Why CoreWeave Must Escape the Intermediary Trap or Die
CoreWeave sits on a brittle capital stack, with a wafer-thin moat, and a clock that's ticking fast
Introduction
I’ve written about CoreWeave before. I think the market is misjudging the company. I don’t see a durable moat. It owns no IP. It doesn’t own real estate. It sits between Nvidia, its supplier, and its customers, which are thus far mainly Microsoft and OpenAI. Whatever margin it is able to extract from that precarious position is consumed by lease and utility payments.
It has become one of the fastest-scaling GPU cloud providers in the world. It leases real estate, secures power in deregulated markets, negotiates chip access with Nvidia, and sells compute capacity to model labs and startups starved for silicon.
On paper, it’s the perfect company for this moment: a nimble arbitrageur of a global GPU famine.
But structurally, CoreWeave is deeply fragile.
It doesn’t own the chips. It doesn’t own the land. It doesn’t generate the power. It doesn’t have proprietary models. And critically, it doesn’t have long-term customer lock-in. It is, at bottom, an intermediary, and one whose margins can be compressed from both directions. Nvidia can squeeze them upstream. Landlords and utilities can extract rent downstream. And customers, once they scale, can walk.
CoreWeave is not a moat. It’s a timing arbitrage. And that window is closing.
I. No Hard Assets, No Moat
Strip away the hype and CoreWeave’s business model is clear:
Capex-light: It leases data centers rather than owning them
Power-dependent: It contracts with utilities but doesn’t generate power
Chip-reliant: It secures GPUs via forward contracts with Nvidia
Customer-disposable: It provides infrastructure, not differentiated services
That chip reliance deserves emphasis. CoreWeave’s differentiation hinges on its access to Nvidia’s most advanced GPUs. This access is secured not through spot purchases, but through forward allocation agreements. These are pre-purchase deals made months in advance, often involving volume guarantees and strategic alignment. In exchange, CoreWeave gets supply priority in a world where availability is gold.
But this is a shaky foundation. CoreWeave is built on a silicon substrate it doesn’t control. Nvidia owns the chip supply chain, the software stack (CUDA), and the long-term customer relationships. At any time, it can:
Raise prices
Redirect inventory to hyperscalers
Launch its own hosted compute platform
Or vertically integrate into CoreWeave’s turf
Meanwhile:
Landlords can raise rent as leases expire
Utilities can reprioritize or renegotiate power deals
Customers can move to AWS, Azure, or self-hosted GPU clusters as capacity normalizes
With no silicon, no land, no power generation, and no IP moat, CoreWeave is a margin sandwich.
II. The Mirage of Speed
Much of CoreWeave’s current advantage stems from a simple fact: it moves faster than the incumbents.
Hyperscalers are slow. Permitting is a bottleneck. Substation lead times are measured in years. CoreWeave wins by cutting deals with real estate developers, retrofitting warehouses, and tolerating fragility that AWS or Azure wouldn’t accept.
But this is a temporal arbitrage, not a structural one. It works only as long as:
Hyperscalers are still building
Sovereigns haven’t fully entered the game
Nvidia continues to prioritize them over others
AI demand outstrips infrastructure growth
Once that gap closes, and it will, CoreWeave’s speed becomes less relevant. At that point, hyperscalers offer better reliability, deeper integration, and far more durable supply chains. The temporary speed premium evaporates.
III. Can Financialization Save Them?
CoreWeave’s best (and perhaps only) strategic escape hatch is to financialize compute.
If they can:
Create standardized, tradable GPU lease contracts
Become a central exchange for GPU capacity
Package GPU exposure into yield-bearing financial products
Act as a custodian or clearinghouse for compute liquidity
...then they evolve into more than a datacenter operator. They become market infrastructure: the CME or ICE of AI compute.
But this path is fraught.
They don’t control the chips. Nvidia does. You can’t financialize what you don’t own.
They don’t own the power. Volatility in power pricing undermines asset pricing.
They don’t have collateral strength. Warehousing risk on an asset you lease from others is dangerous.
They invite disintermediation. The more liquid the GPU market becomes, the easier it is for customers to go direct.
In short: financialization might be CoreWeave’s best shot at escaping the intermediary trap, but it could just as easily accelerate their obsolescence.
IV. The Inevitable Squeeze
CoreWeave is not a sovereign utility. It is not a chipmaker. It is not a hyperscaler. It is an interim solution. It is a well-funded middleman exploiting temporary asymmetries in supply and demand.
Unless it does something radical, like own land, own power generation, vertically integrate into model training, or successfully institutionalize GPU-backed finance, it will be squeezed to death.
Today, it surfs the wave. But the water level is dropping.
Nvidia upstream. Landlords and utilities downstream. Hyperscalers all around. The intermediary trap is closing, and CoreWeave is standing at the center of it.