What A Tangled OpenAI Web We CoreWeave

Wouldn’t it be funny if all of that money that Microsoft spent last year paying neocloud upstart CoreWeave was just to support ever-embiggening AI training workloads at OpenAI as it makes its GPT models smarter?

If that is the case, then the five-year, $11.9 billion contract that OpenAI inked this week with CoreWeave is but a down payment on what OpenAI might ultimately spend at CoreWeave and its neocloud peers. Who, by the way and not at all coincidentally, don’t run myriad applications and data services and which also charge half as much for GPU instances as does Microsoft Azure and Amazon Web Services.

CoreWeave was forced to talk about the OpenAI deal thanks to a report in Reuters on Monday that outlined the deal, confirming not only did that deal cover $11.9 billion in compute capacity but also that ahead of CoreWeave’s initial public offering, CoreWeave is going to issue $350 million worth of stock and give that to OpenAI as a stake in CoreWeave. No money will be changing hands for the stock, which obviously then will be dilutive to all other shareholders in CoreWeave.

OpenAI has hundreds of millions of users of its ChatGPT service, and there is a good chance that this deal is more about inference than it is training, given that CoreWeave has over 250,000 GPUs in total in 32 datacenters peppered around the globe. But there is also very likely to be provisions for AI training for GPT foundation models.

Thanks to the S-1 filing with the US Securities and Exchange Commission ahead of the IPO, we have some insight into CoreWeave’s financials. The company has a fleet of over 250,000 GPUs in those 32 datacenters, and drove $1.92 billion in sales in 2024, up by a factor of 8.4X compared to sales in 2023. Microsoft represented 62 percent of those sales, or $1.19 billion. The OpenAI deal is worth ten times that amount (almost precisely) over a five year period, which means the run rate that OpenAI is looking at is a mere 2X what Microsoft was already spending in 2024 with CoreWeave.

At the current prices that CoreWeave charges for renting capacity on Nvidia “Hopper” H100 GPUs, those 250,000 GPUs could drive $13.49 billion in sales, as we talked about in our story from last week about the IPO. Which means that the CoreWeave GPU fleet is only running at about 14.9 percent of peak capacity. For lots of reasons, it can really only drive 55 percent to 65 percent of capacity, but it gets to charge for time it is not computing as well as time that it is, remember. You pay for cloud GPU capacity in units of time, not in flops performed. Most clouds are only driving 35 percent to 45 percent computational efficiency on their GPU clouds, according to CoreWeave. Including Microsoft Azure, by the way. If this is true, then every dollar that OpenAI spends at CoreWeave can, in theory, make the model go faster and spend less time – and therefore less money – for every unit of AI work.

The wonder is why OpenAI is not soaking up more CoreWeave capacity knowing all of this, and the best reason must be that OpenAI is already getting a CoreWeave price for the Azure capacity, which is a 50 percent discount. That’s probably not a coincidence on the part of CoreWeave, which does not have to support Windows Server and a zillion other features. It just has to run AI models. Period.

Another fun bit is that OpenAI is spending about as much money on CoreWeave capacity over the next five years as the $13 billion that Microsoft invested in OpenAI over the past several years. The details of those Microsoft investments in OpenAI were never detailed, but clearly a lot of the money Microsoft took out of cash to invest in OpenAI was round-tripped right back to buy GPU capacity on Azure, which helped pump up Azure revenues. This has been good for Microsoft, but we can see in its most recent financials that it has been petering out as OpenAI gets some capacity elsewhere.

It is not yet clear how CoreWeave will be a part of the $500 million Project Stargate effort that was announced with President Trump at the White House a month ago, but one can envision the Stargate partners investing in Nvidia GPUs and deploying them on Oracle and now CoreWeave clouds, and eventually shift towards homegrown OpenAI accelerators that are cheaper and more tuned to the GPT workloads. These would also be deployed in systems at Oracle and CoreWeave as they come out of the fabs some years hence. OpenAI has been rumored to be working on its own AI accelerators since the fall of 2023, and Broadcom is rumored to be helping OpenAI bring its AI accelerators to life and get them through the Taiwan Semiconductor Manufacturing Co foundries.