Amazon to sell custom Trainium AI chips to third parties, challenging Nvidia

For a decade, Amazon's custom chips were a closed-loop advantage: designed at Annapurna Labs, deployed only inside AWS, never sold. The talks Bloomberg surfaced break that model — Amazon is now considering selling Trainium directly to other companies for use in their own data centers ^[5]. An AWS spokesperson confirmed the company has historically declined to sell chips directly but may now sell racks to third parties ^[1]. This is the same pivot Google just made with its TPUs, which it announced it would sell into select external data centers ^[2]. The strategic logic is that captive silicon caps your addressable market at your own cloud; a merchant chip lets you monetize every wafer you can fabricate, regardless of whose data center it lands in. DeSantis framed it plainly, saying Amazon views AI infrastructure as rapidly evolving and is constantly looking at ways to reach more customers ^[2]. The signal that demand exists beyond AWS is concrete: third-generation Trainium is described as largely sold out, with fourth-generation parts expected next year ^[4].

The money reframes how seriously to take this. Amazon's custom-silicon business crossed roughly a $20 billion annual revenue run rate in Q1 2026, growing at a triple-digit pace ^[3]. Jassy has gone further, saying that if the chips business were standalone and sold this year's production to AWS plus third parties, the annual run rate would be around $50 billion ^[1]. For scale, that would rival or exceed standalone merchant chipmakers. But the incumbent gap is still enormous: Nvidia's data-center revenue rose 92% year over year to a record $75.2 billion in a single fiscal quarter, against a roughly $326 billion run rate ^[1][3]. The wedge Amazon is betting on is price — it claims Trainium delivers 30 to 50 percent cost reductions versus comparable GPU setups, with some reports putting Trainium3 inference at roughly half the cost of an H100 ^[3]. A $50B run rate against a $326B incumbent is not a takeover; it is the establishment of a credible second source in a market that previously had none.

The hardest objection to Amazon's pitch is software, not silicon. Across social discussion the debate consistently returns to Nvidia's CUDA ecosystem as the real defensive moat — cheaper chips matter little if porting models off CUDA is painful, and the contrarian read is that Trainium shines mainly when frontier labs co-engineer deeply with AWS rather than for general-purpose customers picking it off a shelf. The strongest counter-evidence to that skepticism is adoption by serious labs: Anthropic, OpenAI, and Uber are cited as Trainium users ^[4], and Anthropic's deepening Amazon partnership is widely read as the most convincing proof that the cost wedge can hold at frontier scale. Sentiment splits along these lines. The investor-facing crowd is bullish, treating the merchant-chip move as a deliberate strategic offensive Jassy telegraphed on the Q1 earnings call. The builder-facing crowd is more skeptical-but-watchful, granting that cost is the genuine wedge while insisting CUDA lock-in, existing Nvidia cluster agreements, and shared TSMC dependence blunt how fast that wedge can cut. DeSantis, for his part, dismisses the cannibalization worry outright, saying there is so much underconsumption in AI that he is not worried about it ^[2].

The most overlooked detail is that Amazon is not betting against Nvidia even as it competes with it. The company is still deploying more than one million Nvidia GPUs starting in 2026 alongside its own custom chips ^[3]. That dual posture reveals the actual strategy: Trainium is a margin and supply hedge, not a replacement. By owning a low-cost alternative, Amazon gains pricing leverage on its Nvidia purchases and a fallback if GPU supply tightens, while still offering customers the Nvidia stack they already know. A second-order driver is geography — demand for sovereign AI infrastructure, particularly European pressure to reduce dependence on US technology, opens an external market for locally controlled compute that Trainium racks could fill ^[4]. The near-term market read was mild but positive: Amazon stock rose about 2.5%, closing at $243.56 on the news ^[4]. Analysts framed the development as something for Nvidia investors to monitor rather than fear, since the AI accelerator market is growing fast enough to support more than one winner ^[3].