Jassy: AWS Sold Out Through 2027, 2028 Going Fast

By Daniel Rao

Photo: Andy Jassy, chief executive of Amazon. Credit: Lisi Mezistrano Wolf, via Wikimedia Commons (CC BY-SA 4.0).

There is a particular kind of corporate confession that only sounds like good news. Amazon chief executive Andy Jassy delivered one this week, on a venture capital podcast of all places, and it managed to be both a boast and a warning in the same breath: the world's largest cloud provider has, for practical purposes, run out of cloud to sell. Not in some distant, theoretical sense. Through 2027. With a good chunk of 2028 already spoken for too. If you wanted a single sentence to explain why the artificial intelligence boom feels simultaneously unstoppable and strangely fragile, Jassy just gave it to you, free of charge, on somebody else's show.

The admission came on an Andreessen Horowitz video podcast hosted by Ben Horowitz and Erik Torenberg and released on October 23. Jassy was not announcing a product or apologising for an outage. He was describing, in the flat tone of a man reading out a warehouse inventory, a demand picture so extreme that even Amazon's famously enormous building programme cannot keep pace with it. "2027 is largely sold out. We always leave a certain amount for on demand and our inference service bedrock, or EC2, but largely sold out all the big chunks. A good chunk of 2028 is sold out." Read that again slowly. The big chunks are gone. What remains is the retail shelf: on-demand capacity, kept back deliberately, for customers who turn up without a reservation.

It is worth pausing on how unusual this is. Cloud computing was sold to the world, for the better part of two decades, as infinite. Need more servers? They are already there. That was the pitch, and for most of that time it was roughly true. Jassy has now told us, in public, that the pitch has expired — at least for the customers who matter most, the AI laboratories and the large enterprises signing multi-year deals. Amazon reports its third-quarter results on October 29, and you can be certain that every analyst on that call will now be asking the same question: if demand is this strong, why is supply still the story?

The sold-out admission

Jassy's words matter because of who he is and where he said them. This was not a line buried in an earnings call, hedged by lawyers and wrapped in safe-harbour language. It was a conversation, on a podcast run by one of Silicon Valley's most prominent investment firms, with hosts who know the cloud market well enough to recognise a startling claim when they hear one. Jassy did not flinch. He described a capacity book that is, in his own word, largely sold out for a full calendar year that has not yet begun, with significant commitments already reaching into the year after that.

The timing is doing a lot of work here. Amazon had already told investors earlier this year that it would spend about $220 billion in capital expenditure for 2026, a figure so large it dwarfs the annual budgets of most countries' governments, let alone most companies. And still, Jassy said, the company "could invest a lot more and it wouldn't be a bad use of capital". That is not the language of a chief executive who thinks his spending problem is solved. It is the language of a man standing in front of a queue that stretches around the block, holding a wallet that is already open, and still not being able to serve everyone.

We checked Reuters' account of the podcast against the relayed excerpts and Amazon's reported filing figures, and the picture holds together: the sold-out claim, the $220 billion spending guide, and the October 29 results date all line up. What Reuters did not do — and what nobody on the podcast pressed Jassy very hard on — is explain what "sold out" does to the customers who are not already inside the tent. That is where this story gets less comfortable, because a sold-out cloud is not a neutral fact. It is a rationing decision, and rationing always has winners.

What "sold out" actually means

First, a necessary piece of pedantry, because Jassy's phrasing invites misreading. "Sold out" does not mean that a startup founder cannot spin up a server on Amazon Web Services tomorrow morning. Jassy was explicit that Amazon keeps capacity back for on-demand use and for Bedrock, its inference service. If you want a modest amount of compute, right now, you can almost certainly have it. What is sold out is the other thing: the multi-megawatt, multi-year reservations that AI laboratories and large enterprises sign when they are planning training runs and inference fleets at scale.

Think of it less like a hotel with no rooms and more like a housing market in which every large family home has been bought off-plan, while a handful of studios remain for nightly letting. The distinction matters because the economics of the two markets are entirely different. On-demand capacity is profitable but lumpy. Reserved capacity is what lets Amazon plan power purchases, chip orders and construction schedules years ahead. When Jassy says the big chunks are gone through 2027, he is saying that the planning horizon for the most valuable customers is now fully booked — and that anyone arriving late will pay more, wait longer, or build elsewhere.

There is a second, quieter implication. A cloud provider that is sold out does not need to discount. It does not need to chase. In a normal market, scarcity gives the seller pricing power, and AWS has never been shy about using pricing power. The question for customers is whether "sold out" becomes a polite way of saying "more expensive", and whether smaller AI companies — the ones without multi-year contracts — find themselves squeezed into the on-demand shelf while the giants eat at the reserved table. Jassy did not say that, of course. He did not need to. The arithmetic says it for him.

