Coreweave
CoreWeave is the largest AI neocloud and went public last week. Some commentators have wanted to use the CoreWeave IPO as a way to cast doubt on the ‘Generative AI’ industry. See, for example, a recent Bloomberg Opinion column entitled CoreWeave’s IPO Will Expose AI’s Dirty Secrets…
CoreWeave stands to be a bellwether for the AI industry as a whole — a must-watch stock as questions about return on investment grow ever louder. Any slowdown in demand for CoreWeave’s “compute,” as the term goes, will be seen by Wall Street as a heavy indication of a softening across the board, dragging down Amazon.com Inc., Google parent Alphabet Inc., Microsoft Corp., Nvidia and several others.
This isn’t true at all. CoreWeave is a weird company: it got 62% of its revenue last year from Microsoft, and another 15% from its biggest supplier, Nvidia. This makes it an interesting lens into compute markets, but it cannot be an index of them. The best intuition I have for the company was from a line in Scott Alexander’s Meditations on Moloch…
Las Vegas doesn’t exist because of some decision to hedonically optimize civilization, it exists because of a quirk in dopaminergic reward circuits, plus the microstructure of an uneven regulatory environment, plus Schelling points.
Just as the casinos and hotels we’ve carved out in the desert are the result of a quirk in human desire, Coreweave is carved out too: from Nvidia’s desire to reduce its customer concentration, and Microsoft’s desire to grow its asset base efficiently. Las Vegas was on track to become an irrelevant, middle-of-nowhere railroad town; but historical contingency intervened. In 2022, CoreWeave appeared to be an unprofitable Ethereum miner, but circumstances—AI progress—and a whole lot of capital, intervened.
One way of looking at the company is as a business line for Nvidia. Nvidia owns 6% of the equity, but this understates the depth of the relationship: CoreWeave is arranging tens of billions of dollars in credit facilities, to spend the majority on Nvidia chips and networking, before renting some fraction of this compute back to Nvidia. This could be, partially, a convenience for their internal R&D efforts—it's a hassle to build their own datacentres—but this doesn’t feel like a sufficient explanation, because Nvidia already runs its own small cloud provider.
It is, in part, an effort to weaken the bargaining power of their large customers. About half of Nvidia’s revenue comes from just four customers—Amazon, Google, Meta, and Microsoft—all of whom have internal chip design efforts, so they are vulnerable to these companies changing their orders. This would explain why CoreWeave was the first to offer Nvidia’s newest hardware, the GB200, in February 2025. But this explanation also feels insufficient: if Microsoft is a majority of the revenue, it's hardly diminishing Microsoft exposure that much.
Nvidia benefits from CoreWeave existing. When demand for CoreWeave’s IPO looked shaky, Nvidia backstopped it with an additional $250 million investment. The CoreWeave CEO said they couldn’t have done it without them.
The other lens is through Microsoft. For them, CoreWeave is a tool to manage their datacentre fleet construction. When a cloud provider wants to build a new datacentre, they are looking at about five years to get it running. If you need to build new power, the decision timeline is even further out. Microsoft’s decision to restart a nuclear reactor at Three Mile Island, will start providing power in 2028, and they have a 20-year power purchase agreement. How could they decide that this would be a good investment last year? It is extremely difficult to predict demand on this horizon; you’d need to answer questions like:
How much will hardware improve, in energy efficiency terms?
How much will software improve, in inference cost per token terms?
How many tokens will we want to spend for each query, on average?
How many queries will we want to make, if we have long-horizon agents, or an open-ended AGI?
Where will we want to do inference in the world, so datacentres can be nearby for the lowest latency? (As Satya put it, “[a]t the end of the day, speed of light is speed of light, so you can't have one data center in Texas and say, ‘I'm going to serve the world from there.’” Clearly a subtle jibe at OpenAI’s expectation that Stargate can service three quarters of their compute needs.)
The fortunate thing is that, if you get this prediction wrong, some of the investment can be repurposed. The land, power, and datacentre that was meant to be for AI training can be repurposed for AI inference, Azure CPUs, or storage. Microsoft said five times in their recent earnings that they are building a “fungible fleet”, and they also mentioned that about half of their $80 billion in AI capital expenditure is being spent on long-lived assets, and the other half is short-lived assets, like GPUs.
This is where CoreWeave comes in. It would be ideal for Microsoft if Azure owned all the compute they need for AI inference to serve all their customers in 2029, because the margins on this are better. But doing this requires taking risk on long-lease assets like power-purchase agreements and datacentres. Signing a long-term contract with CoreWeave means they can have access without needing to take on the risk of the long-lease assets. (Renting a piece of hardware for almost its entire useful life is as good as owning it.) You can think of Microsoft’s own fleet as ‘baseload’ compute, which they are more confident they can make a return on, and Coreweave as ‘top up’, for which they accept a reduced margin to ensure they can serve demand without long-term risk. The same pattern will be true for OpenAI.
The challenge then, is finding a price where this works for both sides. The limited useful life of hardware is a strain on this. The cycle which is going to dominate rentals will be:
A new generation of hardware is released, and renters sign up to 2-3 year contracts, while others sign shorter-term 1 year or 6 month deals.
As these contracts expire, the next generation of hardware is around the corner, with better cost efficiency and energy efficiency.
This exerts two downward forces on the rental price for this hardware. First, there’s simply a glut for short leases, because none of that hardware is being signed to longer leases. And second, when the newer hardware is more performant, the incentive to switch is stronger.
Over time, the marginal cost of operation for old hardware will be higher than both the upfront and operating cost of new hardware (in compute per dollar). Jensen also highlighted on Nvidia’s latest earnings call that there’s an opportunity cost for datacentre space and power too:
If you have a 100-megawatt data center, if the performance or the throughput in that 100-megawatt or the gigawatt data center is four times or eight times higher, your revenues for that gigawatt data center is eight times higher. And the reason that is so different than data centers of the past is because AI factories are directly monetizable through its tokens generated.
SemiAnalysis has specific numbers on this improvement in ‘performance against ownership cost’, though it's behind their paywall so I won’t quote them here. Satya gave a rule of thumb on the last Microsoft earnings call about how this dynamic affects their investment decision…
You don’t want to buy too much of anything at one time because of the Moore’s Law every year is going to give you 2x. Your optimization is going to give you 10x. You want to continuously upgrade the fleet, modernize the fleet, age the fleet, and at the end of the day, have the right ratio of modernization and demand-driven monetization to what you think of as the training expense.
At Nvidia GTC, Jensen suggested even more aggressive rates of improvement and went as far to claim, “When Blackwells start shipping in volume, you couldn’t even give Hoppers away.” This annual release cycle compresses the hardware’s effective life, which is an issue for CoreWeave. Their IPO filings say they breakeven on hardware on average 2.5 years after purchase, but this will have been buoyed by the H100 shortage in the earliest days of the ChatGPT-induced investment boom and the delays to the new Blackwell generation. SemiAnalysis has a model for CoreWeave’s payback period again, which I won’t quote specifically, but it seems very difficult to imagine if Nvidia can keep its pace of releases and improvements as high as it intends to, that CoreWeave would be able to earn a suitable return in time.
This is what makes OpenAI’s 5-year contracts with CoreWeave so interesting. Given the rate of improvement in hardware, it seems undesirable to commit to use 2025 accelerators in 2030. What might be going on here?
This might not be for specific hardware, but for capacity, in FLOP terms, or otherwise.
They might be able to get out of these contracts. (It isn’t so clear how long Microsoft’s contracts were, but there is an FT report they’ve been able to step back from some capacity. Note that CoreWeave disputes this.)
Perhaps they are willing to pay for 2025 hardware in 2030, because they know CoreWeave has earliest access to the newest hardware which offers a few months of lower margins.
Or something else…
While we don’t know for sure, the overall picture is clear: CoreWeave’s existence doesn’t depend on an underlying economic engine, but whether it is advantageous to Nvidia and the clouds (including OpenAI). There are a lot of other peculiarities that I’ve left to the side to make this point: CoreWeave’s subsidiaries borrow money to build compute with loans that are secured on the compute which will be worth very little when the debt comes due. CoreWeave had a technical default on a loan because of an admin error. CoreWeave’s founders, per The Diff, have sold $450 million in secondaries, and now own just 2.4% of the equity. Finally, this source casts doubt on CoreWeave’s ability to grow its power supply through a partner, Core Scientific.
At the end of the day, none of these issues answer CoreWeave’s main question: does its existence provide convenience to Nvidia, Microsoft, and OpenAI? If it goes bust in the next few years, this won’t reflect the top of an AI bubble, moreso that it stopped making sense to prop it up.
Otherwise
OpenAI raised $40 billion at $300 billion post-money valuation. Lots of people will be shocked by the valuation — Anthropic, by comparison, raised at $60 billion — but it is further confirmation that research labs are becoming product companies. When Sam Altman was interviewed on Stratechery, Ben asked, “What’s going to be more valuable in five years? A 1-billion daily active user destination site that doesn’t have to do customer acquisition, or the state-of-the-art model?”; Sam’s response, “The 1-billion user site I think.” You can simplify this investment in OpenAI to a simple question: can this company become Meta? In this light, Meta is about five times bigger than OpenAI today, in market cap terms, so a 20% chance seems about right.
xAI bought X (formerly Twitter) in a transaction valuing xAI at $80 billion and X at $33 billion. The story people would like to tell here is that this transaction makes sense most of all for Elon: to offload X’s debt onto xAI, which has a lower cost of capital. Matt Levine will cover this transaction better than I can, but it is worth noting that, in light of Sam’s comments above, that 600 million weekly active users is more than OpenAI’s 500 million (though OpenAI is growing much faster).
Anthropic released two papers on the thought patterns of language models. Circuit Tracing: Revealing Computational Graphs in Language Models and On the Biology of a Large Language Model.
I wrote about why models of explosive economic growth, like Epoch’s GATE model, are misleading here.