A very important part of the CoreWeave story hasn’t gotten nearly enough attention. It’s not the $99.4 billion backlog, even though that number is huge. It’s not the $35 billion in debt, even though that sounds scary. It’s something less loud that tells us more in some ways. Prices are still going up for GPUs from older generations, even though they should be going down as newer models come out. The co-founder of CoreWeave said that. You can’t get rid of that kind of signal once you hear it.
That detail is more important than most quarterly earnings reports for anyone trying to figure out what’s going on in the AI infrastructure boom. In the hardware market, it makes sense that chips from yesterday will be cheaper when chips from tomorrow arrive. When Nvidia comes out with a new generation, the old one slowly loses its value. That’s how things should work. At the moment, it’s not working that way because of high demand for computing power. This is keeping the price of hardware from falling, which tells you something important about how much pressure is building in this industry.
CoreWeave’s business is based on a simple idea: they buy Nvidia GPUs, put them together with custom software and infrastructure, and then rent that computing power to AI companies that need it quickly but can’t build it themselves. OpenAI, Meta, Microsoft, and Anthropic are just a few of the well-known AI companies on the client list. In the first quarter of 2026, sales reached $2.08 billion, up 112% year over year. It’s now $99.4 billion, which is almost 300% more than it was a year ago. Most of the time, those numbers are very high for a company that is growing.
But we should look at the cost side of this equation just as carefully. Venture capitalist Chamath Palihapitiya said on a recent episode of the All-In Podcast that it now costs closer to $100 billion to fully load up a modern one-gigawatt AI data center. CoreWeave plans to spend more than $30 billion on capital projects this year alone, which is more than double the $15 billion they spent last year. The business spends more on infrastructure each year than most S&P 500 businesses make in a whole year. And unlike factories, which stop needing to be built at some point, the need for more computers keeps growing.
It’s possible that this is just the cost of being the first in a generational race to build infrastructure. CoreWeave’s business model is very smart: they rent out physical data center space from outside developers, borrow money against future customer contracts, and only use the borrowed money when capacity is needed. The company has been able to grow at a rate that would have needed a much bigger equity base otherwise. The costs of debt have also gone down. By the end of 2025, the average interest rate on a loan will have dropped from the mid-teens to just below 10%. That’s not nothing.
Still, it seems like the fact that older GPUs are becoming more expensive is a sign that the market hasn’t fully understood yet. Even old hardware is becoming more valuable again, which means that the supply of computing power is still not even close to meeting demand. This is good for CoreWeave’s short-term pricing power. It also means that the money needed to keep building won’t get cheaper. Building a new data center costs more each time. Electrical engineers and other specialized parts are getting harder to find. Land prices keep going up in Northern Virginia, Texas, and the Pacific Northwest, which are all important markets.
A common bear case for CoreWeave is that it will go down in value and accumulate debt. These are valid concerns. In just one quarter, the business paid $536 million in interest. When new generations of GPU hardware come out, the old ones lose a lot of value. Some analysts say that this could make CoreWeave’s adjusted EBITDA margins much smaller when real economic depreciation is taken into account. Also, what happens when AI labs switch from training huge models to running inference at scale? This is a change that tends to favor hyperscalers like AWS and Azure, which have dozens of regions around the world, over a company with 40 facilities mostly in the United States.
But what’s important to think about is what the fact that GPU prices are going up everywhere says about the present. It seems to show that the building of AI infrastructure is still speeding up and not stopping. It means that companies like CoreWeave that are trying to provide more computing power are going to have to deal with high costs that aren’t going away any time soon. It also makes it seem like the risks here aren’t just financial. They’re building designs. As open-source development speeds up, the cost of building AI infrastructure will rise while the cost of models will fall. This could cause the difference between what it costs to build and what the market will eventually pay for compute to get smaller in ways that aren’t good.
In the AI infrastructure space, CoreWeave is still one of the more interesting bets. It’s hard not to notice how sure the market is that the demand for computers will keep growing at the same rate it has been for the last two years. That could very well be true. But you shouldn’t just nod and move on when a co-founder says that the prices of even older chips are still going up. This is to find out what that means for people further down the chain and to see if the foundations of this boom are really as strong as the backlog numbers show.
