Amazon in talks to sell custom AI chips, challenging Nvidia dominance

Bloomberg reports Amazon is in talks to sell its in-house AI chips — its Trainium/Inferentia line — to outside companies, a notable strategic shift from using them purely to power AWS. The motivation is the market structure: Nvidia holds nearly 90% of AI accelerator sales, and persistent shortages plus high costs have left buyers hungry for credible alternatives.

If it materializes, Amazon selling silicon directly would put it in more open competition with Nvidia and other merchant silicon players, leveraging AWS's scale in chip design and manufacturing relationships. Adding to the ambition, AWS SVP Peter DeSantis told CNBC that Amazon expects its own AI models to be competitive with frontier models from partner Anthropic and rival OpenAI within the next year — a striking claim from a company best known as a model host and Anthropic backer.

The context is a chip market where Nvidia's Grace Blackwell superchips remain the performance benchmark and demand vastly outstrips supply. Amazon's pitch would likely emphasize cost-per-token and availability rather than peak performance. Skeptics will note that selling chips externally requires software ecosystems (CUDA's moat is real) and that AWS has historically reserved its best silicon for its own cloud. Watch whether named external customers emerge and whether DeSantis's model-parity claim survives contact with benchmarks.