Read the four biggest compute deals of an AI lab side by side and one of them looks like it doesn't belong. Anthropic has committed more than $100 billion to AWS over ten years for up to five gigawatts of Trainium, and already runs over a million Trainium2 chips under the codename Project Rainier. It has contracted for as many as a million Google TPUs, well over a gigawatt of capacity. It runs Nvidia GPUs on top of both. And then, in a May report from CNBC, there is the fourth thing: early-stage talks to rent Azure servers full of Microsoft's Maia 200 accelerator to serve Claude. No signed deal. No gigawatts. Just a conversation about renting some chips.

The instinct is to file this under "AI lab grabs every GPU it can find in a supply crunch," which is true and boring. The interesting question is why the fourth deal is shaped so differently from the first three — a rental instead of a decade-long reservation, non-exclusive instead of a headline number, still unsigned instead of announced with a press release. That shape is not hesitation. It is the physics of the workload showing through the contract.

Two workloads wearing one word#

We say "compute" as if a lab needs one substance. It needs two, and they behave nothing alike.

Training a frontier model is a single, tightly-coupled job. Tens of thousands of accelerators have to talk to each other at low latency for weeks or months; the interconnect and the cluster's failure rate matter as much as the raw FLOPs; and you cannot casually pick the run up and move it to a different chip because the whole thing is co-designed around the silicon it lives on. If you are going to do that, you reserve the capacity years ahead and you sign for a decade. Hence the shape of the Trainium and TPU deals: enormous, forward-dated, gigawatt-scale, and effectively load-bearing for the roadmap.

Inference is the opposite kind of thing. Serving Claude is millions of independent requests, each mostly stateless, none of which needs to know about the others. That workload is embarrassingly parallel in the literal technical sense — you can spread it across any accelerators that happen to have free cycles, in any data center, from any vendor, and the user never knows which die answered.

Training is a cathedral you commit to building on one foundation. Inference is water — it flows to whichever silicon is cheapest and available this quarter, and you'd be foolish to promise any of it away.

That is why the Maia conversation is a rental. You don't sign a ten-year exclusive for something you can, by its nature, re-route next quarter. You keep it promiscuous on purpose.

The number that makes it worth a phone call#

Microsoft is not offering charity. On its April earnings call, Satya Nadella said Maia 200 delivers "over 30% improved tokens per dollar, compared to the latest silicon in our fleet." Maia 200 is a second-generation, inference-first chip that entered limited preview in early 2026 and still wasn't generally available to Azure customers by mid-year — but "inference-first" and "tokens per dollar" are exactly the axes an inference buyer optimizes, and they are different axes than a training buyer's.

Here is the part that should reframe the whole story for anyone building on top of these models: the reason inference tokens-per-dollar is suddenly the number labs will take a meeting over is agents. A chatbot turn is one model call. An agent task is a loop — plan, call a tool, read the result, re-plan — that can turn a single human request into dozens or hundreds of model calls, and those calls compound as the context grows. The industry's center of gravity moved from training runs you do occasionally to inference you do constantly, at agentic volume. Once inference is the dominant, recurring, price-sensitive cost, arbitraging it across every available accelerator stops being a nice-to-have and becomes the unit-economics game.

What "four chips" means downstream of the API#

Anthropic already frames its own approach, in its own words, as "a diversified approach that efficiently uses three chip platforms — Google's TPUs, Amazon's Trainium, and Nvidia's GPUs." Maia would make four. You can read that as vendor hedging in a supply crunch, and it is. But the deeper reason it's possible to hedge at all is that inference is the portable half — the same trait that lets a lab shop Trainium against TPUs against Maia is the trait that makes the shopping cheap.

For a developer, this is mostly good news arriving quietly. More independent silicon under the same messages endpoint means more supply and more price competition, and you get the benefit without changing a line of code. The catch is subtle and worth internalizing: "the same model" is now a claim about weights, not about the machine that ran them. Trainium, TPUs, and GPUs have genuinely different latency and throughput profiles, and a purpose-built inference ASIC like Maia is tuned for its own sweet spot. If your product has hard tail-latency requirements — a voice agent, an interactive coding loop — don't assume a model name pins a performance profile. Benchmark against the served configuration you actually get, and treat p99 latency as something to verify per endpoint, not inherit from a spec sheet.

The Maia deal may never close; plenty of reported negotiations don't. But the shape of it already tells you where the industry's compute is soft and where it's set in concrete. The training foundations are poured for a decade. The inference layer is water, and every lab is quietly learning to let it run downhill to the cheapest chip that will take it.