AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Anthropic says AI is eroding Nvidia's CUDA lead, cutting an AMD chip switch from months to three days

Collected Nov 11, 2026

Anthropic has told investors that it now trains and runs its Claude models across all four major chip families and that a port of Claude to AMD GPUs took three days, rather than the six to twelve months it says conventional engineering would need. Co-founder and Chief Compute Officer Tom Brown described the multi-chip strategy as an unconventional bet made when most of the industry thought the effort was not worth it. The claims come from an investor presentation and have not been independently verified.

The four families named are Nvidia GPUs, AMD GPUs, Google TPUs and AWS Trainium. Anthropic describes them as a single interchangeable compute pool, with capacity drawn against research, model development, customer demand and internal use of Claude regardless of vendor, architecture or location. Each platform needs its own custom software and inference stack, which is why the work is expensive and slow, but Brown says Anthropic has invested in it for years and has become very good at it. The company also says it extracts more value per dollar of compute than any other frontier AI lab.

The AMD test is the concrete one. The work was done on AMD MI355X GPUs, and the published figures concern prompt processing of an Anthropic model there. Anthropic says its own frontier model, Mythos, wrote or ported nearly all of the kernels itself, with almost no human intervention. Kernels are the low-level GPU routines that implement operations such as matrix multiplication, attention and normalization; in practice, porting a model between vendors means rewriting these for a new instruction set and memory hierarchy. That is the drudgery at the base of the CUDA moat, and it is the work Anthropic says Mythos absorbed.

Background for readers who do not live in this stack: CUDA is Nvidia's programming environment for general-purpose GPU computing, and it is widely seen as the company's strongest lock-in, because a decade of libraries, tooling and debugging habits sit on top of it. Moving a serious training or inference workload to another architecture has historically meant engineers spending months reproducing numerics, chasing performance gaps and stabilizing drivers. Compilers and translation layers exist, but they rarely close the last stretch of performance, and that last stretch is where the money is.

If the numbers hold, the reduction in software overhead is the point. Anthropic frames a multi-chip approach as a way to pick the best price-performance ratio in a tight supply market rather than being tied to one vendor. Different chips suit different jobs, with some better for training and others for inference, so a fungible pool lets capacity be routed to whatever is cheapest or most available at a given moment. Nvidia's margin of more than 70 percent on AI chips, which the presentation calls unusually high for hardware, is the backdrop: any credible threat to its software lock-in touches that pricing power.

The claims sit inside a widening web of financial ties. In November 2025, as part of a $45 billion deal with Microsoft, Nvidia committed to invest up to $10 billion in Anthropic. According to Anthropic's S-1/A registration statement filed with the SEC on November 10, 2026, Nvidia is also taking part in a conditional private placement totaling $26 billion, with seven technology companies subscribing for shares at the IPO price. AMD announced a strategic investment of up to $5 billion in July 2026; in return Anthropic plans to deploy up to two gigawatts of AMD Instinct MI455X GPUs in Helios rack-scale systems, with the first gigawatt phase starting in the first half of 2027. Read plainly, both major GPU vendors now have money riding on the same customer, which makes vendor-neutrality a little less neutral than the phrase suggests.

Two other strands run alongside the chip strategy. Anthropic refers to compute multipliers, meaning research breakthroughs that extract more useful work from the same physical hardware, which it says compound as they stack. The company suggests Claude may eventually find such multipliers on its own, which is recursive self-improvement, an idea much discussed and not yet demonstrated. Separately, the Anthropic Direct program has the company leasing chips and running them in purpose-built data centers rather than renting compute purely as a cloud service, aiming to lower total cost of ownership.

A few caveats are worth keeping in view. The three-day figure covers getting the model running at a usable performance level; it does not by itself say the port matched Nvidia-class throughput across every workload, and the published numbers focus on prompt processing, not the decode-heavy path that dominates chat latency. Onboarding done largely by a model still needs human review, and the real test is whether such ports stay stable under production load and across future driver and compiler releases. There is also a strategic irony: Anthropic is using one frontier model to erode a moat held by a company that is simultaneously an investor.

Why it matters: developers and platform teams that have priced Nvidia-only clusters into their roadmaps now have a public, if unverified, data point that the porting tax is falling. If AI-assisted kernel work compresses onboarding from many months to days, as Anthropic claims, inference capacity becomes something to shop for across vendors and the lock-in argument weakens; my inference is that the near-term effect is more leverage in supply negotiations than a wholesale shift away from Nvidia, since performance parity and ecosystem tooling still have to be proven by teams without a frontier lab's resources.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Anthropic claims it used AI itself to breach Nvidia's CUDA moat. The company's in-house model Mythos reportedly ported Claude to AMD GPUs in just three days, with almost no human intervention. Anthropic now runs Claude on four chip families and aims to cut costs and reduce vendor dependencies. The article Anthropic says AI is eroding Nvidia's CUDA lead, cutting an AMD chip switch from months to three days appeared first on The Decoder .