Amazon Web Services is in active negotiations to sell its Trainium AI accelerator chips directly to third-party data centers — a move that would mark the first time AWS has offered its proprietary silicon outside its own cloud infrastructure and position the company as a direct challenger to Nvidia in the market for AI training hardware.
AWS AI chief Peter DeSantis confirmed the talks in an interview with Bloomberg on June 18, saying Amazon is in 'early discussions' with potential buyers and declined to name them. Amazon CEO Andy Jassy had signaled the direction two months earlier in his April shareholder letter, estimating that if the chip division operated as a standalone business, its 2026 annual run rate would reach approximately $50 billion.
Already $20 billion and growing triple digits
The $50 billion figure is a projection, but the underlying business is already large. Amazon's custom silicon portfolio — Trainium for AI training, Graviton for general-purpose compute, and Nitro for networking — crossed a $20 billion annualized revenue run rate in Q1 2026, growing at triple-digit rates year-on-year, according to figures cited in Jassy's letter.
The headline comparison to Nvidia's current revenue trajectory — Nvidia is on course for roughly $213 billion in fiscal 2026 sales, up 66% — understates how fast Amazon's chip operation is compounding. The gap is real, but the direction of travel matters more than the snapshot.
For context, Nvidia became TSMC's largest customer in 2025, overtaking Apple with a 19% revenue share at the foundry. That dominance over fabrication capacity is precisely the constraint Amazon must navigate if it expands chip sales externally.
The supply problem is the real story
AWS has historically kept Trainium exclusive to cloud customers because the full stack — compute, storage, networking, monitoring — generates higher margin than selling raw silicon. The obstacle to external sales is not strategy; it is wafers.
Trainium3, which reached customers in early 2026, is 'largely sold out,' DeSantis told Bloomberg. Trainium4, the next-generation part, is already absorbing pre-orders despite being roughly 18 months from wide availability. Expanding external sales without additional TSMC wafer allocation would mean diverting chips from AWS's existing cloud customers — a trade-off Amazon has not yet resolved publicly.
DeSantis pushed back on the cannibalisation concern, arguing that 'there is so much underconsumption in AI' that selling chips externally would not shrink AWS cloud revenue. The sovereign AI market strengthens that case: European governments and enterprises seeking to keep AI workloads on domestic infrastructure represent demand that AWS's US-based cloud regions cannot easily serve.
Anthropic and OpenAI already run on Trainium at gigawatt scale
The clearest evidence that Trainium competes technically with Nvidia's H100 and Blackwell families is who is already using it. Anthropic has committed to up to 5 gigawatts of Trainium capacity through AWS; OpenAI has reserved approximately 2 gigawatts as part of a broader commitment of more than $100 billion to AWS infrastructure. Both companies train frontier models on the hardware.
That existing install base is a marketing asset Amazon has not fully exploited in its positioning. When DeSantis approaches hyperscalers and sovereign AI buyers, he can point to two of the world's most scrutinised AI labs already running production workloads on Trainium.
The symmetry with Nvidia's own expansion
The timing is not coincidental. Nvidia CEO Jensen Huang recently declared a new $200 billion market in AI-optimised CPUs, putting Nvidia directly into Intel and AMD territory. Amazon's Trainium push is the mirror image: as Nvidia moves into infrastructure software and general compute, Amazon moves into silicon sales. Both companies are invading each other's margin pools simultaneously.
Jassy drew an explicit historical parallel in his shareholder letter: Graviton, Amazon's custom Arm CPU, is now used by 98% of the top 1,000 EC2 customers after displacing Intel's x86 inside AWS over several years. He expects the same substitution to play out in AI accelerators, and the external sales push is the mechanism that would let that displacement extend beyond Amazon's own cloud.
Engineers push back internally
The chip expansion announcement landed on the same day Amazon confirmed it is investigating three AWS engineers who testified at a Seattle City Council hearing on June 18 in support of a proposed one-year moratorium on new AI data center construction. The engineers, affiliated with Amazon Employees for Climate Justice, argued that data center growth outpaces renewable energy commitments and called for stronger regulation. Seattle's council unanimously approved the moratorium while the city studies infrastructure and environmental impact.
Amazon said employees must follow internal procedures before making public statements representing the company, and that the review 'may or may not' result in action. The engineers filed a civil rights complaint with the Seattle Office for Civil Rights, alleging the investigation constitutes illegal retaliation under the city's political ideology protections.
The juxtaposition captures a structural tension that will follow Amazon's chip ambitions: every gigawatt of Trainium capacity requires a corresponding gigawatt of power, water, and land — resources increasingly contested by local governments and the company's own workforce.
What to watch
Watch whether Amazon announces a named external Trainium buyer before AWS re:Invent in December 2026 — a named customer would signal that early-stage talks have converted into contracts — and whether Amazon secures additional TSMC wafer allocation that would let it supply outside buyers without rationing its own cloud customers.
