A company’s customers would generally prefer that the company’s product be a commodity. If you are the only person who sells the thing, you set the price. If four people sell roughly the same thing, the customer sets the price, or at least gets a vote. So it is mildly interesting that Anthropic, in the prospectus for its initial public offering, describes its AI chips as something close to a commodity. The AI lab says it manages compute “as a single, fungible pool from multiple strategic partners,” using software that sends work to whichever chips are most productive for it at the moment, as Geoff Weiss reported in Business Insider. Analysts he spoke to think the main loser from this could be Nvidia.
The S-1, filed with the SEC on Monday, lists four chip families: Amazon’s Trainium, Google’s TPUs, Nvidia GPUs and AMD GPUs. Its “multi-chip compute architecture” across all four “allows us to orchestrate workloads to optimize for price and performance,” the filing says. Also: “Our compute pool is fungible by design.” Anthropic says it built its own scheduling and capacity-management systems so that capacity can be moved between jobs regardless of who supplied it, what kind of chip it is or where it sits, and that the approach lets it “opportunistically capture the price-performance benefits of each chip family and generation.”
Translated: when one supplier’s chips get expensive, or scarce, or just less good than the new thing, we move the work somewhere else.
That’s the threat. Patrick Moorhead, founder and CEO of Moor Insights & Strategy, put it like this: “The real exposure could be Nvidia’s pricing power, because a customer that can shift work across four platforms negotiates very differently than one that can’t.” Here is a hypothetical to make that concrete. Imagine you sell Nvidia chips to a customer who has spent years making sure every piece of its software runs only on your hardware. That customer will pay whatever you ask, because the alternative is rewriting everything. Now imagine a customer who can, with some effort, shift a big chunk of its work to Google’s or Amazon’s chips by Thursday. It hasn’t left you. It just gets a better price, because both of you know it could. Ashish Nadkarni, an analyst at IDC, made a related point: running on several kinds of chip also leaves Anthropic less exposed to swings in Nvidia’s supply.
None of this is new, exactly. Anthropic has used a mix of chips for a while; as recently as its Google TPU expansion in late 2025 it described a three-platform strategy, before AMD joined. What’s new, Moorhead said, is that the company is now selling the arrangement to investors as a core advantage. The filing says the flexibility “took years and significant investment to build,” and argues that experience is hard to copy. That is how an IPO prospectus works: you take the thing you were doing out of necessity and call it a moat. Anthropic is asking public investors for a lot of money, and “we have unusual leverage over our biggest cost” is a pretty good thing to be able to say when that cost is the main driver of your losses.
Fungible, with an asterisk
The threat has limits, and the analysts said so. Moving work between chips is still hard, Moorhead said, especially for training models as opposed to running them, and only the very largest AI companies can afford to try. Anthropic also keeps spending billions on Nvidia compute. Gil Luria, head of technology research at D.A. Davidson, said other big labs such as OpenAI likely take a multichip approach too, but that it may not be worth the effort and money for smaller players. Brad Gastwirth, global head of research and market intelligence at Circular Technology, said Anthropic’s setup isn’t a major near-term threat, because having more options doesn’t erase Nvidia’s advantages in performance, networking and software.
The scale of the commitments explains why the leverage question matters. According to CNBC’s reading of the filing, Anthropic has committed $110 billion to Amazon’s Trainium chips through 2036 and $114.5 billion to Google TPUs through 2033, said it could buy more than $20 billion of AMD chips, and has long-term deals with SpaceXAI, Nscale and Lambda for access to Nvidia chips. Once all the promised capacity arrives, Anthropic will pay SpaceX about $3.1 billion a month to “access compute powered by NVIDIA GPUs.” That’s about $37.2 billion a year, for one Nvidia-based supplier. OfficeChai tallied the S-1’s committed compute and infrastructure spending at between $542 billion and $567 billion; other readings come out higher, depending on what you count.
CNBC also noted that all four of Anthropic’s main chip designers hold large stakes in Anthropic or options to buy them. So the negotiation goes: Anthropic plays its suppliers off against one another, and the suppliers, as shareholders, take part of the benefit back through the equity. It’s a cozy arrangement for a negotiation.
The airline problem
The bigger risk for Nvidia isn’t one customer with good software. It’s the possibility that AI compute in general becomes a commodity. Citadel CEO Ken Griffin has warned that if supply overtakes demand, rental prices could fall. “GPUs, like airlines, have a very low marginal cost to run,” he told Bloomberg last week. “The moment you have overcapacity, pricing goes into freefall.” The airline comparison is a good one. Once the plane is flying, an empty seat earns nothing, so any fare beats no fare, and in a glut everyone ends up selling seats at marginal cost. On Monday Michael Burry shared a post about Griffin’s warning and, according to Benzinga, wrote: “This is obvious to most with a working brain cell.”
A fungible compute pool is how a buyer gets ready for that world. If GPU-hours become something you shop around for, the customer with routing software that is indifferent to chip type gets to take whatever is cheapest that day. Nvidia did not immediately respond to Business Insider’s request for comment.
Anthropic, meanwhile, has applied to list on the Nasdaq under the ticker “ANT.” It will soon learn what it’s like to have investors who treat you as one fungible option among many.