AWS Sold Out 2027 and a Good Chunk of 2028. The Money Building the Rest Is Getting Harder to Raise.

The most useful sentence in technology this week came out of a podcast, and it was not about models. Andy Jassy told Andreessen Horowitz's Ben Horowitz that AWS cloud capacity is largely sold out through 2027, with significant commitments running into 2028, as Reuters reported. On the show he put it more bluntly: "26 is obviously sold out. 27 is largely sold out. A good chunk of 28 is sold out."

That evening, CNBC reported that CoreWeave, Nebius and IREN are each down well over 30% in October, that CoreWeave's five-year credit default swaps widened past 11%, a level credit desks treat as distress, and that SK Hynix had cancelled the IPO of its Solidigm storage unit, citing market volatility.

Both stories are about the same thing: computing capacity that does not exist yet, and who is paying to have it built. One side has the contracts, the other has the debt. Jassy's remarks land six days before Amazon reports third-quarter results, which is worth remembering; the mechanics he described, though, are checkable.

What sold out actually means

Amazon's 2026 capital budget is about $220 billion, or roughly $600 million a day. Jassy's argument is that even that leaves money on the table. He said the company could invest a lot more and it would not be a bad use of capital, and that he doubts it would satisfy all the demand even then.

AWS grew 36.7% year over year last quarter on a $169 billion annual run rate, by his account. The AI business inside it is over $25 billion in run rate after three years of the generative wave. AWS itself was a $60 million business at the same age. His word for the comparison is "several hundred times faster."

A sold-out year at AWS is not an empty shelf, though. He says the company always holds back a slice for on-demand, for Bedrock and for EC2. The tap does not shut off. What narrows is the margin of comfort around it, and the price of certainty.

The mechanics behind the claim explain why he can make it. Jassy walked through the cash cycle: data centres have to be committed about two years before servers go into them and start earning, and the buildings have a useful life of around 30 years. Servers and networking gear are bought only months ahead, break even in about three years, and last five to six. Most of the deals AWS signs run about five years.

That is a matched book. A five-year commitment against a five-to-six-year asset, with the option to skip the servers if demand fails to appear. It is a different risk profile from building on speculation, and it is why one balance sheet can sell its forward years while the credit market worries about another's.

The memory decision

The constraint he keeps returning to is memory. He said there is only so much of it, and that he does not think the shortage ends in 2027. Asked about the pressure to hand scarce capacity to labs willing to pay almost anything, he described the trade underneath: with a fixed amount of memory, someone decides how much goes to HBM for AI accelerators and how much goes to ordinary DRAM for the rest of the cloud.

That is not a procurement detail. It is an allocation decision made inside one vendor on behalf of every customer who depends on it, with no published rule. His stated principle is that AWS will not squeeze existing customers to make room for AI labs: "We're not going to squeeze our customers." He compared it to the pandemic, when AWS sent field teams to help customers cut bills knowing its own revenue would suffer. That is a promise, not a mechanism. The mechanism is memory supply, and he expects component shortages in some form for several years.

The middle of the barbell is empty

His description of demand explains where the sold-out years go. He calls it a barbell. At one end, AI labs consuming compute at a rate he says he has not seen, plus a handful of runaway applications. At the other, enterprises using AI for cost avoidance and productivity. In the middle sit the current production workloads of enterprises, the overwhelming majority of which, in his words, still are not using inference in any pervasive way.

He expects that middle to become the largest segment of AI usage. He also says why it has not moved: inference costs have fallen, but remain "an order of magnitude too expensive" for enterprises to justify rewriting CPU applications around accelerators. Amazon's answer is Trainium, which he claims delivers 30% to 40% better price-performance than comparable processors.

So the capacity sold through 2027 is largely sold to the two ends of the barbell. The middle, the part expected to be biggest, arrives on a clock set by inference economics rather than by data centre construction. Those clocks are not synchronised, and the gap is where the financing argument lives.

Capacity is becoming a product

Two days before the podcast, AWS shipped something that says more about this market than any forecast. Aurora and RDS customers can now buy On-Demand Capacity Reservations: pick a DB instance class, engine and Availability Zones, and AWS holds that capacity. Reserved capacity is billed at on-demand rates while the reservation is active, whether or not the instances inside it are running. The announcement says why, plainly: the goal is that create, restore, scale and start operations "don't fail because of insufficient capacity."

Capacity assurance is now a line item, and it has reached the database tier. A vendor does not open a reservations desk for a market with slack.

