AI infrastructure investing used to be about counting chips. CoreWeave just changed the question.
The company’s model works like this: buy a large cluster of Nvidia GPUs → sign long-term contracts with AI customers committing to pay regardless of usage (take-or-pay contracts: agreements where the buyer pays a fixed fee even if they consume less than the full allocation) → use those contracts as collateral to borrow at investment-grade rates → use that capital to buy more GPUs.
What changed isn’t the technology. It’s the financial logic behind it.
A GPU cluster sitting in a data center is a depreciating asset. That same cluster tied to a three-year take-or-pay agreement with a creditworthy customer starts to look different: predictable cashflows, financeable structure, institutional-grade collateral. BlackRock and Blackstone have entered this market. Taiwan’s insurance companies, recently permitted to allocate to AI infrastructure, are arriving next.
Securitization — the practice of packaging predictable future cashflows into a borrowable asset — has existed for decades in mortgages and auto loans. CoreWeave applied it to server racks. That’s not a minor twist. It’s the same structural insight that turned home loans into tradeable instruments, now applied to AI infrastructure.
For investors, this reframes the central question. Pure hardware exposure captures value once, when the chip sells. Platforms positioned between AI customers and whoever wins the hardware race — locking in workflows, collecting recurring fees, holding the customer relationship — capture value continuously.
My take: The GPU does the work. The contract holder gets the tip.
Not advice. Just how I see it.