Megawatts for Rent

Programming note: ARPU will return next Friday and take a look at recent moves by frontier AI labs.

Not All FLOPs are Created Equal

If you read the financial press over the last few weeks, you might have noticed a glaring contradiction in the AI compute narrative.

On one hand, companies that built massive internal AI infrastructure are starting to sublet their capacity to the outside market. Meta, having stockpiled billions of dollars in chips, is preparing to open its private AI servers to external customers. SpaceX is executing the exact same strategy, recently locking in deals to rent its Memphis data center to Google and Anthropic for nearly $1 billion a month

On the other hand, we have the "neoclouds"—specialized AI infrastructure startups like CoreWeave and Nebius. Nebius just reported a near eightfold revenue surge and raised its annual CapEx forecast to $25 billion. CoreWeave bumped its own CapEx guidance to $31 billion. Their executives are telling Wall Street that customer demand is utterly insatiable and that they cannot secure Nvidia GPUs fast enough.

So which is it? Is the tech industry drowning in excess compute, or is it starving in a historic shortage?

The Skeptic's Verdict

For the AI skeptics, the answer is obvious: the bubble is bursting.

Ed Zitron, one of AI's fiercest critics, recently pointed to Meta and SpaceX as the canary in the coal mine. His argument is compelling: if AI companies are suddenly subletting their servers to their rivals, it is a tacit admission that they massively overbuilt and have no real internal use for the hardware.

Here's Zitron:

That's the thing with Meta. They are talking about leasing out their compute to someone else. This is a sign they don't need the compute. And it's also a sign that perhaps they overbuilt.

... It begs the question as to who's next and also when Meta will cut capex because at this point, if they keep spending on capex after saying this, what exactly is the plan? Are they going to become Coreweave too?

The bears might eventually be proven right, but there is one major piece of financial data getting in their way: the price of renting a GPU is still rising.

The Pricing Test

If the market were truly drowning in an oversupply of compute, the rental price of a GPU should be crashing. Instead, the exact opposite is happening.

Last week, Amazon Web Services (AWS) raised prices on its EC2 Capacity Block reservations for its most powerful Nvidia GPU instances by 20%, citing supply and demand dynamics. You do not raise reservation prices by 20% in a glut. The AWS hike strongly suggests that, at least for now, the core market remains deeply supply-constrained.

So how do we explain Meta and SpaceX renting out their servers?

The contradiction exists because we are still using the phrase "AI compute" as if it describes a single, fungible commodity. It does not.

Compute Market Fragmentation

In the old days of cloud computing, capacity was relatively simple. A CPU in Virginia was broadly the same product as a CPU in Ohio. But AI compute is highly conditional; it cannot be abstracted into a generic utility.

Instead of a single market, AI infrastructure has fractured into specialized sub-markets. A surplus in one bucket does not solve a deficit in another.

• Training vs. Inference: If you are a frontier lab training a massive model, you need a supercomputer—tens of thousands of chips tied together with incredibly complex, high-bandwidth networking. But once that model is trained, consumers need to use it. That requires "inference" compute, which demands geographically distributed servers optimized for latency and cost-per-token. A cluster perfectly optimized for batch training in Iowa might be highly unprofitable for running low-latency inference in London.
• Reserved vs. Spot Capacity: The market is also split by contract type. Major AI labs need guaranteed, 24/7 access to build their product roadmaps—this is long-term reserved capacity, and it is expensive and scarce. Conversely, short-term burst (or spot) capacity is highly volatile. An abundance of short-term rental capacity does not solve a shortage for a company that requires a multi-year, locked-in guarantee.
• The Enterprise Wrapper: Startups like CoreWeave built massive valuations by offering raw, unadulterated access to GPUs. But a raw GPU at a neocloud startup is practically useless to a global bank whose security and compliance departments mandate Microsoft- or AWS-certified infrastructure. For these regulated customers, the choice isn't just about chip speed; it is about risk management, audit trails, and data sovereignty.

The Dynamics of Idle Capacity

When you combine these fractures with the brutal reality of hardware depreciation, the paradox of excess capacity seems to make sense.

GPUs depreciate on a 5-year schedule. If you commit to building a massive data center years in advance, the hardware might arrive before your internal product roadmap is fully ready to consume it 24/7. In a capital-intensive industry, idle capacity is pure loss.

In a recent interview with Bloomberg, Mark Zuckerberg explained his logic. He denied that Meta has "overbuilt," insisting that they are currently using all their compute. Instead, he framed the move as opportunistic arbitrage, citing Elon Musk’s SpaceX model of short-term, high-premium leases as inspiration:

The SpaceX model I think is quite interesting in terms of just making these short-term deals that are at a big premium... The offers that you get for using the compute are so high that it may make sense, in some cases, to rent out [the capacity]... instead of your own internal uses.

In other words, a way to offset the bills.

The AI industry is becoming less like software and more like the energy grid—location-specific, heavily regulated, reliant on physical infrastructure, and brutally dependent on utilization.

Just like a power grid, you can have a blackout in one city while a neighboring state has a surplus of electricity. It is possible for Meta or SpaceX to have excess capacity in one specific configuration, while a compliant enterprise inference cluster in Frankfurt is completely sold out.

The presence of "excess compute" does not necessarily mean there is a supply glut. It just means the market is getting messy. A company can be long on its current hardware footprint and short on its future ambitions at the exact same time.

If this is indeed where the compute market is heading, then the next phase of the AI boom will not be a simple story of building more datacenters. Instead, it will be about the complex logistics of matching the right chip, to the right power source, under the right contract.

Signal Stack

The operating reality beneath the headlines.

• BlackRock Sees Private Credit Taking Bigger Role in AI Buildout (Bloomberg) – Bloomberg Intelligence projects capital expenditure for the six largest US hyperscalers will reach approximately $820 billion this year, up nearly 80% from 2025's then-record levels, with BlackRock's head of research arguing that companies' needs are now so vast they will have to borrow "just about everywhere" — creating a structural tailwind for private credit even as Apollo and Ares have already begun curbing investor fund redemptions.
• Will Someone Finally Blink in the AI Spending War? (WSJ) – Chip companies in the S&P 500 now account for approximately 18% of the index's total market capitalisation, up from around 5% five years ago—a concentration that means any signal of hyperscaler spending moderation moves the entire market.

📺 On Our Channel

The Extreme Engineering Behind Meta's AI Data Center

In 2014, the entire US data center industry consumed an average of 8 gigawatts of power. Meta is now building a single campus designed to need 5 gigawatts on its own. To get it built, Meta needs new power plants, 240 miles of transmission lines, floors engineered to hold racks heavier than cars, liquid cooling systems running at 120 kilowatts per rack, and 5,000 skilled trade workers in a Louisiana parish of 20,000 people. We broke down the physical systems—building, power, cooling, networking, and coordination—that the AI boom has turned into infrastructure problems at an entirely new scale.

Watch ARPU's deep dive on YouTube (15 Mins)

You received this message because you are subscribed to ARPU newsletter. If a friend forwarded you this message, sign up here to get it in your inbox.