Why Compute Needs a Big Down Payment

Greetings!

If it suddenly feels like every startup is trying to raise money to throw at some kind of AI infrastructure shortage, you’re probably onto something.

We reported last night that the startup behind AI assistant Instinct is in talks to raise $1 billion, as it faces compute capacity constraints after growing to more than 100,000 users on an invite-only basis. Meanwhile Coatue Management has been discussing a multibillion-dollar joint venture involving chip startup MatX, to finance purchases of memory and logic dies and the reservation of manufacturing capacity.

There’s more going on than simply scrounging up a few spare servers or chip components. Getting access to AI infrastructure increasingly means making big, long-term commitments—it’s cheaper for providers to finance new cloud capacity that way, while manufacturers are more willing to allocate scarce capacity to large, established customers. That’s making it tougher for young but fast-growing companies to get a place in line.

Smaller cloud companies often borrow heavily to finance data center capacity and purchases of graphics processing units, while even the biggest are increasingly using debt to supplement balance sheet cash. On top of that, borrowing costs have risen sharply across the economy, with benchmark yields on both long- and short-term debt climbing.

It’s generally cheaper to borrow money with established, long-term customers lined up, especially ones with good credit ratings, than to finance compute capacity to serve less-proven customers. Iren, for instance, said in August that roughly $3.6 billion in financing for GPUs for Microsoft carried an interest rate of around 6%. But a separate $2.4 billion in financing for GPUs for non–investment-grade customers had a rate around 3 percentage points higher.

Some cloud providers are also demanding more up-front payments, which helps lower the amount they have to raise. Iren said recent customers were paying as much as 55% of the cost of GPUs and related equipment up front. Nebius, similarly, said that the share of deals with prepayments reached an all-time high in the second quarter, with roughly 70% of newly inked deals requiring money up front.

Nebius expects to receive more than $9 billion in prepayments this year, compared with projected 2026 revenue of up to $3.4 billion, and called them the “market standard for securing capacity in a supply-constrained environment.”

Renting GPU capacity as needed from startups like Runpod is another option, while larger providers like CoreWeave are introducing flexible ways to buy inference capacity. But that approach can be expensive, and needed capacity might not always be available. Either way, scaling up can get pricey fast for AI-heavy startups, and they might need to reserve capacity before knowing exactly how usage growth will pan out.

In the pre–AI craze days, a consumer app might chase user growth without worrying about revenue, while computing needs were generally less intensive and cheaper, so didn’t require the same kind of hefty upfront payments. Photo sharing app BeReal, for instance, grew to nearly 8 million daily users within months of taking off in 2022, having raised roughly $90 million, though growth later fizzled.

On the chipmaking side, startups are also facing challenges in securing resources ahead of growth, in that manufacturers are more likely to allocate capacity to customers that can make large commitments and have the financial resources to back them up. The joint-venture financing for MatX, which doesn’t expect to have its chip design ready for trial production until next year, would be aimed partly at helping it reserve capacity like customers such as Broadcom or Marvell Technology do.

It’s not clear how much money Coatue would put into a joint venture, which remains in the early stages of discussion, and whether the venture could also pull in outside financing. But Coatue already has some experience helping finance capital-heavy infrastructure deals on the data center side, sometimes in creative setups.

Coatue’s Next Frontier venture participated in a joint venture with AI infrastructure startup Fluidstack to finance a planned Indiana data center campus. A subsidiary of that JV raised $5.7 billion of debt to help buy land and build the facilities and related power infrastructure. (The financing doesn’t cover chips.) Fluidstack will lease those planned facilities to provide AI cloud services, but there’s yet another company involved—Google, which is guaranteeing Fluidstack’s lease obligations.