Hyperscalers vs AI Labs Debate, and the Moving AI Frontier

I thought I would cover Microsoft’s earnings today, but let me first respond to a tweet by SouthernValue and then I will share some thoughts on the state of AI models. I recommend you read his pushback to my yesterday’s piece before you read the rest.

I will start with a clarification. When I said Google Cloud’s growth is more "broad based" than AWS, it was only an observation of 2Q’26 and not a prospective one. What the mix looks like going forward is a different question, and given GCP's smaller base plus upfront recognition on TPU hardware sales, “Southern” may well be right that Anthropic becomes a larger percentage of GCP's revenue than AWS's over the next year. But even if that proves to be true, I'd argue the risk remains more asymmetric for AWS. If lab revenue disappoints, Google can slow capex and absorb capacity into its massive 1P workloads i.e. Search, YouTube, Gemini, DeepMind can all soak up TPUs. AWS has no comparable internal sink; its infrastructure is built to be rented. Higher revenue exposure with a shock absorber can be a better position than lower exposure without one.

On the question of labs being still rather insignificant part of overall cloud revenue, I think a snapshot of today's mix understates where this is heading. Every hyperscaler is reporting hundreds of billions of backlog and that backlog is indeed tomorrow’s revenue. So even if labs are ~10% of AWS run-rate today, the composition of what's contracted to convert into revenue over the next few years is far more lab-heavy than the current mix. Ultimately, a lot hinges on what AI labs’ revenue will be in 3-4 years. If its hundreds of billions, the incremental revenue will have much higher concentration from AI labs. Sothern himself mentioned “Trn/GPU clusters serving labs are basically the lowest rev per GW part of MSFT/AWS business, being long-term wholesale contracts. Clouds charge much higher prices selling to the enterprise (where they also cross sell high margin cloud services), and serving inference tokens.” Indeed, that’s also my concern. The incremental revenue from the capacity coming online in the next few years will be inferior to current and traditional cloud economics. And every year the concentration increases, the labs bargaining power over hyperscalers should increase. There are multiple moving pieces here though. If compute demand remains comfortably higher than supply for years, this increasing bargaining power simply may not materialize in hyperscalers’ financials anytime soon since AI labs will be more busy focusing to get the capacity they want/need. As an investor who cares about moats which plays a critical role in what multiple we should be paying for the earnings streams, I am still alarmed by the evolution of AI labs bargaining power here even if they do not necessarily materialize in the next 2-3 years in current compute constrained environment.

Southern also says “Ultimately the labs will become profitable and want to build their own GPU training and inference infrastructure, so I don’t expect them to get better terms from clouds, just to build their own infra in time”. I’m not sure I understand how that is actually bullish for hyperscalers in the long-term. The possibility of insourcing IS labs’ bargaining power. Once a lab matures to build its own capacity, it will know its all-in self-build cost per token with precision. That number then becomes the ceiling on what it will pay any landlord, plus perhaps a convenience premium for speed, and flexibility. The landlord's margin will become that convenience premium and the premium should shrink as the lab matures, and capital markets get comfortable financing compute against lab offtakes directly. The sweetener of hyperscalers’ circular deals will no longer be relevant and labs will act like a “full adult” in the negotiating table. Admittedly, there are some real barriers for such in-sourcing dreams even if labs are fully capable of financing and operating such data centers. There are potentially multiple bottlenecks in the AI value chain. Power could be a real constraint for transitioning away from 3P hyperscalers to 1P insourcing. So the “convenience premium” could end up being more attractive than it sounds if the incumbent hyperscalers simply have all the pieces of the puzzle. Moreover, we may very well see more concerted political backlash for more and more data centers as the buildout continues.

As you can see, I am much more ambivalent than offering any certainty how this long-term potential economic tussle between hyperscalers and frontier labs may settle in 5-10 years. Given the ambivalence in my mind, especially in current multiples for hyperscalers, I think it makes sense for me to largely be observant for a few quarters or years. I still do have some exposure to this debate via GOOGL, so I’m not quite entirely a distant observer yet.

Frankly speaking, if you listen hyperscaler CEOs closely during this earnings call, they are sort of telling you they are not as nonchalant about labs increasing bargaining power. For example, during the prepared remarks, Satya Nadella started his commentary by mentioning two of the goals Microsoft has. One of those goals is “empowering every organization to build their own continuous learning loop and ensuring that they don’t outsource their core IP.” He later explained during Q&A:

“…at the end of the day, every firm is going to evaluate who are the providers who are helping them with their outcomes and their knowledge creation. I think that, that is now fairly clear, and it’s going to become clearer by the day. This is not going to be about come in and take all my knowledge and benefit yourself, whereas I am not getting anything out of it.

