One number reframes the entire AI chip landscape: $50 billion. That’s the annual revenue run rate Andy Jassy says Amazon’s chip division would command if it sold silicon externally like a traditional chip vendor. Right now, that division — Trainium AI chips, Graviton CPUs, Nitro offload chips — clocks in at $20 billion, growing triple digits year over year. Every dollar stays locked inside AWS pricing. Customers rent compute time; they don’t buy hardware. Jassy’s shareholder letter just cracked that door open, writing that “it’s quite possible we’ll sell racks of them to third parties in the future.”

For anyone watching AI infrastructure costs or holding Nvidia stock, that sentence lands heavy.

The Numbers Behind the Signal

The revenue math in Jassy’s shareholder letter reveals a chip business hiding in plain sight inside AWS.

AWS AI revenue runs at over $15 billion as of Q1 2026 — up from essentially zero three years ago. The demand signals are almost absurd in scale:

• Chip division run rate: $20B+ currently, projected at ~$50B with external sales
• Trainium2: fully subscribed; Trainium3: nearly fully subscribed before shipping
• Two large customers tried to buy all of Amazon’s Graviton capacity for 2026 — AWS declined
• Amazon plans $200 billion in capex for 2026, mostly AI infrastructure
• AWS added 3.9 gigawatts of capacity in 2025, with plans to double that by 2027

That Graviton anecdote deserves a second read. Two customers wanted the entire supply. AWS said no. Demand is moving faster than a sold-out stadium tour — except the venue keeps expanding and still can’t keep up. AWS AI chief Peter DeSantis confirmed to TechCrunch that talks with potential external Trainium buyers are already underway.

Can Trainium Actually Compete with Nvidia?

Economics, not raw performance, are Trainium’s sharpest competitive edge — but the gap on speed remains real.

“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.” — Andy Jassy, annual shareholder letter. Trainium’s pitch has always been about economics, not raw horsepower. Lead architect Ron Diamant told CNBC that Trainium delivers 30–40% better price-performance than competing hardware inside AWS. Trainium3, built on TSMC’s 3nm process with 144GB of HBM3e memory and nearly 5TB/s bandwidth, claims 50% lower cost versus equivalent Nvidia GPU setups. One important caveat tempers that case: Business Insider reported that startups including Cohere found earlier Trainium generations lag Nvidia’s H100 on latency. Price wins; raw speed doesn’t always follow.

For enterprise AI buyers operating at scale, though, this fight isn’t about benchmark bragging rights. Jassy projects Trainium will “save us tens of billions of capex dollars per year, and provide several hundred basis points of operating margin advantage versus relying on others’ chips for inference.”

If Amazon scales external chip sales, every enterprise AI buyer suddenly holds a credible second option. That reshapes Nvidia’s pricing leverage overnight — because the company offering you cheaper silicon also happens to run the cloud underneath half the internet.