AWS ponders selling its home‑grown chips by the rack‑load and is close to selling out AI capacity

Chips on the table

Amazon CEO Andy Jassy’s annual letter sketches a future in which AWS’s in‑house silicon is worth big money. He claimed that “if our chips business was a stand‑alone business… our annual run rate would be ~$50 billion,” and suggested “it’s quite possible we’ll sell racks of them to third parties in the future.” It has been reported that those are not idle boasts — the company already earns about $20 billion from services running on its home‑grown chips, and Jassy framed Trainium and Graviton as core competitive advantages that could tilt the cloud CPU game further toward hyperscalers who design their own silicon.

Demand outstrips supply

Capacity is the drama here. It has been reported that two large AWS customers asked to buy all of the company’s Graviton instance capacity for 2026 — requests that Amazon apparently declined. Trainium demand is also intense: Trainium3 capacity is allegedly “nearly fully‑subscribed,” and Jassy says a significant chunk of Trainium4 has already been reserved even though broad availability is about 18 months away. He argues the home‑designed chips will save “tens of billions” in capex and add several hundred basis points of operating‑margin advantage versus buying outside chips. High stakes. High tension.

Power, drones and robots — all part of the push

Jassy didn’t stop at silicon. AWS added 3.9GW of power in 2025 and plans to double power capacity by the end of 2027 because electricity — not servers — is the bottleneck. The letter also teases a string of other ambitions: a 200‑satellite broadband service slated for mid‑2026, Prime Air drone deliveries targeting 30 million customers, and “over one million robots” already in fulfillment centers. Ambition is not the problem. Capacity is. The question now: will AWS keep hoarding its chips to run its empire, or start selling racks and reshape an already bruising cloud market?

Sources: The Register