Amazon is the everything store, sure — but the machine that actually prints the money is AWS, the biggest cloud on Earth. And buried inside that cloud, Amazon has quietly built one of the largest chip businesses on the planet: a custom-silicon shop running at a $25 billion-plus annualized pace, growing triple digits, while Wall Street still mostly shrugs. This is the story of how the everything store became a chipmaker — and why the AI world is about to find out.
Here's the plain-English version. Amazon designs its own AI processors: Trainium, built for training and inference, and Inferentia, built for inference alone. They're engineered to undercut Nvidia on price-performance, and they've already won the two biggest converts in the industry. Anthropic has committed up to 5 gigawatts of Trainium compute. OpenAI has committed 2 gigawatts starting in 2027. Amazon is the only hyperscaler whose custom silicon runs the workloads of its fiercest AI rivals — the competitors are literally paying the house.
And here's the twist that makes this whole story hum. The AI cycle is shifting from training to inference — from building models to running them billions of times a day. That's exactly the arena where Trainium's price-performance and latency win. AWS CEO Matt Garman puts it bluntly: Trainium3 is "the best inference platform in the world." Hyperbole? Maybe. But the numbers behind it are very real.
Trainium3 hit general availability in December 2025 on a 3-nanometer process, packing 4x the compute of Trainium2 with up to 40% better price-performance. Amazon now has roughly 1.4 million Trainium chips deployed across three generations. Anthropic's Claude runs on more than a million Trainium2 chips, and Project Rainier — a cluster of 500,000-plus chips — is the largest operational AI supercomputer on the planet. This isn't a pilot program. It's production infrastructure.
Inferentia handles the inference side of the house, and it's no slouch either: Inferentia2 delivers 4x the throughput with up to 10x lower latency. Most of the inference flowing through Amazon Bedrock — AWS's AI service layer — runs on Trainium silicon. Add Graviton, Amazon's homegrown server CPUs, and you've got a full stack: training, inference, and general-purpose compute, all designed in-house.
Now the part investors love: the economics. EC2 Trn2 instances deliver 30-40% better price-performance than comparable GPU instances, and comparable GPUs aren't cheap — Nvidia's Blackwell chips reportedly go for $30,000 to $40,000 each. When you're running at AWS scale, shaving that much off the cost of compute isn't a feature. It's a moat.
The customers keep coming, and the demand math is almost embarrassing. AWS's backlog has swelled to roughly $496 billion, and Amazon has raised 2026 capex to around $220 billion — up from $200 billion — after memory costs spiked. Andy Jassy says even that won't satisfy all of 2026's demand. A company spending $220 billion and still not keeping up? That's a demand problem most businesses would kill for.
Then there's the upside nobody has priced in yet: Amazon might sell these chips to the outside world. Jassy has floated the idea of a standalone silicon business worth roughly $50 billion a year. Imagine Amazon taking on Nvidia in the open market — its own fabless designs, its own supply chain, and the world's biggest cloud as a captive customer. That's a chip company hiding inside a retailer's valuation.
So here's why this matters to you. The market still prices Amazon like a retailer with a cloud hobby, while it quietly runs one of the biggest chip businesses on Earth — one that's winning the shift to inference before most investors even see it coming. The everything store built the AI pick-and-shovel play hiding in plain sight. Don't sleep on it.
Disclosure: The Signal holds no position in AMZN. Positions may change. This is not financial advice.
