Here's a number worth sitting with for a second: $50 billion. That's the annualized revenue run rate AWS chief Andy Jassy claims his company's homegrown AI chips would generate if Amazon sold them on the open market the way a real chip company does. For context, that's roughly Intel's entire annual revenue. Not bad for a cloud provider that technically isn't supposed to be in the chip business.

AWS AI chief Peter DeSantis recently told Bloomberg that Amazon is in active talks to sell its Trainium chips directly to other companies for use in their own data centers. He didn't name names—of course he didn't—but the intent is clear. Amazon is no longer content keeping its silicon locked behind an AWS console.

From Internal Tool to External Threat

To understand why this is a big deal, you need to understand how Amazon has traditionally thought about its chips. Trainium wasn't built to be a product. It was built to be infrastructure—a way to process AI workloads at lower cost than renting Nvidia GPUs, while also locking customers deeper into the AWS ecosystem. Every token you process on Trainium isn't just a compute sale. It's a gateway to S3 storage, VPC networking, CloudWatch monitoring, IAM security policies, and a dozen other AWS services that quietly inflate your monthly bill. The chip was the loss leader; the ecosystem was the margin.

Selling chips externally breaks that model. Or at least complicates it significantly. When a competing data center operator buys Trainium racks and runs them in their own facility, Amazon collects a one-time hardware margin instead of the compounding revenue waterfall of cloud services. That's not obviously a better business. So what changed?

Demand Is the Answer—And Also the Problem

Jassy's April shareholder letter dropped a fairly jaw-dropping claim: Trainium capacity sells out almost immediately after it becomes available. Even more remarkable, capacity for Trainium4—a chip that won't ship for over a year—was already spoken for. That was written before AWS formally brought OpenAI onto its platform as a customer, which presumably didn't reduce demand.

So the chip is clearly doing something right. The problem is that "selling out instantly" is not actually a good supply chain story—it's a capacity constraint story. If Amazon wants to sell chips to third parties without abandoning its existing cloud customers, it needs to manufacture more chips than it currently can. And that requires TSMC.

Here's where the geopolitics of silicon get inconvenient. Nvidia recently dethroned Apple as TSMC's largest customer. Nvidia—the company Amazon is ostensibly trying to challenge—is already first in line at the world's most advanced foundry. Getting enough TSMC capacity to supply both internal AWS demand and external chip sales is not a scheduling problem you solve with a strongly worded email to Hsinchu.

How Big a Threat Is This to Nvidia, Really?

Let's not oversell it. Nvidia is currently running at a $326 billion annual revenue pace. A $50 billion Amazon chip business—even if it materialized overnight, which it won't—isn't going to topple that. What it does do is establish a credible second-tier market for AI accelerators that isn't just AMD making hopeful noises about MI300X performance numbers.

Amazon has something AMD and Intel don't: actual proof of workload. Trainium has been battle-tested at hyperscaler scale, not just in benchmark suites carefully curated for press releases. Anthropic runs on it. OpenAI is now on AWS. Apple has reportedly evaluated it. These aren't hobbyist endorsements—they're the most demanding AI customers on the planet giving the chip a passing grade.

The other thing worth noting is the competitive chess being played here. Jensen Huang recently announced Nvidia is moving into CPU territory for AI infrastructure—traditionally Intel and AMD's turf. Jassy is essentially mirroring that move in reverse: a cloud provider stepping onto chip vendor territory. Everyone is eating everyone else's lunch simultaneously, which is either a sign of a healthy competitive market or a sign that the AI infrastructure gold rush has everyone acting a little unhinged. Possibly both.

The Actual Limiting Factors Nobody's Talking About

Here's what the press release version of this story skips: selling chips is genuinely hard. It's not just manufacturing—it's supply chain management, hardware support contracts, firmware updates, driver compatibility across a thousand different server configurations, and the kind of enterprise sales infrastructure Amazon has never had to build because it sold access to hardware, not the hardware itself. Nvidia has decades of that muscle. Amazon is starting from scratch.

There's also the software stack question. Trainium runs on AWS Neuron, a compiler and runtime framework that, while functional, is not CUDA. The AI world runs on CUDA. Pytorch, JAX, most production inference frameworks—they all have CUDA as the happy path. Neuron support exists and has improved, but switching costs are real, and enterprise infrastructure teams do not enjoy rewiring their ML pipelines on a tight deadline.

None of this means Amazon can't pull it off. It means the gap between "Andy Jassy wrote something interesting in a shareholder letter" and "Trainium racks are shipping to data centers outside AWS" is wider than the headlines suggest. Amazon has closed harder gaps before. But let's watch for actual purchase agreements before we start updating the competitive landscape charts.

For now, the most honest summary is this: AWS has built a chip worth selling, the demand signal is real, and the strategic intent is clearly there. The execution, as usual, is where things get complicated.