Anthropic uses AI to port Claude to AMD in 3 days, cracking Nvidia's CUDA moat

Anthropic's Mythos model ported Claude to AMD chips in just three days, threatening the software lock-in that built Nvidia's empire.

Business · Source: Tikr

What happened

Anthropic just dropped a bomb on Nvidia during its IPO roadshow. Co-founder Tom Brown revealed their frontier model, Mythos, ported Claude to run on AMD chips in exactly three days. The AI wrote or ported almost 100 percent of the necessary kernels with basically no human intervention.

This is a massive leap in speed. Bringing up a new chip platform usually takes dozens of engineers six to twelve months of grueling work. Anthropic tested this on AMD MI355X GPUs for prompt processing. They now treat compute from AWS Trainium, Google TPUs, Nvidia, and AMD as one interchangeable pool.

The timing is brutal for Nvidia. The company just saw CEO Jensen Huang resign after reports of a board ouster. Now, Anthropic is showing the world that Nvidia's biggest competitive advantage might be obsolete. Both chipmakers have deep financial ties to Anthropic. Nvidia committed 10 billion dollars in 2025 and is joining a 26 billion dollar private placement, while AMD invested 5 billion dollars earlier this year.

Key facts

Why it matters

Nvidia built a trillion-dollar empire on CUDA. Developers write code specifically for Nvidia hardware, and moving that code to cheaper chips is usually too expensive and painful to justify. Anthropic just proved you can use AI to automate the translation. If models can port themselves across hardware architectures in days, developers are no longer chained to a single vendor. You can hunt for the best price and availability instead of waiting in line for Nvidia GPUs.

This destroys hardware pricing power. Nvidia maintains a gross margin above 70 percent because customers have no choice but to pay up. If switching costs drop to zero, compute becomes a true commodity. Cloud providers and AI labs will pit AMD against Nvidia for every contract. Compute costs will crash, making it drastically cheaper to train and run models at scale. Anthropic is already doing this. They claim to get more out of every dollar of compute than any other lab by routing workloads to the cheapest available hardware.

For builders

Cheaper compute across multiple chip vendors

You will soon have real leverage to negotiate cloud pricing. As infrastructure providers adopt AI-driven porting, they will pass the AMD and TPU hardware savings down to developers. You can stop paying the Nvidia premium.

Nvidia loses its absolute pricing power

Nvidia currently enjoys massive margins because CUDA traps customers. Builders who rely heavily on compute will benefit as hardware margins compress and chipmakers are forced to compete on raw price-to-performance metrics.

Hardware-agnostic AI deployment is here

Stop worrying about vendor lock-in when building your inference stack. You can architect your systems to route workloads to whatever chip is cheapest or most available that day, just like Anthropic does.

My take

I always knew AI would eventually write its own software, but using it to break Nvidia's monopoly is poetic. Nvidia funded Anthropic, and Anthropic turned around and built the exact tool needed to destroy Nvidia's biggest moat. If I am building an AI product today, I am planning for a future where compute is dirt cheap and completely commoditized.

Original reporting: Tikr. This is my rewrite and opinion.

More AI news for builders