So given that direction of travel, we are very, very clear about the architectural sort of design of the platform, which is you’ve got to keep your harness separate from the model. When the harness will ensure that your memory, your context, all of that is external. That means any given model at any given time is swappable. You should and you can use frontier models. There’s no reason not to. But you also can use multiple of them, right? So if you look at some of the stats I gave. It’s a great example of how to use the frontier models for what they deliver, how to use low-cost models for what they deliver, and in fact, train your own model when you don’t want to use any external model itself because after all, you have all the output, you have all the traces, you have all the context. That’s really the enterprise design architecture that we are going to evangelize.”

I am confident that Microsoft will certainly use “swappable” models in their 1P software products/solutions, but not nearly as confident that enterprise customers are nimble enough to follow Microsoft’s footsteps.

In my last Microsoft earnings update, I mentioned while Microsoft now evangelizes a multi-model world, their own disclosures suggested ~90% of Foundry customers were still using models from just one company. This quarter, the disclosure goalposts moved again. Microsoft now says Foundry has 100k customers (up from the 80k disclosed in FY 1Q’26) with revenue more than doubling YoY, and that customers building with models from multiple providers are up 5x since the start of the year. Notice they did not repeat the “used more than one model” count from last quarter, nor the OpenAI+ Anthropic customer count from FY 2Q’26. Each quarter we get a new numerator on a new base, which makes penetration very hard to track; my best triangulation is that the large majority of Foundry customers are still effectively single-model, even if the multi-model cohort is growing quickly off a small base. None of the disclosures (or omission of certain data) in these earnings call is accidental. If it were conducive to the narrative of multi-model or “swappable” model world Microsoft is trying to propagate, we would have consistent data about it. Maybe it’s just too early though; the reality is Anthropic and OpenAI have been on a different gear in 2026 than the rest of the pack. The fact that Gemini and other closed models are increasingly missing from the frontier conversation is a huge impediment to the future Nadella and Jassy hope to see. Of course, Chinese models made material progress, and the idea that Chinese open weight models are closing the gap only got momentum recently. So we may get more adoption of multi-models in a couple of quarters. However, OpenAI has already made it clear that they are very much willing to price aggressively to compete vigorously against open weight models. Ultimately, it seems the best way to hurt frontier models is to actually be at the frontier and make the frontier more crowded space than it is today.

Unfortunately, the frontier keeps moving. Andy Jassy mentioned the models keep leapfrogging each other at the frontier, but in 2026 such leapfrogging has been largely confined to Anthropic and OpenAI. While Jassy was talking about leapfrogging in earnings call, Amazon actually completed their $50 Billion investment on OpenAI yesterday that had some interesting initial conditions:

Indeed, the very moment it appeared Chinese open weight models closed the gap, we may be on the cusp of again further divergence between frontier and open weight models.

OpenAI and Anthropic both seem to have much more capable models internally that may be released soon. Just today, OpenAI mentioned an internal version of Astra (OpenAI’s next major model family), “solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science.” If that didn’t raise your eyebrows, get this: “The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates.” We really are likely entering “an era of proof abundance”.

As impressive as it, I don’t know whether we are on the verge of finding multiple new capabilities unlocking hundreds of billions of revenue for the labs, or the fact that even such complex mathematical problems require $2,000 to solve may mean we are perhaps already in the “intelligence is too cheap to meter” era. So, selling raw intelligence to generate hundreds of billions of revenue every year may prove to be a high bar. I know, I know…”Jevon’s Paradox”!

Let me be very clear. The whole AI space moves so quickly these days that it honestly becomes difficult to have a rigid thought about the future for weeks, let alone months or years. The unfortunate reality for investors is that stocks do get priced for their long-term future and market demands a lot of certainty for the future that I still do not quite share for much of the AI value chain. I am hoping once the labs IPO in the coming months or year, we will have slightly more clarity.

Subscribers get the daily journal and five+ years of Deep Dives, i.e. full-length analyses with financial models on 65+ companies. The daily is just how I think out loud between the Deep Dives!

Current Portfolio:

Please note that these are NOT my recommendation to buy/sell these securities, but just disclosure from my end so that you can assess potential biases that I may have because of my own personal portfolio holdings. Always consider my write-up my personal investing journal and never forget my objectives, risk tolerance, and constraints may have no resemblance to yours.

My current portfolio is disclosed below: