During the Q3 2026 earnings call held on October 29, 2026, Amazon CEO Andy Jassy confirmed that the company will begin selling its Trainium AI chips to third-party data centers outside of AWS by the end of 2027. This will be the first time Trainium is available for open sale as a standalone “merchant” product in physical hardware form, directly challenging Nvidia’s dominance in the AI accelerator market. Buyers, prices, and future supply volumes have not yet been disclosed.
What Happened
On October 29, 2026, during Amazon’s Q3 2026 earnings call, CEO Andy Jassy confirmed that the company will begin selling its Trainium AI chips to third-party data centers outside of AWS by the end of 2027. This will be the first time Trainium is offered as a standalone “merchant” product, rather than solely as compute rentals via AWS. During the same call, Amazon reported its financial results: AWS revenue grew 39.3% year-over-year to $46.0 billion, operating margin reached a record 13.9%, and the combined annual run-rate for the Trainium and Graviton chip businesses exceeded $25 billion, according to the company.
Context
Trainium is Amazon’s proprietary AI accelerator, which until now has operated exclusively within AWS infrastructure: customers could only rent compute on it, not purchase the chips themselves. Selling hardware to third-party operators is the “merchant silicon” market, where Nvidia has historically dominated with an estimated 81–90% market share in AI accelerators. Nvidia’s strength lies not only in its hardware but also in the CUDA ecosystem, under which a significant portion of existing code is written; Amazon’s response is focused precisely here—the open implementation of TorchNeuron allows running existing PyTorch code on Trainium without rewriting it for the Neuron SDK, lowering the barrier to entry. According to Jassy, the only precedent for a major cloud provider combining its own chip for the external market is Alphabet with its TPU accelerators. Characteristics of the generations are currently stated by the vendor itself: Trainium2 offers approximately 30% better price/performance, while Trainium3 operates 30–40% faster than Trainium2.
Why This Matters for the Industry
For the industry, Amazon is transforming from a cloud competitor into a chip market competitor: compute buyers now have a second merchant track for AI accelerators alongside Alphabet’s TPU, and companies with their own infrastructure have the chance to reconsider their dependence on Nvidia when planning capacity for inference-heavy services, including agents and RAG workloads. The signaling effect is already working: the CEO’s public commitment to external sales strengthens the negotiating position of everyone currently bargaining with Nvidia on prices for large batches of accelerators. If the claimed price/performance advantage is confirmed by independent measurements, the economics of training and inference will change most noticeably for mid-size data center operators and neocloud providers, for whom purchasing hardware will become a real alternative to renting. In the time remaining until deliveries, it is logical for engineering teams to prepare platforms for multi-hardware support: introduce a second “hardware profile,” abstractions over serving, and reproducible benchmarks to avoid tying products to a single stack.
Why This Matters for Users
For readers and teams, the announcement promises a more accessible alternative to Nvidia: in addition to renting GPUs in the cloud, there will be an opportunity to purchase physical Trainium chips for their own data centers. Developers do not need to wait—they can already rent Trainium2 in AWS, port their PyTorch code via TorchNeuron, and get their own numbers: throughput, p99 latency, cost per 1 million tokens, and quality after Neuron compilation. Owners of their own data centers will only have access to Trainium racks once external deliveries begin, so it is too early to rely on them in current procurement. Investors should track one leading indicator—the first publicly named external buyer of Trainium, which will show that the announcement is turning into a real merchant channel.
What Is Still Unknown / Limitations
Commercial terms for external sales have not been disclosed: Amazon has not yet named any customers, prices, or supply volumes, so the announcement remains a business declaration. It is unknown which chip generation will be sold externally—Trainium2 or Trainium3—and how software stack support will be organized outside of AWS. Vendor-claimed price/performance characteristics have not been confirmed by independent benchmarks like MLPerf, and the measurement methodology has not been published. The $25 billion annual run-rate figure combines Trainium and Graviton, and the Graviton CPU business alone does not prove the competitiveness of Trainium as an AI accelerator. Amazon will review the balance between sales and rentals approximately annually, so the commitment format may change. Finally, part of the ecosystem remains outside the scope of the announcement: custom kernels, Triton extensions, and debugging tools are still tied to Nvidia, and no equivalents for Trainium are mentioned in the sources.
Sources
Author
Look at AI, editorial team