SIAgentNews.com
← All news

Anthropic uses AI to port Claude to AMD chips in three days

The AI company claims it used its own model to automate the software migration to AMD hardware, potentially breaking Nvidia's software moat.

· First reported by The Decoder

100 percentof kernels ported by the Mythos model

The short version

  • Anthropic claims the porting process took three days instead of the usual six to twelve months.
  • The company used its own Frontier model, Mythos, to write almost 100% of the necessary kernels.
  • Anthropic now treats all four major chip families as a single, interchangeable compute pool.
  • The strategy aims to optimize for price-performance by using different chips for training versus inference.

Why it matters

If true, this suggests that the software complexity of Nvidia's CUDA platform is no longer an insurmountable barrier. It could allow AI labs to switch hardware more freely to find the best prices and availability.

Breaking the CUDA moat

Nvidia has long maintained a dominant position in the AI market because of CUDA. This is a software ecosystem that developers use to run complex calculations on Nvidia's specific hardware architecture.

Because every chip type requires its own custom software stack, moving a large model like Claude to a different brand of hardware is usually a massive undertaking. Most companies avoid this because the manual engineering work is often too expensive and time-consuming to justify.

Anthropic claims it has found a way to bypass this hurdle using AI itself. As reported by The Decoder, the company used its own Frontier model, Mythos, to handle the migration to AMD GPUs with almost no human intervention.

From months to days

Traditionally, integrating a new chip platform into an AI workflow involves dozens of engineers working for several months. Anthropic reports that their new automated approach completed the task in just three days.

The Mythos model was responsible for writing or porting nearly 100 percent of the kernels. Kernels are the specific sets of instructions that allow software to communicate with the hardware's processing units.

By automating this process, Anthropic is attempting to turn hardware into a commodity. They want to treat all chips from different providers and architectures as a single, interchangeable pool of computing power.

A multi-chip strategy

Anthropic's goal is to move away from a single-supplier dependency to find the best price-performance balance. They believe that different chips may be better suited for specific tasks, such as training a model versus running its inference.

This strategy is particularly important because Nvidia currently commands high margins of over 70 percent due to its market dominance. By diversifying, Anthropic hopes to ensure that a dollar of compute goes further than it would at any other frontier laboratory.

To support this, the company is also building its own data center capacities through a program called Anthropic Direct. Instead of just buying cloud services, they are leasing chips and operating them in custom-designed facilities to lower total costs.

Recursive improvements

Beyond just switching chips, Anthropic is focusing on what they call compute multipliers. These are research breakthroughs that allow the same amount of physical hardware to produce more performance than before.

The company views these breakthroughs as stacking, where each new discovery enhances the effectiveness of the previous ones. They are working toward a future where Claude might be able to discover these multipliers on its own.

This concept of recursive self-improvement is a major topic in AI research. While it has not been proven yet, Anthropic is positioning its technology to eventually optimize its own underlying infrastructure and efficiency.

Strategic partnerships

The relationship between these companies is becoming increasingly complex and financially intertwined. Nvidia has committed to investing up to $10 billion in Anthropic as part of a larger deal with Microsoft.

At the same time, Anthropic is making moves to secure hardware from other sources. They have announced a strategic involvement with AMD that includes plans to use up to 2 gigawatts of AMD Instinct MI455X GPUs starting in 2027.

These moves show that while Nvidia remains a key partner, Anthropic is actively building a path toward hardware independence. They are using AI to solve the very software problems that previously kept them tied to a single provider.

Unverified investor claims

It is important to note that these specific claims regarding the speed of the porting process come directly from an investor presentation by Anthropic. Because these details have not been independently verified, the three-day timeline for migrating to AMD hardware remains an internal company report rather than a publicly audited fact.

The presentation highlights how the Mythos model was able to write or port nearly 100 percent of the required kernels. This suggests that the company is positioning its own frontier models as the primary tools for overcoming the technical barriers that usually require months of human engineering.

Complex financial ties

The financial landscape surrounding these hardware shifts is also becoming more intricate. Anthropic is involved in a conditional private placement of 26 billion dollars that was filed with the SEC in a November 2026 registration statement. This deal involves seven technology companies taking shares at the IPO price.

These large-scale financial arrangements show how deeply the major AI labs and chip manufacturers are becoming interconnected. As Anthropic seeks to diversify its hardware, it navigates a web of investments and strategic partnerships that involve some of the largest players in the technology sector.

What to watch

  • Whether other AI labs can replicate Anthropic's success in using models to port kernels.
  • The actual performance and cost savings of the multi-chip strategy compared to pure Nvidia setups.
  • The progress of Anthropic's 'Anthropic Direct' data center initiative.

Sources and links

First reported by The Decoder. 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.

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.

More AI news