Amazon AI Chips Muscle In on Nvidia’s Cash Machine

Amazon AI chips were supposed to be an internal AWS advantage. Now Amazon is testing whether they can become a product that other data centers buy directly, putting the company closer to Nvidia’s core business than ever before.

AWS is in early talks to sell its Trainium AI chips to other companies for use in data centers, Amazon AI chief Peter DeSantis told Bloomberg, according to TechCrunch. DeSantis did not name potential buyers, and Amazon told TechCrunch the discussions are still early.

That caution matters. This isn’t yet a full merchant chip business. It’s a signal that Amazon may no longer be satisfied with renting AI compute only through AWS.

Amazon's AI chip sales pitch targets Nvidia's most valuable bottleneck

The usual assumption is simple: Nvidia sells the chips, cloud giants buy them, and customers rent access. Amazon is now probing a different model. It wants to design the AI silicon, run it inside AWS, and possibly sell racks of it to third-party data centers.

That would push Trainium beyond AWS’s walls. It would also shift Amazon from being a huge chip customer into a more direct chip supplier.

The strategic bet is not that Amazon must crush Nvidia on every workload. The more practical goal is narrower: offer enough performance, enough availability, and enough cost appeal for large AI customers that don’t want every training or inference job tied to Nvidia GPUs.

That is where the tension sits. Nvidia still dominates the high-end AI chip market in the source reporting. But Amazon has a distribution advantage Nvidia does not have in the same form: AWS already sells compute, storage, networking, security, and monitoring to the same customers that need AI infrastructure.

For context on the physical infrastructure pressure behind AI buildouts, XOOMAR has also covered Amazon Data Centers Clash Lands Engineers in Crosshairs. The chip race is not just about silicon design. It’s about who can get capacity into usable data centers fast enough.

The $50 billion Amazon AI chips opportunity Andy Jassy wants AWS to capture

Amazon CEO Andy Jassy put a hard number on the opportunity in his annual shareholder letter in early April.

"If our chips business was a standalone business, and sold chips produced this year to AWS and other third parties (as other leading chips companies do), our annual run rate would be ~$50 billion. There’s so much demand for our chips that it’s quite possible we’ll sell racks of them to third parties in the future."

That quote is the center of the story. A $50 billion annual run-rate chip business would not “tank” Nvidia, TechCrunch notes, because Nvidia is currently on a $326 billion revenue run rate if it keeps delivering quarters like the last one. But $50 billion is still large enough to resemble Intel’s annual revenue scale, according to the same report.

The business model change is the point.

That second model is more aggressive. It also creates a problem Amazon can’t talk around: if Trainium capacity is already scarce, selling chips outside AWS could mean choosing between cloud customers and hardware buyers.

Trainium gives Amazon a credible opening, but not a clean shot at Nvidia

The source reporting centers on Trainium, Amazon’s homegrown AI chip for AWS. It does not establish a direct external sales plan for other Amazon AI chips, so the clean factual read is narrower than the hype: Amazon is talking about selling Trainium capacity or racks, not announcing a full open chip catalog.

Jassy’s shareholder letter makes the supply issue stark. He said current Trainium capacity sold out almost instantly. He also said capacity for the next chip, Trainium4, had already sold out, even though it would not be available for more than a year.

That creates the sharpest contradiction in Amazon’s plan.

Before vs. after the talks:

• Before: Trainium strengthened AWS by giving cloud customers an in-house alternative to outside chips.
• After: Trainium could become a product for other data center operators.
• Constraint: Amazon may need more supply before it can serve both groups without leaving current AWS customers waiting.
• Risk: More external sales could weaken the “AWS first” advantage that made the chips valuable in the first place.

TechCrunch also notes that selling more chips would depend on manufacturing partners such as TSMC. That is not a small obstacle. The report says Amazon would have to “elbow Nvidia out of the way” at TSMC, where Nvidia has recently supplanted Apple as the foundry’s largest customer.

