AI models could change the dynamics of the chip industry that powers them.
Anthropic has revealed that its Claude Mythos model single-handedly brought Claude up and running on AMD’s chips in just three days, a task that would once have taken dozens of engineers months to complete. The disclosure came during Anthropic’s IPO roadshow, as the company laid out the thinking behind one of its earliest and most unconventional infrastructure bets: training and serving Claude across all four major chip families instead of standardizing on one.
What Anthropic’s Said
An Anthropic executive, speaking in the roadshow video, explained that the company now trains and serves Claude across all four major chip families, something made possible only because it placed an early, unconventional bet on a multi-chip strategy at a time when many in the industry doubted it was worth the effort. Each chip requires its own tailored software and inference stack, an expensive and difficult undertaking, but years of doing this have made Anthropic very good at it.
According to the executive, this strategy carries two key advantages. In a supply-constrained compute market, it lets Anthropic chase the best price-performance regardless of which provider offers it, and it allows the company to match each chip to the job it’s best suited for, since some chips are stronger at training and others at inference. The underlying goal is to treat every chip Anthropic uses, across providers, architectures and locations, as one single, fungible pool of compute.
That same muscle, the executive said, has also made Anthropic unusually fast at bringing up entirely new chip platforms. This year, the company worked with AMD GPUs for the first time. In the past, bringing on a new chip platform took dozens of engineers months of work. This time, Anthropic let Mythos try it instead, and the model got Claude running on AMD in just three days with basically no human intervention, porting or writing almost 100% of the kernels itself.
Separate reporting on the episode has filled in some texture: Claude Mythos was reportedly set loose on AMD’s MI355X GPUs, built on the CDNA 4 architecture, with an Anthropic engineer simply connecting the model to the new hardware and telling it to get going. AMD CEO Lisa Su has reportedly confirmed that the extra engineering support AMD had lined up to assist turned out not to be needed at all.
This AMD milestone sits inside a much larger hardware relationship. Anthropic and AMD had already signed a deal committing Anthropic to up to 2 gigawatts of AMD’s Instinct MI450-series chips in AMD’s Helios rack-scale systems, with AMD agreeing to invest up to $5 billion into Anthropic as deployment milestones are hit. Part of that arrangement includes using Claude itself to help optimize ROCm, AMD’s software stack, which is widely seen as the weaker link compared to Nvidia’s CUDA.
A Pattern Anthropic Has Been Building For Years
The AMD sprint is the latest chapter in a hardware diversification strategy Anthropic has pursued methodically. The company’s compute pool already spans Nvidia GPUs, AWS’s Trainium chips and Google’s TPUs, a relationship that began with a deal for up to a million Google TPUs and has since deepened further. Anthropic has also been quietly building out an in-house chip design team, with the roadshow disclosing that custom silicon is now part of the long-term plan alongside, not instead of, its merchant-chip relationships.
That hiring push has notably pulled talent from rival chip programs, including engineers who previously worked on OpenAI’s custom chip effort with Broadcom.
The Bigger Implication: What Happens If Chips Become Truly Fungible
The idea of treating GPUs, TPUs and custom accelerators as one interchangeable compute pool, rather than committing to a single vendor’s ecosystem, is a significant departure from how most AI labs have operated. If a major model developer can genuinely treat chips from different architectures as fungible, the implications ripple across the entire semiconductor industry.
For chipmakers, fungibility shifts the basis of competition. Instead of locking in customers through years of accumulated software investment, vendors would increasingly have to compete on raw price-performance, availability and power efficiency, metrics that can change quarter to quarter rather than remaining fixed for a decade. That is a much more contestable, and much less comfortable, position for any single supplier to be in.
It also changes the calculus for customers. A buyer no longer has to make a multi-year, all-in bet on one chip architecture, which lowers the risk of adopting a newer or less established platform. That is precisely the dynamic that has helped AMD, and could eventually help other challengers, win business that would once have defaulted to the incumbent almost automatically.
For the industry’s current supply crunch, fungibility is also a practical hedge. When compute is scarce, as Anthropic’s own roadshow disclosures about its enormous compute commitments make clear, being able to redirect workloads to whichever chip happens to be available protects a company from being hostage to a single vendor’s production bottlenecks.
What This Means For Nvidia’s CUDA Moat
The most consequential part of this story is what it implies for Nvidia’s software moat. CUDA has anchored Nvidia’s dominance for nearly two decades precisely because switching away from it has been so costly: code has to be rewritten, engineers retrained and pipelines revalidated, with AMD’s rival ROCm stack historically seen as less mature and less reliable by comparison.
What Anthropic is describing cuts directly at that cost structure. If an AI model can port and rewrite nearly all of the low-level kernels needed to run on a new chip architecture in three days with minimal human oversight, then the single biggest expense behind the CUDA moat, the enormous engineering effort required to migrate off it, starts to shrink dramatically. Anthropic’s own commitment to use Claude to help improve AMD’s ROCm stack only reinforces that trajectory, since it points toward AI-assisted tooling closing the software gap between Nvidia and its challengers faster than organic developer adoption ever could.
That does not mean Nvidia’s position evaporates overnight. Nvidia has long argued, including in response to Google’s TPU momentum, that its GPUs offer broader versatility and fungibility than purpose-built accelerators designed for specific frameworks or functions. Raw performance per chip, networking and system-level integration, and the sheer scale of Nvidia’s existing developer base remain real advantages that a fast AMD bring-up does not immediately erase. Throughput and efficiency on the new hardware, not just the ability to get it running at all, will determine how much workload shifts away from Nvidia in practice.
But the direction of travel is notable. As more frontier labs gain the tooling to port their models across architectures quickly, the premium customers are willing to pay for staying inside Nvidia’s ecosystem is likely to come under growing pressure, forcing Nvidia to compete increasingly on hardware economics rather than the switching costs that have protected it for most of the last two decades.