The AI chip market is undergoing a significant shift that could challenge Nvidia's dominance. Anthropic recently demonstrated that its Claude Mythos model could transition from Nvidia's GPUs to alternative AI chips in just a few days, a process that previously took months and required extensive engineering. This development weakens Nvidia's CUDA moat, especially as competitors like Alphabet and AMD gain ground.
Alphabet's Tensor Processing Units (TPUs) are now outperforming Nvidia's GPUs on a performance-per-dollar basis, according to Semianalysis, though some dispute these findings. AMD's GPUs are also gaining traction, benefiting from larger memory capacity and improved software, making them ideal for inference tasks. These advancements suggest that AI models can be more easily shifted to different AI accelerators, eroding Nvidia's pricing power and margins.
While Nvidia still holds advantages in its ecosystem and supply chain, the commoditization of AI chips could benefit other companies. Taiwan Semiconductor Manufacturing (TSMC) stands to gain from lower AI accelerator prices and increased demand for alternatives. Alphabet, with its decade-long head start in custom chips, could see significant gains as its lower-cost offerings attract external customers. AMD, operating with slimmer margins, is well-positioned to capitalize on the shift towards cost-effective and fungible AI chips.
The market's focus on inference and agentic AI further favors AMD, whose chips excel in memory capacity and cost-performance. With the introduction of its Helios system, AMD has a strong opportunity to establish itself as a viable alternative to Nvidia. As the AI chip landscape evolves, these three companies, TSMC, Alphabet, and AMD, are poised to benefit, while Nvidia faces growing pressure on its market share and pricing.