During the Anthropic IPO roadshow, there was a case that made many people take a look at Nvidia’s stock price.
According to Anthropic, they used Claude Mythos Preview to help adapt AMD's AI accelerator. It only took 3 days for the internal model to complete prompt word processing on AMD MI355X, and the performance reached usable levels. Almost all the GPU kernel code in this work was ported or written by Claude, with very little manual intervention.
For comparison, Anthropic said that based on past experience, using traditional engineering methods to integrate a new chip platform usually requires dozens of engineers to work for several months, and it actually takes 6 to 12 months.
Why is this matter sensitive? NVIDIA's deepest moat has never been just the chip, but CUDA: a set of development tools, computing libraries and GPU-optimized software for it. Once the code is written for CUDA, if you want to move it to another chip, you need to rewrite a lot of underlying kernels and re-tune it for matrix operations, memory access, and multi-card communication. This migration cost is the real reason why many companies dare not change their cards.
If AI can reduce this matter from half a year to a few days, the threshold for changing chips will be greatly reduced. The actual chips used by Anthropic now include Nvidia, AMD, Google TPU and Amazon Trainium. It and AMD also previously announced that they will deploy up to 2GW of MI450 series computing power, with the first batch of 1GW planned to begin in the first half of 2027.
Looking at it calmly, there is still a gap between "reaching usable performance" and "running to optimal performance", and 3 days is just Anthropic's unilateral statement. But the direction is clear: after writing the underlying code itself is accelerated by AI, the lock-in effect of the hardware ecosystem will become weaker and weaker.