Startups find Amazon's AI chips 'less competitive' than Nvidia GPUs, internal document shows (www.businessinsider.com)

An internal July Amazon document obtained by Business Insider shows several AI startups — notably Cohere and Stability AI — found Amazon’s in-house Trainium 1/2 chips lagged behind Nvidia’s H100 GPUs on latency and overall cost-performance. Cohere reported extremely limited access to Trainium 2 and frequent service disruptions; Stability AI called Trainium 2 “less competitive.” Other customers flagged Inferentia 2 (for inference) as less cost-efficient than Nvidia A100s, and independent tests found AWS G6 servers with Nvidia GPUs often outperform AWS’s custom chips. Nvidia today owns roughly 78%+ of the AI accelerator market, while AWS chips sit near ~2%, illustrating a steep performance and ecosystem gap (CUDA familiarity, mature tooling, and developer expertise favor Nvidia). The findings matter because AWS is banking on custom silicon to protect margins and scale cloud AI services without buying expensive Nvidia GPUs. If startups and large customers insist on Nvidia hardware, AWS could face higher costs and slower path to profitable AI growth. Amazon says Trainium/Inferentia have strong wins with some customers (Anthropic is a major Trainium user and Project Rainier will deploy hundreds of thousands of Trainium chips), touts 30–40% better price-performance claims, and promises Trainium 3 previews later this year. For the AI community the story underscores that raw hardware performance, stability, availability, and ecosystem/tooling (not just chip specs) determine adoption — and that Amazon must close those gaps to break Nvidia’s near-monopoly.