The message moving through Amazon’s AI hardware supply chain this quarter is simple: hurry up. AWS has told the manufacturers that assemble its AI servers that it is raising third-quarter 2026 shipment plans by 20% to 30% over original targets, according to people familiar with the matter. The increase points to stronger-than-expected demand for Trainium3, the third-generation custom chip Amazon designed for training and running large AI models.
Trainium3 is the latest product of Annapurna Labs, the chip unit Amazon acquired in 2015, and it sits at the center of Amazon’s effort to reduce its dependence on Nvidia, the dominant supplier of AI accelerators. The revised targets cover the server systems built around the chip, which Amazon has positioned as a lower-cost alternative to the graphics processors that power most of the industry’s AI computing. Supply-chain partners were notified of the increase in recent days, the people said, and component orders for power, cooling, and packaging have been adjusted upward to match.
The scale of the change matters as much as the direction. A 20% to 30% increase on top of an already aggressive build plan suggests AWS sees Trainium3 clearing a hurdle that has tripped up earlier custom-chip efforts: actually winning over the internal teams that train Amazon’s largest models. Amazon runs some of the biggest AI workloads in the world across its retail, advertising, and cloud businesses, and it has said repeatedly that its own silicon will handle an increasing share of that work. The higher shipment targets imply that promise is becoming budget reality.
The move is part of a broader shift across the industry. Google has built its TPU line into a workhorse for both internal and external AI workloads, Microsoft is scaling up its Maia accelerators, and Meta has its own MTIA family in development. Analysts who track the accelerator market say custom chips, or ASICs, are taking a growing share of AI computing, especially for the largest cloud operators, where the economics of running models at massive scale favor chips designed for the specific job. Nvidia remains the default choice for most developers, but supply constraints and cost pressure have given custom silicon a durable opening.
For Amazon, the stakes are financial as much as technical. Training and running AI models on in-house chips cuts the single largest variable cost in its cloud business, and AWS has been renting out Trainium capacity to outside customers through its cloud services as well. The company has said capital spending will keep climbing this year, with the largest share going to data centers and the computing gear inside them. A higher Trainium3 build means more of that spending flows to Amazon’s own design rather than to Nvidia’s profit line.
The risks are familiar ones. Nvidia’s advantage is not just hardware but software: its CUDA programming platform has become the industry standard, and developers are slow to move to alternative stacks no matter how attractive the price per chip. AWS has invested heavily in making Trainium easier to program, and it argues that the savings are large enough to justify the switch for customers running big, predictable workloads. Outside adoption, however, remains modest compared with Nvidia’s installed base.
There is also the question of timing. The revised targets cover one quarter, and hyperscalers have a record of adjusting build plans sharply in either direction as demand shifts. If the AI build-out cools, the same suppliers now being asked to accelerate will be asked to pause. For the moment, though, the direction of travel is clear: AWS is confident enough in Trainium3 to push its supply chain harder, and the companies that make the servers, power systems, and substrates around the chip are the first to know it.
AWS has been here before, in a smaller key. Its first Trainium generation was treated largely as an internal experiment, and the second became a meaningful source of capacity as Amazon pushed its largest workloads onto custom silicon. The third generation is the first that AWS has positioned as a full-fledged alternative to Nvidia’s newest accelerators for outside customers, and the shipment increase suggests the strategy is gaining traction inside Amazon’s own cloud business as well.
People familiar with the matter cautioned that shipment plans are revised routinely as orders firm up, and that a single quarter’s increase does not establish a trend. But they said the magnitude of the change — two to three tenths on top of an already-large build — reflects real demand signals rather than speculative inventory building. Component suppliers have been asked to hold higher buffer stocks, and assembly capacity has been reserved further into the year.
The broader signal for investors is that the AI compute market is becoming a multi-supplier contest. Nvidia’s share of accelerator shipments has been dented by custom silicon from the very customers it supplies, and every design win for Trainium, TPU, or Maia is a data point in that story. Amazon’s decision to raise its own build is the latest and most direct evidence that the largest buyers of AI chips are determined to have alternatives — and are willing to fund them.
