AWS Plans to Sell Its AI Chips to External Data Centers
Amazon Web Services is in active talks to sell its custom Trainium and Inferentia AI chips to third-party data centers, a move CEO Andy Jassy says represents a $50 billion opportunity. The strategy marks a significant shift from using the chips exclusively for internal AWS workloads.
Original sourceAmazon has long designed custom silicon — its Trainium training chips and Inferentia inference chips — primarily to reduce its own dependence on Nvidia GPUs and lower the cost of running AI workloads internally. Now AWS is reportedly in talks to sell those chips externally to other data center operators, a move that would put Amazon in direct competition with Nvidia not just as a cloud provider, but as a chip supplier.
CEO Andy Jassy has publicly framed the external chip sales as a $50 billion opportunity, signaling that Amazon views this as a serious revenue line rather than a side experiment. The pitch to external buyers would likely rest on cost and availability advantages — Nvidia's H100 and H200 chips have faced chronic supply shortages, and any credible alternative with competitive performance-per-dollar could find willing buyers among hyperscalers and colocation operators.
The competitive dynamics here are meaningful. Nvidia's dominance in AI compute has been sustained by its CUDA ecosystem as much as its hardware — developers write to CUDA, and switching costs are enormous. Amazon's chips run on a different software stack, which means external adoption isn't just a procurement decision; it's a re-platforming decision for any serious ML team. AWS has made progress with its Neuron SDK, but closing the software ecosystem gap with Nvidia remains the central challenge.
If Amazon succeeds in establishing an external market for its chips, the implications extend beyond chip revenue. It would give AWS leverage in enterprise AI deals, create a new distribution channel for its broader cloud services, and potentially reshape how AI infrastructure costs are structured across the industry. Whether the $50 billion figure reflects genuine signed pipeline or optimistic TAM math is the question the market will be asking.
Panel Takes
The Futurist
Big Picture
“The thesis here is falsifiable and specific: CUDA's moat is software, not hardware, and if AWS can get enough external operators buying Trainium, it seeds an alternative ecosystem that makes Neuron a viable target for ML framework optimization. The dependency chain is brutal though — this only works if PyTorch and JAX teams prioritize Neuron support at parity with CUDA, which requires adoption volume that doesn't exist yet. The second-order effect that nobody's pricing in: if Amazon becomes a chip supplier to other data centers, it gains visibility into competitor workloads in a way that will eventually become a regulatory conversation.”
The Founder
Business & Market
“The buyer here is a data center operator or colocation provider who is currently paying Nvidia spot prices or waiting six months for allocation — that's a real, liquid pain point with a real budget attached. The moat question is whether AWS can create switching costs on the software side before Nvidia restores supply availability, because the moment H100s are in stock at list price, the urgency for every procurement conversation evaporates. Jassy's $50 billion number needs scrutiny: that's a TAM claim dressed up as a pipeline claim, and those are very different things when you're deciding whether to build out a chip sales organization.”
The Skeptic
Reality Check
“The specific scenario where this breaks is any ML team running serious training workloads — they're not switching off CUDA because their entire toolchain, their existing model checkpoints, and their ops team's institutional knowledge lives there. Amazon has been shipping Trainium since 2021 and Neuron SDK adoption outside of AWS's own infrastructure remains marginal, which is the data point that matters more than Jassy's $50 billion aspiration. What would change this verdict: a major foundation model lab publicly commits to training a flagship model on Trainium externally and publishes the benchmark comparison with methodology included.”
The Builder
Developer Perspective
“The primitive is a non-CUDA AI accelerator you can rack in your own data center, and the DX bet Amazon is implicitly making is that Neuron SDK has closed enough of the gap that migration isn't a rewrite. I've worked with Neuron and the honest answer is that it handles inference on standard transformer architectures reasonably well but falls apart the moment you're doing anything custom — kernel fusion, exotic attention patterns, anything off the beaten path. Until AWS ships a compatibility layer that lets teams run existing CUDA code without rewriting it, this is a chip for workloads you're willing to redesign, not workloads you already have.”