Another Tottering AI Data Center Play: Coreweave
Coreweave is a startup AI Data center play. It makes Oracle look downright healthy and athletic. While Oracle's legacy software business generates a lot of cash flow, Coreweave totally depends on external financing in the form of stock sales and debt. Moreover, Coreweave has exactly the same problem as Oracle, namely high customer risk. Coreweave's main clients are Microsoft, META, and OpenAI. Now Coreweave claims to have a $100 billion backlog. But the Devil is in the Details. No contract is valid if Coreweave cannot deliver power, GPUs, and space to fulfill them. And that is exactly the probem. Demand vastly exceeds supply with no relief in sight for years to come. Secondly, it is likely that OpenAI and Anthropic are issuing more compute requests than they actually need given their soft commitment contracts. So if you add up the backlog across all customers, it probably vastly exceeds real demand.
These key fundamental factors are the same as in the case of Oracle. An AI data centre rack's total capex including servers can total $6 million versus $56,000 to $100K for a standard interconnection data centre or telecom hotel. So capex is astronomical for a large AI data facility. But the per rack power requirements of a telecom hotel average from 5Kw to 10Kw versus as much as 132kW for its AI counterpart. While the large Interxion 2 data center in Vienna consumes 40 megawatts, an AI data center with the same footprint could consume as much 400 megawatts or even more. So AI data center operating expenses are very high due to the immense amount of energy required and steadily increases power rates.
If we look at Coreweave, its gross margin is 65%. On the surface that looks healthy. But the company depreciates GPUs using linear five year depreciation rates. That is a stretch. When META estimated one of its Llama LLMs, it experienced 419 hardware failures during the 54 day estimation run with half due to heat. It is hardly likely in these environments that GPUs will last on average five years. A large language model has trillions of neural net parameters to be estimated. GPUs do massive parallel matrix calculations in this context.
So my view is that the skyrocketing Coreweave debt plus uncertainty about its true losses are driving the stock price down. I am not alone in this assessment. Michael Burry, of Big Short fame, believes the real GPU depreciation rate is two to three years versus Coreweave's five year depreciation rule. He views Coreweave as dramatically underestimating its losses via optimistic depreciation rates. Burry is currently shorting the company and probably doing quite well given the downward stock trajectory. It is worth noting that Coreweave's debts total 3.5x its annual revenues. This is a severe burden for a company that has no positive operating cash flows.


Comments
Post a Comment