Jassy: AWS 2026 capacity is sold out, and 2027 is "largely sold out"

Amazon CEO Andy Jassy said on the a16z podcast released October 23 that AWS is running out of room fast: 2026 capacity is “at this point, sold out,” 2027 is “largely sold out,” and “a good chunk” of 2028 is already booked. The episode, hosted by a16z cofounder Ben Horowitz and Erik Torenberg, is titled “Amazon CEO Andy Jassy on the $220 Billion AI Buildout.” Reuters picked the remarks up the same day at 23:07 GMT under the headline “Amazon cloud capacity sold out through 2027,” while The Information went with “AWS Largely Sold Out Of Capacity Into 2028.” All three headlines compress the same quote — and none of them compresses it quite the way Jassy said it.

Key points

• The quote is a three-step qualification, not a blanket sellout. 2026 is “obviously, at this point, sold out”; 2027 is “largely sold out,” because AWS deliberately holds back capacity for on-demand customers and its Bedrock inference service; for 2028 it’s “a good chunk of ‘28 is sold out.” “Largely” is not “entirely,” and 2028 is a fraction, not a year.
• The shortage now has a stated end date: none. Jassy said “I don’t think that that shortage ends in 2027,” pinning it on three bottlenecks: “There’s only so much data center and power today. There’s only so much memory.”
• The $220 billion is context, not news. The roughly $220 billion 2026 capex figure dates from the July 30 earnings call; the episode notes add a new claim — demand could eventually make AWS “a $1 trillion annual run-rate business.”

What Jassy actually said

The line-by-line account below follows the verification pass published by shattered.io on October 24. The fullest version of the quote: “2027 is largely sold out. We always leave a certain amount for on-demand and our inference service Bedrock, or EC2… largely sold out, all the big chunks. A good chunk of ‘28 is sold out.” A shorter version circulating on social media pairs the two years: “2026 is obviously, at this point, sold out. ‘27 is largely sold out.”

Each qualifier carries weight. “Sold out” applies to 2026’s committed capacity. “Largely” applies to 2027, and Jassy explained why himself: on-demand customers and inference workloads on Bedrock and EC2 need capacity available on short notice, so that pool never gets pre-sold. “A good chunk” for 2028 is a conversational fraction, not a financial disclosure — Amazon offered no number for how much 2028 capacity is under contract.

Headline math

Outlets split the same sentence three ways. Reuters stopped at 2027: “sold out through 2027.” The Information stretched to 2028: “Largely Sold Out Of Capacity Into 2028.” Some Chinese outlets ran “capacity sold out through 2027, most of 2028 snapped up,” and at least one wire summary dropped “largely” altogether. In Jassy’s actual words there is no unqualified “sold out” covering all of 2027, and 2028 is only “a good chunk.” Anyone making procurement or budget calls off these headlines should read the quote itself — it is more conservative than any of the headlines.

The numbers behind it

None of this is the first signal. On the July 30 earnings call Jassy already said Amazon would spend roughly $220 billion of capex in 2026 and still lack the capacity to meet all demand, that 2027 would be similarly tight, and that 2028 demand was already “striking.” We covered the CFO’s side of that in AWS CFO: AI server payback under three years — the Q2 10-Q shows about $496 billion in unrecognized multi-year commitments with a 6.4-year weighted-average remaining life, meaning customers had already been locking in capacity with long contracts. The podcast extends the same curve into 2028.

Memory, one of the bottlenecks Jassy named, has already pushed Amazon’s spending forecast up once (the RAM shortage runs through 2028), and foundry relief is not expected before 2027 (TSMC’s record Q3). One arithmetic note: shattered.io converts $220 billion into roughly $600 million of spending per day and flags it as “our own arithmetic, not an Amazon statistic” — quote it with that caveat or not at all.

What others say

Coverage landed across 24 hours from Reuters, The Information, India Today, Livemint, CNA (via UA.NEWS syndication) and Korea’s Maeil Business Newspaper; on the Chinese side, Wallstreetcn, East Money, BlockBeats and Sohu all picked it up. Most English reports kept the “largely” qualifier; some headlines compressed it away. As of publication, no outlet has disputed the sellout claim, and none can independently verify how much of 2028 is actually gone — that figure exists only as Jassy’s spoken words, with no contract or filing behind it.

Timing is worth noting. The episode dropped days before Amazon’s Q3 earnings on October 29, and the show notes carry the “one trillion-dollar run-rate” framing. Telling the market that capacity is booked out and demand is spilling over is, among other things, a way to set expectations before the print. Read the two together.

Our take

For developers doing capacity planning on AWS, the practical read is more concrete than the headlines: if “the big chunks” of 2027 are spoken for, the on-demand pool only gets tighter, and spot price volatility will likely widen. Teams that need large, stable GPU allocations have two moves — negotiate multi-year commitments now, or design workloads to fail over to another cloud or their own hardware. Jassy’s counterpoint is that inference costs keep falling, which grows demand; both ends pressing at once makes it hard to see the capacity gap closing inside 2027.

Two doubts to keep. First, “largely sold out” and “a good chunk” carry no numbers — they are a CEO’s podcast phrasing, not a disclosure. Second, the claim serves a purpose: earnings land on October 29. What is checkable today is thin and real — the episode exists, dated October 23, and the $220 billion capex figure is from the July call. The rest waits for the October 29 earnings and the next 10-Q.