Amazon Web Services is in early-stage discussions to sell its Trainium AI accelerator chips directly to other companies for use in their own data centers – a potential shift from captive internal use to open-market availability that would represent the most direct competitive challenge Amazon has yet posed to Nvidia’s dominance in the AI chip market. Amazon’s AI chief Peter DeSantis confirmed the direction in a public interview, describing a commercial logic rooted in the conviction that demand for AI computing is so large that external chip sales would not displace AWS cloud revenue. “There’s so much underconsumption in AI. I’m not worried about it,” DeSantis stated, describing external sales as a way to reach more customers rather than a reallocation of supply from existing ones. NEWSCENTRAL assesses this move as strategically significant regardless of its ultimate scale, because the credible availability of a lower-cost alternative to Nvidia’s GPUs changes the pricing power dynamics of the AI chip market even before a single Trainium chip is sold externally.
The commercial foundation for external sales is more substantial than a typical early-stage exploration would suggest. Amazon’s custom silicon business – spanning Trainium, Graviton, and Nitro chips – crossed a $20 billion annual revenue run rate in the first quarter of 2026, growing at a triple-digit pace with demand from major AI customers already committed. OpenAI has agreed to approximately 2 gigawatts of Trainium capacity through AWS, and Anthropic has signed agreements for up to 5 gigawatts of current and future Trainium chips. Uber is also a confirmed customer. Trainium3 is nearly sold out, and pre-orders for Trainium4 – which will not be available for more than a year – have already been largely reserved. In his April shareholder letter, CEO Andy Jassy stated that if Amazon’s chip business operated as a standalone company selling to both AWS and third parties, its annual run rate would reach approximately $50 billion. That figure is not a projection of external sales revenue; it is a measure of the gap between the value Amazon’s chips currently generate and what they would generate if priced to the external market.
The cost argument for Trainium is Amazon’s primary competitive lever against Nvidia. The company has consistently argued that Trainium performs equivalent AI workloads at a materially lower cost than Nvidia’s GPUs, with some estimates placing the price advantage at 80% for comparable tasks. That claim has been validated by the customer commitments already in place: enterprises including Anthropic and OpenAI, which have access to both Nvidia and Amazon silicon through their own procurement channels, have signed substantial Trainium agreements on commercial rather than exclusively strategic grounds. Lucas Grant, Semiconductor and Manufacturing Strategy Analyst at NEWSCENTRAL, points out that the cost differential is not primarily a function of chip design sophistication but of vertical integration: Amazon controls the software stack, the data center infrastructure, and the chip architecture simultaneously, enabling optimization across the full system rather than at the component level alone. That integration advantage is difficult for Nvidia to replicate, just as Nvidia’s software ecosystem advantage through CUDA is difficult for Amazon to replicate.
The supply constraint is the most immediate obstacle to external sales at meaningful scale. Trainium capacity has sold out almost immediately upon availability, meaning that selling chips externally would require either manufacturing a surplus or creating a waitlist for existing AWS customers – neither of which is straightforward. TSMC, which manufactures Trainium chips, has recently displaced Apple as its largest customer, and Nvidia is simultaneously the foundry’s most strategically important client for advanced node capacity. Amazon expanding external chip sales without securing a significant increase in TSMC allocation would create supply allocation conflicts that could damage relationships with the cloud customers whose satisfaction is the foundation of the AWS business. The discussions DeSantis described as early stage likely include, as an unspoken precondition, the negotiation of additional TSMC capacity that would make external sales additive rather than competitive with existing commitments.
The competitive implications for Nvidia are real but require proportional framing. Nvidia’s data center revenue rose 92% year-on-year to $75.2 billion in its most recent reported quarter, and its business continues to accelerate rather than plateau. Amazon entering the external chip market as a merchant silicon provider would add a credible low-cost alternative to a market that is currently dominated by a single vendor, creating pricing pressure that benefits every AI infrastructure buyer and constraining the premium that Nvidia can command for GPU supply. Freddy Miller, Senior Analyst at NEWSCENTRAL, argues that the more consequential effect may not be on Nvidia’s near-term revenue but on its long-term CUDA ecosystem advantage: if Trainium chips become widely deployed in external data centers, the incentive for software developers to optimize their workloads for Trainium rather than CUDA increases, potentially beginning the slow erosion of the software moat that has been Nvidia’s most durable competitive asset.
The most important qualifier in the entire discussion is the word “early.” The conversations DeSantis described are preliminary, the supply constraints are genuine, and the timeline for any meaningful external chip sales program is measured in years rather than quarters. What NEWS CENTRAL finds significant about the announcement is not the imminence of external sales but the signal it sends: Amazon has decided that the strategic and commercial case for open-market chip sales is strong enough to pursue, and that decision will shape how the company allocates manufacturing capacity, builds software tooling, and prices its cloud compute products for the rest of this decade.