Amazon is in talks to sell its custom artificial intelligence chips directly to outside buyers as part of a bid to undercut Nvidia's grip on the AI accelerator market, Bloomberg reported.

The discussions would, if they lead to merchant availability, mark a strategic shift for Amazon Web Services. AWS has spent years building custom AI silicon under the Trainium training accelerator and Inferentia inference chip product lines — both manufactured by TSMC — but has so far restricted those parts to customers running workloads inside AWS infrastructure. Direct sales would put Amazon into head-to-head competition with Nvidia, the merchant chip vendors chasing the same customers, and the cloud businesses of Google and Microsoft.

Why the timing matters

Demand for AI compute has outstripped supply for two years running. Nvidia's data center GPUs continue to command premium prices and rationed allocation, and large enterprise buyers have publicly called for alternatives. AWS has both the chip-design depth and the foundry relationships to credibly serve that demand — and a financial incentive, because merchant sales would amortize the fixed R&D and mask-fabrication costs of Trainium and Inferentia across a larger volume base.

Bloomberg's report did not disclose pricing, target customers, geographic scope, or whether AWS would sell to its cloud competitors. Those details will define the actual commercial impact.

How Trainium and Inferentia fit

Trainium and Inferentia are not positioned as direct performance peers to Nvidia's flagship GPUs. AWS has consistently marketed them as offering competitive throughput for specific workload profiles — primarily large-batch training and high-volume inference — at lower price points. That positioning maps onto a buyer segment that exists but has had limited alternatives: enterprises that need meaningful AI capacity but cannot justify Nvidia's premium pricing for every workload.

If AWS prices standalone Trainium or Inferentia parts materially below comparable Nvidia products, the move could compress margins across the AI accelerator segment. If AWS prices them at parity, the move would still serve a volume goal but would not materially undercut Nvidia's economics.

The hyperscaler competitive picture

All three major U.S. cloud providers now operate custom AI silicon programs:

• Google's Tensor Processing Unit, accessible externally through Cloud TPU since 2018
• Microsoft's Maia chip, deployed at scale inside Azure data centers
• AWS's Trainium and Inferentia, used inside AWS and offered through EC2 instances

Amazon's reported merchant push would represent the most aggressive external commercialization among the three. Google has effectively monetized TPUs as a cloud service rather than a standalone product. Microsoft has not yet sold Maia to outside buyers.

An AWS merchant move would sharpen the question of whether Microsoft and Google follow — and on what timeline.

What remains to be seen

The talks remain just that — talks. Bloomberg's report contains no announcement of pricing, no customer commitments, no product SKUs, and no launch timing. The next inflection points are the ones that will determine commercial impact:

• Formal announcement of standalone chip SKUs for purchase outside AWS
• Pricing benchmarks versus comparable Nvidia parts
• Nvidia's response on pricing or allocation
• Whether Microsoft or Google announce their own merchant silicon strategies

Until those signals materialize, the Bloomberg report functions as a strategic signal rather than a market event. But a signal from a hyperscaler with the chip-design depth, the foundry access, and the installed customer base of Amazon carries commercial weight regardless of how thin the details are.