For years, one company has quietly taxed the entire AI boom: Nvidia, whose GPUs power almost every major AI model. On June 18, 2026, Amazon signaled it wants to break that grip — confirming it's in talks to sell its own Trainium AI chips directly to outside companies, not just rent them out through AWS.
It's a deceptively big move. Amazon going from cloud landlord to direct chip seller takes aim at the very heart of Nvidia's business — and it's pitching comparable AI power at roughly half the cost. Here's the full breakdown of what's happening and why it could reshape the AI hardware market.
What Happened
Amazon's AI chief, Peter DeSantis, told Bloomberg that Amazon is negotiating to sell its Trainium3 processors directly to external data centers. He declined to name potential buyers, but the intent is clear: for the first time, Amazon would sell its proprietary AI accelerator hardware outside of AWS.
That single change flips Amazon's role. Instead of only offering Trainium as a cloud service, it would compete with Nvidia the way Nvidia competes with everyone — by selling chips.
What Is Trainium?
Trainium is Amazon's family of custom AI accelerator chips, designed in-house specifically to train and run large AI models. Until now, the only way to use Trainium was through Amazon Web Services. The latest generation, Trainium3, launched in late 2025 and was built to go toe-to-toe with Nvidia's GPUs on both performance and price.
Custom AI silicon like this is how the cloud giants reduce their dependence on Nvidia — and Amazon's has matured into a serious product, not a science project.
Trainium3 vs Nvidia: The Pitch
Amazon's sales pitch comes down to price-performance. Here are the headline claims.
| Metric | Amazon's Claim |
|---|---|
| Performance | ~4x the performance of Trainium2 |
| Cost | ~half the cost of conventional GPUs |
| Availability | Near capacity since late-2025 launch |
| Chip division revenue | ~$20 billion annual run rate |
If those numbers hold up in real workloads, Trainium3 offers a compelling deal: similar AI compute for roughly half the money. In a market where AI infrastructure costs are exploding, that's exactly the lever buyers care about — a theme we explored in the 2026 AI price war.
Why Nvidia Should Worry
Plenty of companies build custom AI chips. What makes this different is the business model shift:
- It attacks Nvidia's core, not just the cloud. Until now, Trainium only competed inside AWS. Selling hardware directly puts Amazon in Nvidia's actual market — the open sale of AI accelerators.
- Price is the weapon. Nvidia's GPUs command premium margins. A credible rival at ~half the cost pressures both Nvidia's market share and its pricing power.
- Amazon has scale and trust. A $20B chip business and a global logistics/enterprise footprint make Amazon a far more threatening entrant than a startup.
- The whole industry wants alternatives. Every major AI buyer is desperate to reduce its Nvidia dependence; Amazon is offering an exit.
The $225B Demand Signal
How real is the appetite? Reportedly, Trainium commitments from OpenAI and Anthropic alone have reached around $225 billion. That's a staggering signal that the biggest AI labs already want Amazon's silicon at scale — and helps explain why Trainium3 has run near capacity since launch.
That demand is also the strongest argument for selling externally: if customers want the chips badly enough to commit hundreds of billions, limiting them to AWS leaves enormous money on the table. The compute behind every frontier model is now a strategic battleground (see our June 2026 AI model roundup).
What It Means for AI
Step back and the implications are big:
- Cheaper AI compute. More competition at the chip level should push down the cost of training and running models — good for startups, enterprises, and ultimately users.
- A less Nvidia-dependent world. A genuine second source for high-end AI chips reduces a major bottleneck and single point of risk in the AI supply chain.
- The cloud giants go vertical. Amazon joining Google (TPUs) and others in custom silicon shows the hyperscalers want to own the full stack — chips, cloud, and models.
- Caveat: it's early. This is still at the talks stage, with no named customers or launch date, and Nvidia isn't standing still. Execution will decide how much it actually dents Nvidia.
Frequently Asked Questions
What is Amazon doing with its Trainium AI chips?
Amazon is in talks to sell its Trainium3 AI chips directly to outside companies for use in their own data centers, rather than offering them only through AWS. Announced by Amazon's AI chief Peter DeSantis on June 18, 2026, it would be the first time Amazon sells its proprietary AI accelerator hardware outside its cloud — putting it in direct competition with Nvidia.
What is Amazon Trainium?
Trainium is Amazon's family of custom AI accelerator chips, designed in-house to train and run large AI models. Until now it has been available only via Amazon Web Services. The latest generation, Trainium3, launched in late 2025 and is built specifically to rival Nvidia's GPUs on performance and cost.
How does Trainium3 compare to Nvidia GPUs?
Amazon says Trainium3 delivers about four times the performance of Trainium2 at roughly half the cost of conventional GPUs. That price-performance edge is the core of Amazon's pitch: comparable AI compute for substantially less money than Nvidia's hardware, which has dominated the market.
Why is this a threat to Nvidia?
Nvidia's dominance rests on selling AI GPUs to everyone. If Amazon starts selling Trainium hardware directly to other data centers — at roughly half the cost — it attacks Nvidia's core business model, not just competes in the cloud. With Amazon's chip division already at a $20 billion annual revenue run rate, it's a credible challenger.
Who would buy Amazon's Trainium chips?
Amazon hasn't named specific buyers, but the demand signal is huge: Trainium commitments from OpenAI and Anthropic alone have reportedly reached around $225 billion. Likely customers are AI labs, large enterprises, and data-center operators seeking cheaper alternatives to Nvidia GPUs for training and running big models.
Won't selling chips hurt Amazon's AWS cloud business?
Amazon executives say no — they don't view selling Trainium hardware outside AWS as a meaningful risk to cloud revenue. The logic is that expanding into the hardware market grows the overall pie, and many buyers who run chips on-premises were never going to be AWS cloud customers anyway.
How big is Amazon's AI chip business?
Amazon's custom-chip division has reached a roughly $20 billion annual revenue run rate, and Trainium3 has run nearly at capacity since its late-2025 launch. Selling directly to outside customers could expand that business substantially and reduce the entire industry's dependence on Nvidia.
When will Amazon start selling Trainium chips externally?
As of June 2026 the plans are in the talks/negotiation stage, confirmed by Amazon's AI chief in a Bloomberg interview. No firm launch date or customer list has been announced, so timing and terms could still change before any external sales begin.
Final Thoughts
Amazon selling Trainium chips on the open market would be one of the most consequential shifts in the AI hardware landscape since the boom began. For the first time, Nvidia faces a rival with the scale, the product, and the price advantage to take real share — and an industry that's eager for an alternative.
It's still early, and talks aren't shipments. But the direction is unmistakable: the AI chip market is opening up, compute is getting cheaper, and Nvidia's near-monopoly is finally being tested. We'll keep tracking the AI hardware war as Amazon's plans firm up.