Amazon to Sell Trainium AI Chips Directly — Nvidia Challenged

Amazon confirmed on June 18 that it is in active negotiations to sell its custom-designed Trainium AI chips to external customers for deployment in third-party data centers — a significant strategic shift from its previous AWS-only model. CEO Andy Jassy has framed the direct silicon sales opportunity as worth approximately $50 billion in potential revenue.

The Trainium3 chip is Amazon's most advanced AI accelerator, reportedly delivering training performance at roughly 80% lower cost than Nvidia's H100 GPUs. AWS has already secured $225 billion in multi-year compute commitments from major AI clients including OpenAI and Anthropic, giving Amazon leverage to bundle or unbundle chips as it sees fit.

The move directly threatens Nvidia's near-monopoly on AI training and inference hardware. If Amazon follows through, it becomes the first hyperscaler to sell merchant silicon in direct competition with chip companies rather than exclusively wrapping chips into cloud services. This would give customers a credible alternative for large-scale AI compute without depending on Nvidia's supply-constrained GPUs.

For the broader market, the implications are significant. AI infrastructure costs have been a bottleneck for startups and enterprises alike. A well-capitalized alternative with proven silicon could compress margins across the AI hardware stack — from training to inference. The question now is whether Amazon will offer Trainium at price points that make non-AWS deployments economically rational, or whether the chips will come with AWS-ecosystem strings attached.

For MENA data center operators — particularly in Saudi Arabia and UAE where sovereign AI infrastructure is a national priority — a credible Nvidia alternative at lower cost could accelerate regional AI buildouts.