GPUs Are Scarce; for Some, the Money Is Scarcer

NEW YORK, Oct. 22 — OneOne fear behind this month’s slide in AI stocks has been a glut: too many chips, too little demand. “At the moment, we have more compute capacity than we have demand,” John Lovelock, Gartner’s chief forecaster, told CNBC on Monday.

Current rental prices offer a different signal. Renting an Nvidia H100 cost a median $3.70 an hour in the week of Oct. 19, up from $2.98 a year earlier, by GetDeploying’s count of cloud list prices. The makers of chips and chipmaking tools say capacity is their limit. Harder to find may be long-term money to build more, priced very differently for a hyperscaler than for a heavily indebted neocloud.

That difference shapes who profits from tight supply. Companies that own capacity and can borrow cheaply collect the rising rents, and some of Nvidia’s suppliers are being paid in advance to build more. Among the neoclouds, the specialist clouds that rent out Nvidia’s chips, the most indebted face costly capital, and their lenders carry the risk that rents turn before the debt comes due.

The Rents

By GetDeploying’s count the H100 rented for a median $3.70 an hour across 40 providers in the week of Oct. 19, against $2.98 across 36 providers in the week of Oct. 20, 2025, up 24 percent in 12 months. The H200 cost $5.00, against $3.50, and the B200 $7.99, against $5.59, each up about 43 percent. Across the 21 models priced at both ends of the year, on-demand prices rose a median 18 percent; 17 rose and four fell.How the Index WorksGetDeploying counts price per GPU per hour: an eight-GPU H100 machine renting at $40 an hour counts as $5.00. It takes each provider’s own median first, then the median across providers, and leaves out spot and reserved prices and any model listed by fewer than four providers that week.

The site’s own warning is plain: “These are medians across providers, not a rate you can book.” These are list prices for on-demand use; across its comparisons, at the median, a one-year reservation runs about 28 percent cheaper, and a three-year one about 49 percent, it says. A median also moves when a provider starts or stops listing a chip, and the site calls its early history approximate: prices were updated by hand until February 2026, which makes the year-earlier figures the softest in the comparison.

A second gauge agrees. Ornn’s Compute Price Index, a volume-weighted average of rentals actually transacted, settled on Wednesday at $3.09 an hour for an H100 SXM, up 10.75 percent in a month; its B200 rose 21.92 percent, to $9.40.

So does one company that rents chips out. Nebius, a neocloud, ran a capacity auction on Tuesday that showed “continued upward mobility in terms of the kind of price per megawatt,” Marc Boroditsky, its chief revenue officer, told CNBC.

The Makers

On Wednesday Nvidia said it was working with six suppliers to expand capacity in substrates, packaging, testing and optical engines, and paying some of them in advance. It is giving King Yuan Electronics, which tests chips, a $1.1 billion capacity reservation advance for new testing facilities in the U.S.The SixAmkor, Ibiden, King Yuan Electronics (KYEC), Samsung Electro-Mechanics, Siliconware Precision Industries (SPIL) and Unimicron. Ibiden’s multibillion-dollar advance was agreed earlier this year. Nvidia expects its testing capacity to double from 2026 to 2027. “We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint,” Jensen Huang, Nvidia’s founder and chief executive, said in the announcement.NVDANasdaq204.64▼9.02−4.2%Nvidia · Friday’s session

Demand is extraordinary. Capacity is the constraint.

The toolmakers describe the same limit. “As we move into calendar year 2027, we see clean room space availability and supply chain constraints for certain components continuing to limit chip makers and equipment companies’ ability to fully meet end demand,” Tim Archer, chief executive of Lam Research, said on its earnings call on Wednesday, according to a transcript published by The Motley Fool.

Haviv Ilan, Texas Instruments’ chief executive, said on its call, according to the Motley Fool’s transcript, that TI wanted to “aim higher” in its capital spending but was “seeing lead times extending on tools and everything, especially on the back end side.” And on Wednesday Susan Li, Meta’s finance chief, told analysts that Meta was “facing the same kinds of capacity constraints as everyone else.”

Sellers of chips and tools have every reason to speak of demand. A Gartner report last week found that “AI revenue will be insufficient to support profitable returns on all planned AI capital expenditures” from 2027 to 2030, CNBC reported. Mr. Lovelock’s worry is partly about who pays: “we have a lot of the tokens and processing power being given away in freemium solutions,” he told CNBC. Rents are rising for now. That does not settle whether all the capacity being planned can earn a return.

The Money

The strain shows in the traded debt of CoreWeave, the biggest of the neoclouds: its unsecured debt yields about 15 percent, and the spread on its five-year credit-default swaps last stood at nearly 1,067 basis points, CNBC reported on Wednesday. That yield is what investors demand to hold the debt, not what CoreWeave pays on what it has borrowed; in a presentation filed in September, it put its weighted average interest rate for the second quarter at 8.3 percent. But it suggests what new unsecured borrowing could cost. CoreWeave has about $20.7 billion of debt coming due through 2030 and another $14.9 billion after that, CNBC reported.The Bill Coming DueCoreWeave had $5.5 billion of cash and equivalents at the end of the second quarter. Rob Oliver, an analyst at Baird, estimates that it will need to raise roughly $30 billion in 2027, CNBC reported.

Unnamed sources on credit desks told CNBC that the high-yield debt market was effectively on hold until after the midterm elections, given concerns about political pushback against data centers; one said investors “want permitting and regulatory requirements to be substantially resolved” before they commit financing.

CoreWeave declined to comment to CNBC. In a memo to employees that CNBC obtained, Michael Intrator, its chief executive, wrote: “Our financing model combines infrastructure-backed borrowing, corporate debt, equity, and customer prepayments.” Analysts at Goldman Sachs, JPMorgan Chase and Barclays issued bullish notes on CoreWeave this week, CNBC reported, citing its GPU-backed loans.

The hyperscalers borrow on other terms: Alphabet, Amazon, Meta and Oracle have issued about $228.4 billion of bonds this year through Oct. 20, according to LSEG data cited by CNBC. Rick Rieder, BlackRock’s chief investment officer for global fixed income, told CNBC that he saw the best opportunity in investment-grade corporate bonds, hyperscalers’ bonds included, and urged caution outside investment grade, AI names included. “There’s parts of the data center supply that’s not hyperscaler guaranteed or hyperscaler backed,” he said. “That we’ve been more cautious around, and we’ll continue to be.”

Even the giants have felt the change. Meta’s preliminary information statement, filed on Wednesday, records that Moelis, financial adviser to a special committee of its board, briefed the committee on Aug. 4 on “the recent tightening of credit conditions, particularly for hyperscalers,” which showed in less demand for the July bond of Meta’s Texas data-center joint venture than for other issues.

Tight supply, then, favors those who own capacity and can borrow cheaply; the hyperscalers also charge more for the same chip.The MarkupFor an H100, GetDeploying finds that AWS, Microsoft Azure and Google Cloud charge a median $7.89 an hour on demand, about 75 percent more than dedicated GPU clouds’ $4.50. Rising rents may help the neoclouds too, but their financing risk differs: Nebius’s Mr. Boroditsky told CNBC that it was not in CoreWeave’s position because “we’re not carrying the kind of debt that they are.”

The next test is CoreWeave’s third-quarter report, which Nasdaq, using an algorithm based on past reporting dates, estimates for Nov. 9. Before then come the midterms on Nov. 3, which, CNBC’s sources said, the high-yield market is waiting out. Investors are also watching whether Nscale, a British neocloud backed by Nvidia, goes ahead with the New York listing it filed for last month, CNBC reported.