On April 9, 2026, Amazon CEO Andy Jassy published his annual letter to shareholders, and the number infrastructure people zeroed in on was not a marketing metric: the combined annual revenue run rate of Graviton, Trainium, and Nitro — Amazon’s three custom silicon lines — has passed $20 billion, growing at triple-digit percentages year over year. Jassy immediately added a caveat: that figure is an undercount, because those chips are currently monetized only through EC2. Priced the way other chipmakers sell, to AWS and outside buyers alike, the run rate would be roughly $50 billion.
The letter repositions Amazon from “the cloud giant that buys enormous quantities from NVIDIA” to something closer to a silicon power in its own right, and for the first time floats the idea that Amazon might “sell racks of them to third parties in the future.” Amazon shares rose more than 5% on the day the letter landed.
$20 Billion Is an Undercount
Accounting treatment is the key. These chips never get their own invoice — all of the revenue hides inside EC2 pricing — so the $20 billion run rate only reflects how much Amazon consumes itself. The demand-side signal is even more blunt: Jassy disclosed that two large customers asked to buy all of Amazon’s 2026 Graviton capacity, and Amazon declined. The backdrop is AWS’s own trajectory — 24% year-over-year growth in Q4 2025 at a roughly $142 billion annualized run rate, with AWS AI services passing a $15 billion annualized run rate in Q1 2026.
In other words, custom silicon has graduated from “an internal cost lever” to a product line with an independent valuation. Amazon has simply chosen, so far, not to sell it externally.
The Trainium Roadmap: Sold Out Through the Generations
The letter is unusually concrete about the generational cadence:
- Trainium2 delivered roughly 30% better price-performance than comparable GPUs and is largely sold out
- Trainium3, shipping since early 2026, is 30–40% better than Trainium2 and nearly fully subscribed
- Trainium4 is about 18 months out, and part of its capacity is already reserved
The economics justify the pace. Jassy estimates that at scale, Trainium saves “tens of billions” of capex dollars per year and carries a several-hundred-basis-point operating margin advantage over relying on other vendors’ chips. In an era of $200-billion-class AI capex budgets, that return is calculable, not speculative.
Defending $200 Billion of Capex
The most-quoted line of the letter: “We’re not investing approximately $200 billion in capex in 2026 on a hunch.” Jassy’s argument is that demand is already contracted — the over-$100-billion OpenAI commitment plus other customer agreements cover a substantial portion of 2026 spend, with monetization landing mostly in 2027 and 2028.
The cost is stated just as plainly: free cash flow fell from $38 billion to $11 billion, driven by a $50.7 billion year-over-year increase in property and equipment purchases. On the capacity side, AWS added 3.9 gigawatts of power capacity in 2025 and plans to double total capacity by the end of 2027. Jassy’s summary: “We’re not going to be conservative in how we play this.”
Selling Racks: Redrawing the Line Between Cloud and Chips
“It’s quite possible we’ll sell racks of them to third parties in the future” is the sentence with the most second-order implications. If it happens, Amazon competes with NVIDIA and AMD at the system level, and customers buy not an EC2 instance but a physical rack stuffed with Trainium. TechCrunch noted the letter also takes aim at NVIDIA, Intel, and Starlink by name. Set against Anthropic’s multi-gigawatt TPU procurement deal with Google and Broadcom announced the same week, the compute market is polarizing in both directions at once: mega-buyers locking up external capacity on one end, and a cloud vendor that might start selling hardware directly on the other.
What It Means for Developers and Product Teams
Three practical effects. First, expect the price-performance gap on EC2 to keep widening; the Trainium family’s leverage is most direct on inference costs, and it belongs in your benchmark matrix. Second, capacity scarcity is now the norm — two consecutive generations sold out and generation four already partially reserved means teams with large-scale needs should plan ahead rather than assume capacity on demand. Third, if racks do go on sale externally, multi-vendor architectures gain one more candidate — but it also means supply-chain risk math now includes a supplier that both sells chips and runs a competing cloud.
Sources
- CEO Andy Jassy’s 2025 Letter to Shareholders — Amazon
- Amazon CEO defends AI spend — CNBC
- Amazon CEO takes aim at Nvidia, Intel, Starlink — TechCrunch
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.