AWS wants the chip sale, but the cloud bundle is still the richer prize

AWS has resisted selling its AI chips directly for a reason. The chip itself is only one part of the economics.

When customers run AI workloads on AWS, Amazon can charge not just for compute, but also for storage, security, networking, and monitoring services. TechCrunch describes this as a “waterfall effect.” The chip creates the first transaction. The surrounding cloud stack creates the rest.

That makes external chip sales both tempting and awkward.

A direct hardware sale could unlock buyers that do not want to run inside AWS. Yet it could also reduce the pull of AWS as the default place to use Amazon silicon. Amazon would gain a chip customer while potentially missing the broader cloud revenue that comes from hosting the workload.

This is the strategic gap beneath the headline. Nvidia sells into other people’s data centers by design. AWS built Trainium to make its own data centers more attractive. Turning that internal advantage into an external product changes the incentive structure.

For the AI application layer that ultimately consumes this infrastructure, see XOOMAR’s coverage of Amazon Crowds Into Odyssey World Models in $1.45B Race. The spending logic is connected: models need compute, and compute scarcity shapes who can scale.

Nvidia faces a new kind of rival, not a mirror image of itself

Amazon would not be a conventional Nvidia challenger. It is not starting from chip sales and moving into cloud. It is starting from cloud demand and moving outward into chips.

That distinction matters. AWS already knows where AI workloads strain capacity because it operates the infrastructure directly. It also has a captive internal buyer for Trainium: AWS itself. Most chip companies need external buyers first. Amazon can absorb chips internally before deciding how many, if any, to sell outside.

But Nvidia’s position remains powerful in the facts supplied. TechCrunch frames Amazon’s potential $50 billion opportunity against Nvidia’s $326 billion revenue run rate, and also notes Nvidia CEO Jensen Huang recently described a new $200 billion market for Nvidia in selling CPUs for AI, not only GPUs.

So both companies are moving across boundaries. Nvidia is pushing further into CPU territory. Amazon is pushing further into Nvidia’s AI chip territory.

AWS spokesperson Doron Aronson confirmed the door is open, while keeping the language careful:

"While we’ve historically declined requests to sell chips directly, Andy noted it’s quite possible we’ll sell racks of them to third parties in the future."

That is not a product launch. It is a warning shot.

Amazon's chip challenge will rise or stall on supply, buyers, and trust

The next evidence will matter more than the ambition. Amazon needs named buyers, clearer delivery timelines, and proof that it can expand Trainium supply without starving AWS customers who already face sold-out capacity.

The strongest version of the thesis is selective, not total. XOOMAR analysis: Amazon AI chips could win workloads where customers value cost, availability, and AWS integration more than access to Nvidia’s dominant hardware path. That would still be meaningful. A partial win in AI infrastructure can be worth tens of billions if Jassy’s run-rate framing holds.

The weakest version is also clear. If talks stay unnamed, if Trainium4 capacity remains locked up before launch, or if TSMC supply cannot stretch, Amazon’s external chip business may remain mostly theoretical.

The practical watch list is short:

• Named buyers: Third-party data centers would validate demand outside AWS.
• Capacity evidence: Amazon must show it can supply outsiders without deepening AWS waitlists.
• Trainium4 timing: Sold-out future capacity makes execution harder, not easier.
• Revenue treatment: Investors will need to see whether chip sales add profit or merely shift value away from AWS services.
• Nvidia response: The key question is whether Nvidia treats Amazon as another cloud customer with custom ambitions, or as a direct hardware rival.

Amazon does not need to replace Nvidia to change the AI chip market. It only needs to prove that AWS silicon can travel outside AWS. That is the line to watch.

The Bottom Line

• Amazon selling Trainium directly would move AWS closer to Nvidia’s core chip business.
• Large AI customers could gain another option for training and inference workloads beyond Nvidia GPUs.
• The talks are still early, so this is a strategic signal rather than a confirmed merchant chip push.