Bloomberg Tech

Michael Intrator on why AI compute is sold out for years

Michael Intrator· Co-founder and CEO of CoreWeave at CoreWeave
·~7 min·English·Bloomberg
AI InfrastructureGPUBusiness StrategyInferenceTraining
TL;DR

CoreWeave CEO Michael Intrator argues demand for AI compute is structurally durable — not cooled by paced model releases or data-center moratoriums — even as his NVIDIA infrastructure stays sold out for years and customers pull the company up the stack from raw GPUs into AI tooling.

01Moving Up The Stack

Pulled Up The Stack

CoreWeave says its own customers are pulling it beyond renting GPUs into cloud services and developer tooling like Forge, to solve the distinct problems that come with artificial intelligence.

the systematic pull by the clients that use our infrastructure up the stack. And they are looking for a cloud solution to the distinct and unique problems that are associated with artificial intelligence.

— Michael Intrator, Bloomberg Tech
Key Insight
The pull up the stack framing is a margin story in disguise. Renting GPUs commoditizes toward the price of power and silicon, so CoreWeave has to climb toward tooling, where switching costs and differentiation live. Forge is less a product launch than a hedge against its own base layer becoming a commodity.
02Positioning

The AI Cloud, Not a Neo Cloud

Intrator rejects the neo cloud label as a point-in-time description and casts CoreWeave as a purpose-built AI cloud, engineered from the ground up for one kind of computer and meant to outlast the current buildout.

Neo Cloud is something that exists at a point in time, but what we have built is the cloud solution for artificial intelligence.

— Michael Intrator
Key Insight
Calling itself the AI cloud rather than a neo cloud is a durability claim aimed at investors. Neo cloud implies a rentals wave that could crest, so the rebrand asserts the business survives even if the current capital-expenditure boom slows. The ground-up specialization is the defensibility argument underneath the branding.
03The Demand Narrative

Pacing Does Not Kill Demand

The story that pacing frontier model releases will cool compute demand is, in Intrator's words, deeply misunderstood, and he says the procurement of infrastructure keeps coming regardless.

So, I think that narrative is misunderstood, deeply misunderstood, actually.

— Michael Intrator
Key Insight
Intrator quietly redefines pacing from slow down to stay ahead safely, which conveniently leaves demand intact. The deeper point is that compute demand is decoupled from release cadence: training runs, serving existing products, and new entrants all consume GPUs whether or not the frontier labs ship the next model this quarter.
04Siting And Power

Moratoriums Move The Map, Not The Demand

Data-center moratoriums do not shrink the demand for compute, Intrator argues; they only relocate where it gets built.

moratoriums do not alter the demand for compute. They simply move where the computer is going to be built.

— Michael Intrator
Key Insight
This reframes moratoriums as a NIMBY lever, not a demand lever: they change the map, not the total. The tell is the enter communities early posture. The binding constraint on AI is now power and land, so the winner is whoever is the least-resisted neighbor, not just whoever has the most chips.
05Supply vs Demand

Sold Out For Years

CoreWeave is largely sold out of NVIDIA capacity for years; Intrator says he watches alternative accelerators but cannot build enough NVIDIA infrastructure to meet demand.

We are overwhelmed. We are largely sold out. We are sold out, you know, over time for several years.

— Michael Intrator
Key Insight
Being sold out is itself the answer to the non-NVIDIA question. Intrator meets the AMD and custom-silicon pressure with demand rather than deployment specifics: he monitors other accelerators but leads with how far demand outruns NVIDIA supply, which leaves CoreWeave's actual move on alternative chips unresolved. The years-long backlog is also his strongest rebuttal to the demand-is-softening narrative from earlier in the same interview.