Summary: In his April 9, 2026 shareholder letter, Andy Jassy said Amazon’s Graviton, Trainium, and Nitro chip lines had passed a $20 billion annual revenue run rate. He argued the number understates the business because revenue is folded into EC2 pricing and estimated about $50 billion under chipmaker-style sales. The letter also discussed about $200 billion of 2026 capex and possible future rack sales.
April 9, 2026 was the day Amazon CEO Andy Jassy’s annual letter to shareholders went out, and the number that carried the most weight was a revenue figure: Graviton, Trainium, and Nitro — Amazon’s three in-house chip lines — have together crossed a $20 billion annual revenue run rate, expanding at triple-digit percentage rates year over year. Right behind it came his caveat: the figure undersells the business, because these chips are monetized only through EC2 today. Priced instead the way other chipmakers price — selling to buyers inside AWS and outside it — the annual run rate would land at roughly $50 billion.
The letter rewrites Amazon’s positioning: it no longer reads as “the cloud giant that buys enormous quantities from NVIDIA” but as a silicon force of its own, and for the first time it floats the possibility that Amazon might “sell racks of them to third parties in the future.” The stock gained more than 5% the day the letter was published.
Why $20 Billion Undersells It
The catch is in how the revenue gets booked. These chips never appear on an invoice of their own — the money is folded into EC2 pricing — so the $20 billion run rate captures nothing beyond Amazon’s own usage. The demand picture is starker still: the letter recounts how two large customers tried to buy up Amazon’s entire 2026 Graviton capacity and were refused. AWS’s own numbers supply the context: 24% year-over-year growth in Q4 2025 on an annualized run rate of about $142 billion, plus AWS AI services clearing a $15 billion annualized run rate in Q1 2026.
Custom silicon, in other words, has outgrown its job as an internal cost lever and now rates as a product line with a valuation all its own — one Amazon has so far simply declined to sell externally.
Trainium, Generation by Generation: Nearly Nothing Left to Buy
Rarely does a shareholder letter get this concrete about cadence:
- Trainium2 beat comparable GPUs on price-performance by roughly 30% and is essentially sold out
- Trainium3 has been shipping since the start of 2026, runs 30–40% better than Trainium2, and has almost all of its capacity spoken for
- Trainium4 sits about 18 months out, yet a slice of its capacity is already reserved
There is arithmetic underneath the schedule: by Jassy’s estimate, Trainium at scale keeps “tens of billions” of dollars a year off the capex bill and adds several hundred basis points of operating margin advantage over leaning on other vendors’ chips. With AI capex budgets now sized in the $200-billion class, that is a return you can put a number on.
Making the Case for $200 Billion in Capex
The letter’s most-quoted sentence: “We’re not investing approximately $200 billion in capex in 2026 on a hunch.” The argument is that the demand is already under contract — an OpenAI commitment worth more than $100 billion, plus the rest of the customer agreements, covers a sizable share of 2026 spending, and the revenue mostly arrives across 2027 and 2028.
The bill comes with the same directness: free cash flow slid from $38 billion to $11 billion, mainly because property and equipment purchases rose $50.7 billion year over year. Supply is being built to match — AWS switched on 3.9 gigawatts of new power capacity in 2025 and expects to double its total capacity by the end of 2027. Jassy’s closing note on the topic: “We’re not going to be conservative in how we play this.”
Racks for Sale: Where Cloud Ends and Silicon Begins
Of everything in the letter, the line with the widest second-order reach is “It’s quite possible we’ll sell racks of them to third parties in the future.” Acted on, it would put Amazon across the table from NVIDIA and AMD at the system level; what a customer buys would be not an EC2 instance but a physical rack packed with Trainium. TechCrunch observed that the letter also calls out NVIDIA, Intel, and Starlink by name. Place it beside Anthropic’s TPU deal with Google and Broadcom, spanning multiple gigawatts from the same week, and the compute market’s pull toward both extremes is visible at once: mega-buyers tying up outside capacity on one end, a cloud provider weighing direct hardware sales on the other.
The Read for Developers and Product Teams
Three effects land in practice. One, the price-performance spread on EC2 should keep opening up; the Trainium family’s leverage bites hardest on inference costs, so it earns a place in your benchmark matrix. Two, scarcity is the new baseline — two straight generations sold through and part of the fourth already reserved mean teams with heavy capacity needs should line things up early rather than count on on-demand availability. Three, should racks actually go on sale outside, multi-vendor architectures pick up another candidate — and the supply-chain risk column gains an entry for a vendor that both sells chips and runs its own competing cloud.