46.7 billion in 2027, up 13.9% YoY. 

AI Segment: 

In Q3, GEV signed just over 12GW of new gas equipment contracts with ~1GW going directly to orders and ~12GW going into what’s called a slot reservation agreement. During the quarter, the company also converted 7GW of SRAs into orders and shipped 4GW of equipment.  

Management had previously stated they would exit the year with 60GW “at better margins with significant momentum into ‘26.” Here is the breakdown from that comment:  

33GW are in the backlog, up 4GW sequentially.   

Slot Reservation Agreements (SRA) grew from 25GW to 29GW.  

Total backlog including SRAs is 62GW, up from 55GW last quarter. 

There was a discussion in Q2’s call that this represents about 3 years of backlog: “And we've talked about the fact that we'll get to at least 60 gigawatts by the end of the year. So that's directionally 3 years of backlog.” There was mention of eventually seeing 80GW to 100GW in the backlog but no date or other details were discussed, other than that’s the goal over time.   

Earnings: 

Q3 GAAP EPS came at $1.64, missing estimates by a (18.6%) but up from ($0.35) in the year-ago quarter. Despite the miss in Q3, earnings growth is projected to be strong over the next few quarters, with GAAP EPS up 87.3% to $3.24 in Q4 and accelerating to 128.6% to $2.08 in Q1 2026. 

Analysts continue to expect strong EPS growth in the coming years. For 2025, analysts expect GAAP EPS to grow 40.9% YoY to $7.86, and 62.8% and 44.8% YoY in the subsequent two years, reaching $18.53 in 2027. 

Margins: 

Q3 adjusted EBITDA grew by 234% YoY to $811 million, driven by strong growth in the electrification and power segments, and a strong rebound in wind. Adjusted EBITDA margin improved 540 basis points YoY to 8.1% driven by profitable volume, better pricing, and productivity gains. The company is witnessing an annual EBITDA margin expansion, increasing from 2.4% in 2023 to 5.8% in 2024, and management has further guided expansion in the range of 8-9% for 2025.  

Q3 operating margin was 3.7%, down slightly from 4.2% in Q2 but up from (4.0%) in the year ago quarter.  

Cash: 

Most importantly, management has maintained its FY free cash flow guidance from Q2, where it was raised from a range of $2.0 billion to $2.5 billion to a new range of $3.0 billion to $3.5 billion. This was primarily driven by a higher profit outlook and increased down payments due to rising orders. Through Q3, the company has generated $1.9 billion in free cash flow, implying a strong Q4 to finish the year inside its guided range. 

The company generated free cash flows of $732 million in Q3 compared to $968 million in the same quarter last year. On a sequential basis, this was a sharp increase from $194 million in Q2. 

Valuation: 

Similar to other names in this group, GEV’s valuation has reset higher as the company demonstrates clear AI product–market fit. The stock is trading at 4.2 forward PS despite trading as low as 1 earlier this year.  

The forward PE ratio offers room in the valuation. The stock is trading at 83 forward PE ratio yet has traded as high as 130 in the past. 

Risks: 

Of the companies featured here, GEV offers less risk as the sheer GWs it can provide are desperately needed for AI data center buildouts. However, it’s not a hypergrowth stock like the others. It’s included here to say – we are eyeing the stock as one that could outpace the legacy FAANGs for example, as it’s a leader within one of our largest and most timely thematic trends. 

#3 Bitcoin Miners 

Bitcoin Miners offer an exceptional risk/reward as these companies are pivoting from unprofitable Bitcoin mining operations to the high margin business of supplying powered shells for AI data centers. According to CoreWeave, powered shells are the biggest constraint in the AI build out, marking a critical shift to where Nvidia’s GPUs are no longer the primary constraint. With existing power, cooling, and network infrastructure, miners can deploy AI-ready capacity faster than new construction—offering Big Tech a critical shortcut in the race to scale. 

Our Discovery Tier highlights a Bitcoin miner that is expected to report 325% YoY growth in the upcoming quarter with a positive operating margin. Knox sees significant upside and is ranking this Bitcoin miner as his #1 among all Bitcoin Miners.  

Learn more about Discovery here. 

Conclusion: 

The best way to train for a marathon is to run a marathon. You can consider this report a training exercise while the real test is portfolio returns. Hopefully after reading this 43-page report, a few things are evident – we have a strong grasp of where the AI market plans to go next, that our process is nimble enough to capture winners in niche micro-trends, and we are capable of offering a level of conviction rare among AI investors. Consider that Nvidia is not one of our top performers this year (so far) – yet it’s been one of our strongest years to date. 

With one quarter left, we look forward to making it our best of 2025. 

Want to know what the I/O Fund is eyeing next for a new entry? Our Discovery tier surfaces new ideas in an effort to provide a significant edge to AI investors. Last week, we published our Top 10 New Ideas List for our Discovery members that pinpoints the stocks we are most likely to add to our portfolio next. Discovery was first to discover Bloom Energy, Core Scientific and Oklo to our portfolio, and these new additions became some of our biggest wins (thus far) in 2025. Current Pro and Advanced Members: To subscribe to Discovery with 30% off, click here to email us or email premium@io-fund.com and mention code DISCOVERY30.click here to email us or email premium@io-fund.com and mention code DISCOVERY30. 

Advanced Members – stay tuned! This week, you will be receiving technical setups from Knox in his Quarterly Positions Report offering a complete picture of how we plan to enter or layer into the stocks listed above. These setups will also be covered in Knox’s upcoming webinar this Thursday at 4:30 PM Eastern. If you’re a Pro Member wanting information on Advanced or Discovery tiers, please email us at premium@io-fund.compremium@io-fund.com

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

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 I/O Fund’s Top 15 Stocks for Q3 2025
• GE Vernova: All Roads Point to the Nat Gas Behemoth
• Why Power is Critical for Data Centers and their Hyperscaler Customers
• Free Bitcoin & Broad Market Webinar Replay – August 21, 2025