Amazon Weighs Selling Its Trainium AI Chips Directly to Challenge Nvidia

Amazon may be about to do something it has never done: sell its in-house AI chips to other companies. On June 18, 2026, reporting revealed that Amazon is in early talks to sell its Trainium AI chips directly to data centers, a sharp break from a strategy that has kept the silicon locked inside Amazon Web Services. If it happens, it would put Amazon into more direct competition with Nvidia.

Key takeaways

• Amazon is in early talks to sell its Trainium AI chips directly to outside data centers, per June 18, 2026 reporting.
• This breaks from Amazon's long-standing approach of offering Trainium only through AWS cloud rentals.
• AWS AI chief Peter DeSantis confirmed the company is in discussions about selling the chips externally.
• CEO Andy Jassy had floated the idea in his April shareholder letter, framing it as a roughly $50 billion opportunity.
• Amazon's third-generation Trainium3 is said to reach rack-scale performance roughly on par with Nvidia's Blackwell NVL72.

What happened

For years, Amazon designed its own AI chips, Trainium for training models and Inferentia for running them, and offered them exclusively as rentable capacity inside AWS. Customers could use the chips by running workloads in Amazon's cloud, but they could not buy the silicon to install in their own facilities. The June news signals that this may change.

According to reporting, AWS AI chief Peter DeSantis said the company is in talks to sell Trainium to other companies for use in their data centers. The discussions are early, and no deal had been signed as of late June. But the direction is significant: it follows comments from CEO Andy Jassy in his April shareholder letter, where he said the homegrown chips were so sought-after that selling them directly could be a roughly $50 billion opportunity.

The technical context matters. Amazon's third-generation chip, Trainium3, is described as reaching rack-scale performance roughly comparable to Nvidia's Blackwell NVL72 system, reportedly the first time Amazon's silicon has closed that gap at the rack level. Closing in on Nvidia's flagship is what makes external sales plausible in the first place.

Here is how the two approaches stack up for a data center operator weighing the options:

Factor	Buy Trainium directly	Rent Trainium on AWS	Buy Nvidia (Blackwell)
Upfront cost	High capex, potentially lower per-chip	None, pay-as-you-go	Highest capex, premium pricing
Software maturity	Neuron SDK, still catching up	Neuron SDK, AWS-managed	CUDA, the industry default
Availability	Not yet sold (talks only)	Available now in AWS	Constrained, long lead times
Lock-in	Your hardware, your facility	Tied to AWS cloud	Vendor lock to CUDA ecosystem
Best for	Operators wanting an Nvidia alternative	Teams already on AWS	Anyone needing the mature toolchain today

Why it matters

Nvidia has dominated the market for AI training hardware so thoroughly that its chips are the default, and often a bottleneck, for anyone building large models. A credible alternative would matter enormously. By selling Trainium directly, Amazon could offer data centers a potentially lower-cost option and chip away at Nvidia's grip on the most important hardware in technology.

It would also be a strategic shift for Amazon itself. Keeping Trainium AWS-only made the chips a reason to use Amazon's cloud. Selling them externally trades some of that lock-in for a chance at a large new hardware business, the kind of $50 billion opportunity Jassy described. That tension, cloud differentiation versus hardware revenue, is exactly why the talks are still early.

The move fits the broader scramble to build AI compute at scale, the same forces behind the enormous capital commitments we covered in Alphabet's $80 billion buildout and the memory boom lifting Micron. More competition at the chip level could ease supply pressure across the whole industry.

What is next

Talks are early, and a lot has to happen before Trainium shows up in third-party data centers. The questions to watch:

1. A signed deal. Whether the early discussions turn into actual sales agreements.
2. Software support. How easily customers can move workloads to Trainium versus Nvidia's mature ecosystem.
3. Pricing. Whether Amazon can undercut Nvidia enough to win buyers without eroding AWS demand.
4. Performance proof. How Trainium3 holds up against Blackwell in real, large-scale deployments.

Frequently asked questions

What is Amazon Trainium?

Trainium is Amazon's custom-designed chip for training AI models. Until now it has been available only by renting capacity on AWS rather than buying the physical hardware.

What changed in June 2026?

Reporting on June 18, 2026 revealed Amazon is in early talks to sell Trainium chips directly to outside data centers, a break from its AWS-only model. No deal had been signed as of late June.

How does Trainium compare to Nvidia?

Amazon's third-generation Trainium3 is said to reach rack-scale performance roughly on par with Nvidia's Blackwell NVL72, reportedly the first time Amazon's silicon has matched Nvidia at the rack level.

Why would Amazon sell chips it could keep exclusive?

CEO Andy Jassy framed direct sales as a roughly $50 billion opportunity. The trade-off is between keeping Trainium as an AWS draw and pursuing a large new hardware revenue stream.

Nothing is signed yet, and Nvidia's lead is formidable. But the mere fact that Amazon is considering selling its chips on the open market is a signal of how much the AI hardware landscape is starting to shift.