Thoughts on CoreWeave
What’s the risk profile of CoreWeave? Why has capital flocked to the company when it has seemingly high risk (at least from the outside)?
CoreWeave has been given preferential treatment by Nvidia due to their relationship that dates back to when CoreWeave was a crypto miner. CoreWeave’s GPU capacity allowed them to secure multi-year compute contracts with Microsoft and likely other hyperscalers1. This in turn got BlackRock, Carlyle, and Blackstone comfortable enough to give them GPU-backed loans at single digit interest rates.
Yet, Nvidia controls GPU prices. Their supply, product iteration, and CUDA platform gives them this control. It’s reminiscent of the early days of OPEC when its control over global oil supply allowed it to influence prices by setting collective production quotas.2
However, just like OPEC losing control of the market as advanced oil extraction techniques emerged, Nvidia is exposed to similar risks. While AMD and Intel are struggling to gain market share, TPUs are getting better by the generation and ARM-based chips continue to be adopted.
Unlike the oil market, there are limited hedging methods in the GPU market. In addition to contracts with Microsoft, CoreWeave needs to enter into contracts with emerging startups to increase utilization. There are no guarantees that these startups will still be around in a couple of years. Despite this counterparty risk, CoreWeave doesn’t have financial instruments like collars and swaps to hedge their risk. The risk could very well be small if most of their compute is offloaded to hyperscalers, but that puts CoreWeave at a different type of counterparty risk where too few customers contribute to a large part of their revenue. It’s especially worrisome since these hyperscalers are technically competitors that can build data centers for 2-3x cheaper.
What makes risk more difficult to assess is their GPU-backed loans. CoreWeave is using its GPUs as collateral, making them more vulnerable to margin calls and ensuing domino effects if prices fall. The stamp of approval by BlackRock et al. gives an illusion of safety for traditional infrastructure investors who want AI exposure. This in turn injects more equity investment into the company that, when combined with the cash CoreWeave is generating from customers, improves debt metrics. This allows CoreWeave to take out even more loans and purchase more GPUs. It appears that CoreWeave has been efficient in executing this business model in the past year or so, and as the latest valuation suggests, investors are giving quite a high multiple and assuming that the pace of growth will continue. But the market is heating up from all sides for CoreWeave, including competition for powered shells and energy generation/transmission.
As an outsider looking in, I can’t help but wonder what I may be missing here.
Despite hyperscalers having their own GPUs, when multiple internal teams are competing for compute time, it’s better to outsource some of the work to third-party cloud providers like CoreWeave so that there aren’t
Specifically during a time period when OPEC controlled >55% of the global oil supply. That number has been fluctuating between 30% and 40% in recent years, and OPEC has much less influence today