GPU Rents Hit a Record. Read the Fine Print.

3 days ago

5 min read

A rising price tells you demand is beating supply. It doesn't tell you which side moved. Right now that difference decides who wins the next leg of the AI trade.

The Read

Blackwell Ultra (B300) rental prices just rose 17 days in a row to a new all-time high, according to an index of verified rental transactions from U.S. neoclouds. That index hit $7.45 an hour on October 14, up from a low of $4.55 last November. That's a 64% climb in under a year. The popular read writes itself: compute is sold out, demand is endless, buy everything with "AI" in the description.

Same week, CoreWeave's 2032 bonds traded below 80 cents on the dollar, with yields above 15%. Think about that for a second. The product these companies sell is at a record price, and lenders are pricing one of the biggest sellers like it might not pay them back. Both things can't mean "endless demand." One of them is telling you something the headline isn't.

The Chart Everyone's Sharing

B300 GPU rental price index (DESKB300), dollars per GPU-hour, Oct 2025 to Oct 2026, hitting a record $7.45 on Oct 14

B300 GPU rental index (DESKB300), $ per GPU-hour, Oct 2025 – Oct 2026. Source: Bloomberg.

What a Price Actually Tells You

A price goes up when buyers want more than sellers can deliver. That's all it tells you. It can happen because demand surged. It can also happen because supply got stuck while demand just kept doing what it was already doing.

Those two stories produce the exact same chart. They don't produce the same winners. In a demand boom, everyone in the chain gets paid: the chipmakers, the builders, the lenders, the landlords. In a supply squeeze, only the people who already have working capacity get paid, and everyone with capacity stuck in the pipeline gets squeezed along with the supply.

The Bottleneck Moved

This is the part of how AI is evolving that most people haven't caught up to. In 2023 and 2024, the constraint was chips. You couldn't get H100s at any price, and Nvidia was the whole story. That's not the constraint anymore. The constraint is now power, permits, land, and the physical buildings to put the chips in.

Satya Nadella said it out loud last year: if you can't get the power and the buildings ready, "you may actually have a bunch of chips sitting in inventory that I can't plug in." That's the CEO of Microsoft telling you the chip isn't the hard part anymore.

And it's getting harder, not easier. On September 24, Oracle sent a force majeure notice on Project Jupiter, the $165 billion, 2.45-gigawatt Stargate campus in New Mexico, citing delays in securing power. That project has also run into a denied energy permit and loud local opposition. That's the political pushback against data centers I've been talking about, showing up in an actual contract.

Here's the nuance on the B300 specifically. A lot of B300 deployment needs liquid cooling and far more power per rack than older data halls can deliver. It mostly goes into new or retrofitted facilities. So a record B300 rental price is, in part, a gauge of how slowly those new buildings are showing up.

Why the Bond Market Isn't Buying the Headline

If rents were rising purely on demand, the neoclouds renting out GPUs would be swimming in cash, and their credit would be getting cheaper. Lenders would see less risk. The opposite is happening.

The GPUs already plugged in and running are a gold mine at these prices. Lenders aren't worried about those. They're worried about the billions borrowed for data centers that aren't finished or powered yet. That debt costs money every month. The building doesn't earn a dime until it's switched on. If power, permits, or a town council stall it, the company carries the interest with nothing coming in to pay it.

One more wrinkle: that rental index tracks spot and short-term transactions. Most neocloud capacity is locked into multi-year contracts at fixed rates. So the record price on the screen often isn't what those companies actually earn, while their construction and borrowing costs keep climbing. Higher rents can show up on the chart and still not show up in their bank account.

Not Every "Can't Build" Is the Same

This is where I'd push back on the lazy bearish read too. "We can't build fast enough" comes in two flavors, and they hit the builders in opposite directions.

Flavor one: can't get permission or power. Permits denied, interconnect queues, gas pipelines delayed, towns voting no. That's bad for builders. The backlog looks huge on paper, but the revenue keeps sliding to the right, and the lenders behind those stuck projects take the hit.

Flavor two: not enough crews and equipment. Everything's approved and powered, there just aren't enough electricians, transformers, and contractors. That's great for builders. PWR, STRL and the like get pricing power and fatter margins. So when you look at a name like STRL that round-tripped its whole earnings pop, the real question isn't "is the buildout over?" It's "which flavor is it stuck on?" One makes it a value spot. The other makes it a value trap.

Who Wins When Compute Is Scarce

If this is a supply squeeze, the winners are the people who already have what everyone else is waiting for. Owners of powered, running capacity, especially on short contracts that can reprice higher. Owners of the power itself. And the efficiency trade: anything that squeezes more output out of every watt matters a lot more when the watt is the thing you can't get.

It's also a tailwind for the Great AI Repatriation I've been talking about since January. When cloud compute is scarce and expensive, more companies decide to just own the box. That's DELL, which I've been rebuilding as a core holding, and HPE.

The losers: neoclouds stuck in fixed-rate contracts while their costs rise, lenders to stalled projects, and eventually Nvidia's unit growth. You can't keep shipping GPUs to buildings that don't exist.

How I'll Know Which One It Is

Older-chip rents. If H100 rental prices are firm too, demand is broad and real. If only the newest chips are tight, it's a deployment bottleneck.

Capacity coming online. Are new data centers being switched on when promised, or are the dates quietly sliding?

Rents versus spreads. Rents up and credit spreads tightening is a demand boom. Rents up and spreads blowing out, which is what we have right now, is a supply problem with a debt bill attached.

Hans's Take

Yeah, I know, record GPU prices sound like the most bullish thing in the world. Sometimes they are. But a price that rises because you can't build is a very different animal from a price that rises because everyone wants more. The AI demand story is alive. The buildings, the power and the debt are where the next round of winners and losers gets decided. Own what's already running. Be picky with what's still waiting for a plug. Trade smart, stay hedged.