CoreWeave, which is in the business of buying GPUs that other users can’t get their hands on and then renting them out to those users, is raising another round at a $23 billion valuation, which may include $400-500m for employees. CoreWeave’s fundamental insight is that there’s a slice of the AI capital structure that 1) generates immediate cash flow, 2) isn’t something Nvidia feels the need to do right now, and 3) benefits Nvidia by keeping GPU ownership, and thus AI, fragmented.

H100 costs about $25K right now, let’s say. CoreWeave prices an H100 at $4.25 an hour. Assuming the GPU is in use for about 3 years, and it’s in use every hour, total GPU-Hours over 3 years comes to total hours per month * number of months in 3 years = ~26K. Let’s assume a utilization rate of 50% (which is quite high, given almost every data center in the world is dramatically overprovisioned), giving us around 13,000 utilized GPU-hours.

Assuming facility building and operating costs are at 50%, the variable costs per GPU-Hour come to 50% of $25,000 / 13,000 = $0.96, while depreciation costs are $25,000 / 13,000 = $1.92, bringing total cost per GPU-Hour to $2.88.

That’s a gross margin of ~32% for the hottest chip on the market.

Previously specializing in ethereum mining, CoreWeave expected to generate approximately $1.5B in revenue in 2024, according to a report by Bloomberg News. Their main USP seems to be cost, claiming to be “upto 80% less expensive than generalized public clouds”, which I’m guessing stems from having solved GPU utilization rates at scale owing to their tryst with crypto mining.

Clearly, with billions in credit facility extended, the opportunity cost of Nvidia not doing this, and the premium they’re placing on keeping GPU ownership fragmented is recipe for massive investor love.