Anthropic claims to reduce model porting to AMD to three days
Anthropic states it has cut the porting time of Claude to AMD GPUs from several months to three days, thanks to its Mythos model. The company aims to loosen Nvidia's grip and treat multiple chip families as a single compute pool.
Anthropic claims to have reduced the porting of Claude to AMD GPUs to three days, compared to the usual six to twelve months, by letting its frontier model Mythos write or port almost all kernels. This announcement, drawn from an investor presentation, is significant because it challenges the software lock that has long protected Nvidia, CUDA, and its chips. If confirmed, it could make supplier changes much smoother for major AI labs.
Le vrai sujet, c’est le verrou logiciel
Nvidia’s position is not only due to hardware. Its advantage also comes from CUDA, the software environment that forces developers to adapt their models to its architecture. Porting a model to another chip usually requires a lot of engineering, time, and entire teams. Anthropic claims to have bypassed this cost with a multi-chip strategy, treating GPUs as an interchangeable compute pool.
Tom Brown, co-founder and chief compute officer at Anthropic, states that the company trains and runs Claude on the four major chip families. In this logic, AWS Trainium, Google TPU, Nvidia GPU, and AMD GPU no longer form silos but a single compute pool. The benefit is simple: choose the best price-performance ratio, or the chip available at the right time, without rewriting all the software each time.
Trois jours, si les chiffres tiennent
The AMD case is the most striking. Anthropic says it used AMD GPUs for the first time this year and entrusted Mythos, its frontier model, with almost all the adaptation work. The company talks about nearly 100% of kernels ported or written by the model itself. This does not mean humans disappear from the picture, but that the most tedious layer of porting has been largely automated.
The contrast with traditional methods is clear: six to twelve months to integrate a new accelerator, according to Anthropic, versus about three days with this workflow. The gap is large enough to change how major AI players negotiate their dependence on chip suppliers. One key point remains: these figures come from an investor presentation and have not been independently verified.
Une stratégie pensée pour la rareté du calcul
In a market where compute capacity remains tight, flexibility counts almost as much as raw power. Anthropic also explains working on “compute multipliers,” advances that increase the yield of the same amount of compute. The company claims these gains accumulate, allowing it to produce more with the same infrastructure. This kind of optimization makes cloud architects smile and GPU buyers grimace.
Added to this is the “Anthropic Direct” program, through which the company rents chips and operates its own data centers designed for them. The stated goal is to reduce total cost of ownership, rather than only buying capacity as cloud service. This rise in autonomy comes as financial ties with Nvidia and AMD deepen: Nvidia announced in November 2025 a potential $10 billion investment in Anthropic as part of a $45 billion Microsoft deal, while AMD announced in July 2026 a strategic investment up to $5 billion.
Ce que ce basculement changerait pour le marché
If Anthropic truly manages to industrialize this accelerated porting, the effect goes beyond its own use case. AI publishers and labs could negotiate more freely between Nvidia, AMD, Google, and AWS, instead of enduring the schedule and prices of a single supplier. For Nvidia, the stake is clear: CUDA’s dominance remains strong but no longer appears totally untouchable. For AMD, it is another showcase in the battle to exist against the market leader.
The future depends on two things: the reproducibility of these results and Anthropic’s ability to maintain this flexibility as compute demand rises further. For now, the company simply says it has found a way to turn AI into a hardware porting tool. Compute, this time, no longer just runs models. It helps them change homes.
Key points
- Anthropic claims to have ported Claude to AMD in three days.
- The porting reportedly involved almost 100% automated kernels.
- Tom Brown describes four chip families as a single pool.
- The figures come from an investor presentation, without independent verification.
- AMD and Nvidia announced investments in Anthropic in 2026.
By the numbers
- 3 days — porting to AMD, according to Anthropic, 2026.
- 6 to 12 months — usual delay announced for a new accelerator.
- 100% — nearly all kernels ported or written by Mythos.
- $10 billion — potential Nvidia investment, November 2025.
- $5 billion — strategic investment announced by AMD, July 2026.