Anthropic uses AI to port Claude to AMD hardware in three days
The AI company used its own model to automate complex software migrations, potentially breaking Nvidia's software moat.
The short version
- Anthropic used its 'Mythos' model to write nearly 100% of the required kernels for AMD GPUs.
- The migration took three days, compared to the traditional six to twelve months required for new chip platforms.
- Anthropic now runs Claude on a pool of hardware including AWS Trainium, Google TPUs, Nvidia, and AMD.
- The move could allow companies to choose hardware based on price and availability rather than software compatibility.
Why it matters
By using AI to automate the hardest part of hardware integration, Anthropic is making it easier to switch away from Nvidia. This could lower costs and increase competition in the AI chip market.
Automating the migration
Anthropic recently revealed that it successfully ported its Claude AI model to run on AMD graphics processing units in only three days. This achievement was made possible by the company's own frontier model, known as Mythos. According to an investor presentation, Mythos wrote nearly 100 percent of the necessary kernels for the move with very little human intervention.
Kernels are specialized pieces of software that tell the hardware how to perform specific mathematical operations. Traditionally, moving a large AI model to a new type of chip is a massive engineering undertaking. It requires rewriting these complex instructions to ensure the software can communicate correctly with the new hardware architecture.
Before this AI-driven approach, the process of getting a model running on a new chip platform typically took between six and twelve months. This manual work usually required the involvement of dozens of engineers. By using Mythos to do the heavy lifting, Anthropic has drastically reduced the time and human labor needed for hardware compatibility.
Breaking the software moat
Nvidia has long maintained a dominant position in the AI market due to its CUDA software platform. CUDA is the standard software layer that developers use to run AI workloads on Nvidia chips. Because it is so deeply integrated, many companies stay with Nvidia even if other chips are cheaper or more available because moving away is so difficult.
The ability of Anthropic to quickly port models to different hardware suggests that the difficulty of switching chips is shrinking. If AI can be used to automate the transition between different chip architectures, the software advantage held by Nvidia becomes less of a barrier. This could lead to a market where companies choose hardware based on price, availability, or specific performance needs.
As it-daily.net reported, this shift could have significant implications for the industry's economics. Nvidia currently sees gross margins of more than 70 percent. If customers can more easily switch to competitors like AMD, it could put pressure on those high margins and encourage more diverse hardware spending.
A multi-chip strategy
Anthropic is currently operating Claude on a diverse pool of hardware that includes AWS Trainium, Google TPUs, and GPUs from both Nvidia and AMD. Tom Brown, the company's Chief Compute Officer, described this as an unconventional bet on a multi-chip strategy. The company treats the total computing capacity from these different providers as a single, interchangeable pool.
This strategy allows Anthropic to distribute its research, model development, and customer demands across various architectures and locations. They also utilize what they call Compute Multipliers, which are research findings that allow them to extract more performance out of the same hardware. This approach helps them manage the massive scale required to run modern AI models.
The company also manages its own infrastructure through a program called Anthropic Direct. This allows them to rent chips and operate them in their own data centers. By not relying on a single hardware provider, Anthropic gains more flexibility in how they scale their operations and meet user demand.
Financial and hardware ties
Both major GPU manufacturers have deep financial ties with Anthropic. Nvidia has committed to investments of up to $10 billion by November 2025 and is participating in a conditional private placement worth $26 billion. These investments highlight the importance of Anthropic as a primary customer for high-end AI hardware.
On the other side, AMD has also secured a significant relationship with the company. AMD announced an investment of up to $5 billion in July. In exchange, Anthropic plans to use up to two gigawatts of AMD Instinct MI455X GPUs in Helios rack systems, with the first phase beginning in the first half of 2027.
Market research from Forrester suggests that this trend toward hardware diversity may continue. The firm expects that by 2027, a quarter of AI infrastructure accelerator tenders will specifically request hardware from manufacturers other than Nvidia. Anthropic's success in quickly porting Claude to AMD could be a key driver in this transition.
What to watch
- Whether other AI companies can replicate Anthropic's method of using models to write hardware kernels.
- The actual performance of Claude on AMD MI355X GPUs compared to Nvidia's offerings.
- The percentage of AI infrastructure tenders that move away from Nvidia by 2027.
Sources and links
- Reportingit-daily.net: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by it-daily.net · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
