CoreWeave CEO Brannin McBee on record Q2, GPU scarcity, and why demand for 4-year-old chips is still growing

CoreWeave Q2, GPU lifespan, and the financing edge

CoreWeave posted a record second quarter, with founder Brannin McBee describing the result as a "phenomenal beat" and flagging that the back half of 2025 looks stronger still. The company operates 51 data centers and has raised north of $40 billion in debt and equity over the past 24 months.

The real bottleneck isn't power

The constraint isn't electrons — it's delivered data center capacity. McBee points to powered shell, backup battery supplies, transformers, and above all, electricians. Skilled trades take years to develop, and that pipeline won't clear quickly. Meanwhile, McBee says his inbox is flooded daily with unsolicited offers of powered land across the US. That supply isn't the problem.

On local politics, McBee is blunt: county-level pushback doesn't suppress demand, it just redirects construction elsewhere. CoreWeave is already operating in Canada and Europe, recently announced an APAC expansion, and expects to keep adding international capacity while maintaining a domestic focus.

“Q2 was a phenomenal beat for us. We have 51 data centers in operation today. We signed a contract in the quarter that goes out through 2029 for A100 SKUs — that'll be a nine-year-old chip. Pricing on A100s for us have held solid since early 2025. We've raised north of $40,000,000,000 in debt and equity over the last twenty-four months.”

The chip lifespan argument

The most pointed commercial claim McBee makes is about GPU useful life. CoreWeave signed a contract in Q2 for A100 compute — an Ampere-generation chip from 2020 — running through 2029, meaning the silicon will be nine years old by the time the deal expires. It's a fixed-price, take-or-pay structure, and McBee says A100 pricing has held since early 2025.

His argument is that different workloads suit different GPU generations, and there is no single chip that dominates every use case. CoreWeave has consistently applied a six-year depreciable life to its infrastructure and believes material useful life beyond that is realistic. The A100 contract is the clearest evidence he can point to.

The financing innovation

CoreWeave's most recent debt facility — DDTL 5, a term loan B — introduced something new for the company: the contract duration is shorter than the amortization period, meaning investors in the facility absorb renewal risk rather than being fully covered by committed contract revenue. In prior facilities, a $5 billion loan would have $7 billion in contracted revenue behind it. This one flips that ratio.

The practical effect is that CoreWeave can now compete credibly for shorter-duration enterprise contracts, which typically run at the lower end of the five-to-six year range that AI labs and hyperscalers prefer. McBee frames the ability to finance shorter contracts as proof that the enterprise segment — historically underserved by CoreWeave's longer-duration focus — is now a viable commercial target.

To staff the financing function, CoreWeave is recruiting heavily from private equity, private credit, and investment banking, with a team concentrated in New York.

Positioning

McBee's differentiation claim rests on execution — timeliness of delivery, cross-workload performance across training, fine-tuning, and inference, and what he describes as the best performance-adjusted compute available. On the broader "is this a bubble" question, his clearest rebuttal is the A100 contract: inference is profitable for CoreWeave's client base, customers are returning, and a nine-year chip commitment is a hard data point against the argument that this generation of compute has a two-to-three year useful life.