AI Data-Center Spending Shrinks as Compute Stays Sold Out

By Amara Osei

AI data-center spending may reach $3 trillion to $4 trillion through 2030, down from an earlier projection of $7 trillion, even as operators report sold-out computing capacity.

Stephanie Link of Hightower told CNBC on November 11, 2026: “Maybe we’re not going to spend $7 trillion on data centers between now and 2030. That was what we were projecting. Maybe it’s 3 to 4 trillion. That’s still a lot of money.”

Lower spending projections can reflect rising financing costs and customers buying hardware themselves. They do not, by themselves, show weaker demand for working AI compute, because power access, permits and cluster commissioning can limit usable supply.

The two signals can coexist because financing terms determine who pays for hardware while infrastructure limits determine when it becomes operational.

What does CoreWeave’s backlog show?

CoreWeave’s third-quarter revenue rose 168.3% year over year to $3.66 billion, above the $3.54 billion consensus. Its $1.27 loss per share came in below the expected $1.55 loss. Revenue backlog reached about $132.5 billion, and the company raised full-year revenue guidance to $13.0 billion-$13.3 billion.

CEO Michael Intrator said on the earnings call: “We are largely sold out through 2027.” Prices for older Ampere and Hopper chips rose 30% to 50% quarter over quarter.

On Mad Money, Jim Cramer relayed that CoreWeave’s CEO told him 2027 demand was 10 times oversubscribed and added 2028 demand was a multiple of available supply. Cramer also said the CEO would begin earning more next year on Nvidia chips bought four years ago than in any of the prior four years.

The results lifted AI cloud stocks and Nvidia as the figures eased bubble fears. Investors still face funding risk: CoreWeave plans to raise at least $4 billion through a high-grade loan. Its shares were down 23.4% year to date, despite a 30.8% gain over the previous week.

Demand is not CoreWeave’s immediate constraint; funding and usable power are.

Why does Oracle’s pricing look tight?

Oracle’s fiscal first-quarter cloud infrastructure revenue rose 121% to $7.39 billion, while remaining performance obligations reached $664 billion. The company booked more than $30 billion in new AI contracts, ran its GPUs at 97.9% usage and renewed expiring capacity at a 20% premium.

Co-CEO Clayton Magouyrk said on the September 10 call that prices do not fall when demand exceeds supply; they rise.

Oracle expects fiscal 2027 revenue of at least $90 billion. Its shares were down 34.4% year to date. Most new contracts use prepayment or bring-your-own-hardware terms, which can reduce the capital Oracle must provide even when contract commitments remain high.

Oracle can report firm pricing while investors wait for its backlog to become recognized revenue.

How long does capacity take to become usable?

Elon Musk said on X on November 11 that usable capacity is the relevant measure. He estimated that working continuously takes about two months to make a large compute cluster reliable, while others may need closer to six months.

That delay separates spending on buildings and equipment from working compute. Vertiv sells the power and cooling equipment that helps operators bring clusters online.

Vertiv’s third-quarter revenue rose 61.2% to $4.31 billion, while organic orders grew more than 100% year over year. Full-year guidance calls for net sales of $14.45 billion-$14.85 billion and adjusted EPS of $7.10-$7.20.

CEO Giordano Albertazzi said customers care about time to first token. He identified power availability, permitting and skilled labor as constraints. Vertiv shares were up 26.62% this year.

Vertiv ranks among seven power, cooling and networking companies identified as AI infrastructure beneficiaries rather than chipmakers. Its order growth shows equipment demand, but site and labor constraints still govern deployment.

Vertiv has the orders, while power, permits and labor determine when that equipment produces usable compute.

Where does the AI spending cycle stand?

Former Cisco CEO John Chambers told CNBC on November 11 that AI is here to stay and described the market as being in its second inning.

His view leaves open the central investor question: whether constrained capacity and pricing power can persist after spending forecasts reset lower. The figures point to near-term scarcity, but they do not remove execution risk from financing, power delivery or backlog conversion.

The argument has shifted from whether AI demand exists to how quickly spending becomes working compute.

What will decide which story wins?

Three operating and financial tests will separate a spending slowdown from a demand slowdown:

• CoreWeave: Track whether the company reaches its target of more than 1.85 GW of active power by year-end and the pricing on its new loan.
• Oracle: Track whether about half of its $664 billion backlog converts to revenue within 36 months, as the company expects.
• Vertiv: Compare actual results with fourth-quarter sales guidance of $4.31 billion-$4.51 billion. Continued increases in GPU renewal prices would indicate that capacity remains sold out.

For CoreWeave, power and funding matter most; for Oracle, contract conversion; for Vertiv, equipment deliveries and customer capacity.

Financing and delivery will determine whether lower forecasts reflect efficiency or weaker demand.