The $220 billion answer — and why it may not be enough

Let us put the $220 billion in context, because numbers of that size stop meaning anything unless you anchor them. Amazon's reported filing figures, relayed via EconoTimes, show AWS sales of $42.2 billion in the second quarter, up 36.7% year on year — a $169 billion annualised run rate. The AI business and the chips business are each above a $25 billion annual run rate. Cash capital expenditure in the first half of the year reached $96.3 billion, including $53.1 billion in the second quarter alone. Trailing twelve-month free cash flow, meanwhile, turned negative $7.6 billion, against positive $18.2 billion a year earlier, attributed mainly to property and equipment purchases for AI. Read those figures together and the story writes itself: Amazon is spending at a pace that would terrify most boards, and demand is still outrunning it.

This is the part of the story that should unsettle even the optimists. Capital expenditure is supposed to be the answer to scarcity. You spend, you build, scarcity eases. But data centres are not software. You cannot ship one overnight. You need land, permits, grid connections, transformers, cooling systems and chips, in that order, and every one of those links has its own queue. Power, above all, is the binding constraint: you can order all the servers you like, but if the local grid cannot deliver the megawatts, you own a very expensive, very quiet building. It is the same constraint, in a different accent, as the one behind the news that Microsoft walked out of Nscale's Essex data centre after a power row. The industry's bottleneck has moved, over the last three years, from chips to electricity, and electricity does not scale like software.

That pressure is one reason behind Amazon's reported plan to offload $8 billion of Nvidia chips into an investor vehicle, a structure that would let outside capital carry some of the balance-sheet weight of the buildout while Amazon keeps operating the capacity. Whether that deal happens exactly as reported matters less than what it signals: even a company generating AWS revenue at a $169 billion annualised pace is looking for ways to fund the next phase without carrying every dollar itself. When the richest cloud provider on earth starts sharing the bill, the bill is genuinely enormous.

Trainium and Annapurna: the cost weapon

Jassy's other big claim on the podcast was about cost, and it is arguably the more important one for Amazon's long-term position. He described the acquisition of Annapurna Labs, the Israeli chip designer behind Amazon's Trainium and Graviton processors, as "possibly the best deal in Amazon's history". That is a striking thing for the chief executive of Amazon to say about any deal, given the company's history includes the acquisitions that built its advertising and logistics empires. But the logic is straightforward: if you cannot buy enough of somebody else's chips, and you cannot build data centres fast enough, the next best thing is to make your own chips cheaper than anyone else's.

Jassy claimed Trainium offers 30–40% better price-performance than comparable processors — his claim, not an independent benchmark, and worth labelling as such. The company also points to more than $225 billion in future Trainium revenue commitments, including from Anthropic and OpenAI, as reported. If even a large fraction of that holds, it changes the shape of the AI infrastructure market: the most important customer of Amazon's chips would be Amazon itself, and the most important supplier to the AI laboratories would increasingly be a company that also competes with them in models, through Bedrock and its own work. That vertical integration is exactly what makes rivals nervous, and exactly what makes Amazon's cost structure so hard to match.

Photo: A server room, the kind of capacity now booked out through 2027. Credit: BalticServers.com, via Wikimedia Commons (CC BY-SA 3.0).

It was also the same week AWS showed it could park 75,000 idle agents on a Raspberry Pi, a stunt that looked like a party trick and was actually a flex about efficiency: if your agents are cheap enough to idle by the tens of thousands on a single-board computer, your cost floor is lower than your competitors'. Put the two stories together — custom silicon on one side, ruthless agent efficiency on the other — and Jassy's cost weapon comes into focus. The sold-out headline is about demand. The Trainium story is about who profits from it.

The memory wall: the shortage Jassy says won't end in 2027

Jassy's warning goes beyond data centres and power. Covering the same podcast, Reuters reported him saying: "There's only so much data center and power today. There's only so much memory, and I don't think that shortage ends in 2027." The bottleneck, in other words, is not just concrete and megawatts — it is the high-bandwidth memory inside the accelerators, and the DRAM the rest of Amazon's cloud still runs on.

That is also why the Trainium claim matters: Jassy put its price-performance advantage at 30% to 40% over comparable processors, which is the difference between inference that is merely possible and inference that is cheap enough to sell at scale. If memory stays tight into 2027, the sold-out picture does not ease just because Amazon pours more concrete.

The barbell: labs at one end, enterprises at the other

Perhaps the most revealing part of Jassy's podcast appearance was not a number at all but a shape: a barbell. Demand, he suggested, is heavy at both ends and thin in the middle. At one end sit the AI laboratories, spending at a pace that would have seemed delusional five years ago. At the other sit large enterprises, finally moving from experiments to production systems. In the middle, underserved and overlooked, sits mid-market software — companies that want AI capability but cannot yet justify the cost, the complexity or the commitment.

Jassy's explanation for the missing middle was blunt: inference costs are, in his view, "an order of magnitude too high" for broad enterprise adoption. That is a remarkable admission from a man whose company sells inference by the token. It is also, if you think about it, a statement of strategy. If costs fall by a factor of ten — through cheaper chips, better models, more efficient agents — the middle of the barbell fills in, and the market gets much bigger than it is today. Amazon is betting, with Trainium and with Bedrock, that it will be the company that delivers that cost collapse, and that it will capture the upside when it happens.

