Anthropic Just Delivered Great News for AMD Investors and Bad News for Nvidia

Nvidia (NASDAQ:NVDA) created one of the widest moats in the semiconductor space for its graphics processing units (GPUs) through its CUDA software platform. However, it looks like Anthropic has built a bridge across this moat, which should be a huge benefit to Nvidia rival, Advanced Micro Devices (NASDAQ:AMD).

Nvidia's CUDA moat could be shrinking

Nvidia's GPUs have become the main chips used to power artificial intelligence (AI) workloads, and its GPUs particularly dominate in the AI model training market.

This dominance came in part because Nvidia had the foresight to develop the CUDA software platform, which makes it easy for developers to program its GPUs for tasks beyond their original purpose of speeding up graphics rendering in video games. While other markets for GPUs were slow to develop, the company smartly seeded CUDA into institutions that were doing early work on AI, such as universities and research labs. The result was that an entire generation of developers was trained on its software, and most original AI code was written on CUDA and optimized for its chips.

In the years since, Nvidia has expanded on its CUDA moat with CUDA-X, an immense ecosystem of AI libraries and tools to help optimize, deploy, and scale high-performance computing workloads. It has also given university researchers free GPUs and training to make sure the next generation of developers are also accustomed to using its software platform.

However, with one slide in its latest investor deck ahead of its initial public offering, Anthropic has potentially weakened Nvidia's CUDA moat significantly. The AI lab said its Claude Mythos Preview model took just three days to successfully port and optimize Claude to run on AMD's MI355 GPUs, with Claude porting or writing 100% of the kernels. A kernel is AI code that tells a GPU how to execute a mathematical operation. It said that, under normal circumstances, it previously would have taken a dozen engineers between six and 12 months to onboard the use of new AI accelerators.

The goal for Anthropic was to create a "fungible compute pool," where basically it could bring any AI chip up to speed within days to run its models, whether GPUs from Nvidia or AMD, or custom silicon like Amazon's (NASDAQ:AMZN) Trainium chips or Alphabet's (NASDAQ:GOOGL) (NASDAQ:GOOG) Tensor Processing Units. This allows it to buy chips based on the best price performance regardless of vendor and to use the chips for what they are best at, like AI model training or inference.

This is good news for AMD and comes at a great time, with its new Helios system rolling out and Nvidia suffering reputational damage following its now-former CEO's dalliance with an alleged Chinese spy. One of Nvidia's biggest advantages has always been that software written on CUDA did not run on competitors’ chips and was difficult to modify for them.

If Claude Mythos can greatly reduce the time, cost, and complexity of allowing AI models to run on alternative chips despite their having different software and inference stacks, this meaningfully levels the playing field.

Image source: Getty Images.

AMD has a lot to gain

Where AMD has the advantage over Nvidia, meanwhile, is cost. Nvidia's Vera Rubin platform can cost around $40 billion per gigawatt, while AMD's Helios system sells for roughly $20 billion per gigawatt. Even if Helios' overall performance lags that of the Vera Rubin platform, AMD could take a significant share based on the price-to-performance calculation. This is a huge market, so AMD has a lot to gain.

At the same time, Nvidia has a lot to lose. If the company's CUDA moat significantly shrinks and it starts losing share, this could force it to lower prices, which would impact sales, gross margins, and profitability. Now, I wouldn't write Nvidia off yet, as it still has the premier AI infrastructure solution and the best overall ecosystem, and has locked up a lot of supply-constrained high-bandwidth memory, which all these AI chips still need to be packaged with to optimize their performance.

Nonetheless, AMD is the top AI semiconductor stock to buy on this development, as it wouldn’t need to take that large an additional share of this huge market to achieve significant sales growth.

Geoffrey Seiler has positions in Advanced Micro Devices, Alphabet, and Amazon. The Motley Fool has positions in and recommends Advanced Micro Devices, Alphabet, Amazon, and Nvidia. The Motley Fool has a disclosure policy.