The money side of the same story

None of this makes the credit market wrong. It makes it a story about a different balance sheet.

CoreWeave is the clearest case. Its five-year CDS spread has been trading above 1,100 basis points, up 39% in October and 179% over a year, and its 2031 and 2032 unsecured notes have traded above 15% yields. It carries about $20.7 billion of debt maturing through 2030 and another $14.9 billion after that, against $5.5 billion of cash at the end of the second quarter. Analysts expect third-quarter negative free cash flow of $10.1 billion, according to FactSet. Baird's Rob Oliver estimates it needs roughly $30 billion in 2027.

Its defence is real: a backlog above $100 billion, around two-thirds with investment-grade counterparties, and more than two-thirds of last year's revenue from Microsoft. Goldman Sachs upgraded the stock this month, arguing that "not all neoclouds are created equal." The dispute is not whether the contracts exist. It is whether the cash arrives before the refinancing does.

Around that sit the markers of a repricing rather than a rout. NextDC fell 39.5% in a day after cutting its earnings forecast. Meta posted negative free cash flow above $2 billion while raising the floor of its 2026 capex to $135 billion. Spreads on its five-year CDS hit a record above 124 basis points. Nvidia's touched 126 and Oracle's 343. Oracle carries about $132 billion of net debt and a BBB- rating from S&P, which makes it the bridge between junk-rated neoclouds and the hyperscalers. Credit desks told CNBC the high-yield market is effectively on hold until after the midterm elections, with investors waiting on permitting and regulatory clarity.

Short sellers have booked about $6 billion in paper profits this month, with short interest across five neocloud and data centre names running at 20% to 26% of free float, according to Reuters. That is not a market pricing a shortage. It is a market pricing the cost of funding one.

Dispersion, not collapse

The bulls have not gone quiet. Jensen Huang said in Berlin that "demand for the neoclouds is off the charts," and that demand for capital exceeds its availability, which he expects to be solved. Nebius's chief revenue officer told CNBC his company is not in CoreWeave's position because it does not carry that kind of debt, and that its 2027 capacity would sell out if brought online, with price per megawatt still rising.

Gartner's forecast is the sharpest counterweight, and it is worth stating precisely because it gets flattened in the retelling. Up to $7 trillion of AI infrastructure spending by 2030 would need $8.5 trillion of revenue to cover capex, financing costs and a 12% return, against expected AI revenue of about $3.5 trillion. The firm expects more than 80% of neoclouds to cease independent operations by 2030, which its analysts define to include acquisition and partnership, not only failure. AI-optimised infrastructure-as-a-service reaches $174 billion in 2030, from $66 billion in 2027.

That distinction is the one buyers keep missing. Consolidation is not disappearance, and a vendor absorbed into a larger company is a contract question rather than an outage.

What a buyer should take from it

Retire the elastic assumption first. Forward capacity is a planning item now, not a background condition. If the large blocks of 2027 and part of 2028 are committed, the workload that arrives late, or grows faster than the model said, competes for the held-back on-demand slice and for whatever price clears it. Any roadmap resting on "we will scale when we need to" should be tested against a reservation quote.

Then look at the contract and the counterparty behind it. AWS is matching five-year commitments to five-to-six-year servers, which is a conservative structure. A neocloud with a $100 billion backlog and $35 billion of debt maturing by 2030 is making a different promise with similar language. Ask what happens to your reservation, your prepayment and your data if the company is acquired or restructured. Gartner's Will Sommer put the wider version of this plainly: "For the first time in really a generation, the math is starting to move toward 'build' and away from 'buy' or 'partner.'" HyperFrame's third-quarter survey found neoclouds in use at 2% of respondents, against 59% for AWS and 51% for Azure. Most enterprises are exposed to the hyperscaler version of this story, not the neocloud one.

And the cost curve is no longer free. Inference being an order of magnitude too expensive is the vendor's own framing, and the fix he proposes is cheaper silicon, not cheaper capacity. Those are different bets, and only one of them runs on a two-year construction schedule. Model compute the way you model components: lead times, allocation rules, and a supplier who can commit to a date.

The crunch that matters to a hardware business rarely announces itself as a shortage. It arrives as a longer lead time, a reservation fee, or a supplier who will not commit to a date at all. At DMC we work with hardware companies on exactly that, from sourcing strategy and cost modelling to production ramp planning when demand outruns the plant. If your roadmap depends on compute you cannot book yet, let's talk.