There is a human detail in the podcast worth noting, because it captures Jassy's style. He did not deliver the sold-out line as a dramatic reveal, with a pause for effect. He delivered it as inventory — flat, factual, almost bored, as if reporting that the warehouse was empty were the most ordinary thing in the world. That flatness is itself a signal. When a chief executive treats a sold-out year as unremarkable, it tells you how thoroughly the extraordinary has become the baseline inside Amazon. The boom is no longer an event to Jassy. It is the weather.

The sceptical read: booked is not banked

Now for the scepticism, because a sold-out sign is not the same thing as a bank balance. Amazon's reported filing figures include roughly $496 billion in unrecognised performance obligations, with a 6.4-year weighted-average remaining life. That is an enormous backlog, but it is not a map of capacity by year, and it is not cash in hand. Customers commit, then renegotiate. Laboratories sign, then merge, then pivot. Multi-year AI commitments are, by nature, bets on a technology that is changing monthly, and some of those bets will be unwound. Anyone who treats the entire backlog as guaranteed revenue is making the same mistake as someone who treats a restaurant's reservation book as its profit-and-loss account.

The second sceptical point is power. A data centre that cannot get a grid connection is not capacity, whatever the contract says. Across the United States, utilities are quoting connection timelines measured in years, not months, and local opposition to new facilities is hardening — a subject we return to elsewhere in this publication. Amazon can sign all the customers it likes; if the megawatts do not arrive, the revenue does not either. Jassy's "sold out" claim is, in that light, partly a statement about demand and partly a bet about delivery. The demand part is verifiable. The delivery part will be tested, quarter by quarter, in the filing figures.

The third point is the simplest and the most easily forgotten: free cash flow is negative. Amazon is funding this buildout by spending more cash than its operations currently throw off, on a trailing twelve-month basis. That is defensible when you believe demand will persist — and Jassy clearly does — but it means the company has less room for error than the headline numbers suggest. If AI demand cools, or if inference prices collapse faster than volumes grow, the company will own a very large, very fixed cost base. The sold-out sign would then read less like a triumph and more like a warning that was hiding in plain sight.

Our take: a boom, and a bottleneck wearing a boom's clothes

So which is it: proof that the AI boom is real, or evidence that it is hitting its limits? Our position is that it is both, and that pretending otherwise is how investors get hurt. Jassy's admission proves demand in a way that no analyst forecast ever could — customers do not sign multi-year capacity deals for a fad. But it also proves that the infrastructure of the boom is now the boom's binding constraint. When the world's largest cloud provider cannot sell you a large reservation for a year that has not started yet, the story is no longer "AI is growing fast". The story is "AI is growing faster than the physical world can accommodate", and the physical world does not care about your roadmap.

That distinction will decide the next two years. If power, chips and construction keep pace, the sold-out book becomes revenue, Trainium's cost advantage compounds, and the barbell's missing middle fills in as inference gets cheaper. If they do not, the sold-out book becomes a queue of disappointed customers, a negotiating weapon for rivals with spare capacity, and a political problem in every community asked to host a data centre it did not ask for. Jassy, to his credit, has put the fact on the record before either outcome is decided. The least the rest of us can do is take him at his word — and then watch, sceptically, whether the megawatts follow the money.

Frequently asked questions

Is AWS really sold out?

For large, reserved capacity: largely yes, through 2027, with a good chunk of 2028 already committed, according to Jassy. On-demand capacity and Bedrock inference capacity are still held back for customers who need smaller amounts without a long-term reservation.

What does Amazon plan to spend on this buildout?

Amazon guided about $220 billion in capital expenditure for 2026. Jassy said the company could invest considerably more and that it would not be a bad use of capital, which suggests the spending programme is demand-led rather than capped by an arbitrary budget.

Why is Trainium important?

Trainium is Amazon's own AI training chip, designed by Annapurna Labs. Jassy claims 30–40% better price-performance than comparable processors, and Amazon reports more than $225 billion in future Trainium revenue commitments. If those claims hold, custom silicon is how Amazon keeps costs down while rivals pay full price for merchant chips.

Does "sold out" mean prices will rise?

Not automatically, but scarcity gives the seller leverage, and AWS has historically used its scale to set terms. Smaller customers without reservations are the most exposed: they compete for the on-demand shelf while the largest buyers lock in multi-year deals.

When will we know if Jassy is right?

Amazon reports third-quarter results on October 29, and its AWS growth, capital expenditure and backlog figures will be the first hard test. The longer test is physical: whether new data centres secure power and come online on schedule through 2026 and 2027.

Sources: Reuters' October 23, 2026 account of Andy Jassy's Andreessen Horowitz podcast appearance; relayed excerpts from the podcast hosted by Ben Horowitz and Erik Torenberg; Amazon's reported filing figures as relayed by EconoTimes (second-quarter AWS sales, capital expenditure and performance obligations).

Daniel Rao is a technology journalist covering cloud computing, artificial intelligence and the business of the internet.