The Neocloud Was a Moment, Not a Moat
Compute demand is real and enormous. That was never the question.
The question is who ends up owning it — and who merely passes it through.
There is a kind of company that appears at the seam of a technological dislocation, thrives because the seam is new, and is quietly absorbed once it closes. The neocloud is that company. It bought Nvidia’s chips and rented them to the AI labs at the moment the labs could neither finance the silicon nor build the buildings to house it — and for three years it was one of the great trades of the cycle. CoreWeave took its cost of borrowing from roughly SOFR plus a thousand basis points in 2023 to an investment-grade SOFR-plus-225 by early 2026, against the very same chips whose rental price was, over the same window, falling from about eight dollars an hour toward three.
The trade is not ending because the demand is fake; the demand is the realest thing in the economy. It is ending because the neocloud was never really a cloud. It was a financing-and-development structure in a cloud’s clothes, assembled to bridge a specific and temporary gap — and the thesis here is narrower than the bubble-callers’. It is not that neoclouds die. It is that the pure pass-through — the entity whose only job is to stand between the lab and the chip and route someone else’s risk on someone else’s credit — does not survive the gap closing. The ones that own something, that build and hold real infrastructure, turn into something else and live. The question was never whether the compute gets built. It is who ends up owning it.
To see why, start with the problem the neocloud was invented to solve.
Contents
I. 2023: Enter the Neocloud
II. What Changed
III. What’s Left for the Neocloud
IV. What Does the Future Look Like?
I. 2023: Enter the Neocloud
It is 2023. ChatGPT has detonated and the demand for training compute has gone vertical. The frontier labs need GPUs at a scale that breaks the old way of buying them. The constraint is not ambition. The constraint is the balance sheet — and, more precisely, five things a lab cannot or will not do.
1. It will not open its books. The lab is private, valued on a software multiple, and has just tripled that valuation; the last thing it wants mid-repricing is to show the public — or its competitors — its true financials and balance sheet, least of all the mountain of depreciating silicon underneath the software-margin story. Capital intensity is not a fact it wants to disclose. It is a fact it wants to put somewhere else.
2. It cannot finance the chips. No bank in 2023 writes a term loan against a warehouse of H100s for a company with no profits, no history of operating a cluster, and an asset with no residual-value curve — no comparable, no recovery model, no precedent. To a lender who doesn’t already understand them, the chips are uncollateralizable.
3. It does not know how to build the thing. Financing the silicon is one problem; standing up the data center is another entirely. Site selection, power procurement, the interconnection queue, the shell, the modular power and cooling units, the MEP, the substation — none of this is a lab’s competence. It is a construction-and-development company’s competence.
4. It cannot move fast enough. Even if a lab could build, starting from zero costs it years it does not have. In a land-grab for compute, time-to-energization is the whole game — and the neocloud already knows how to find a site with power and stand it up at speed. That head start is its own kind of scarcity.
5. It has no privileged access to Nvidia. In a supply-constrained market, getting the chips at all is its own problem, and the lab is at the back of the line.
Enter the neocloud. It secures the Nvidia allocation — because Nvidia wants it to exist and steers silicon toward it. It raises the capital and buys the chips. It finds the site, builds the cluster fast, runs it, and signs the lab to a multi-year take-or-pay. The GPUs now sit on the neocloud’s balance sheet, not the lab’s. The capital intensity is the neocloud’s problem. The build risk, the operating risk, the duration mismatch between a five-year lease and a three-year useful life — all of it lives at the neocloud. The lab gets compute as a contracted service cost and keeps its books to itself.
That is the whole proposition, and in 2023 it is a genuinely good one. The neocloud is a risk-absorption machine. It takes the things the lab cannot or will not hold — disclosure, financing, development, speed, allocation, residual-value risk — and holds them, for a spread.
But notice the one thing that makes any of it possible. None of this debt was financeable on the strength of the chips. There was no financing solution for these data centers without Nvidia standing behind them — and Nvidia has said so, on the record. Asked on Dwarkesh Patel’s podcast why it doesn’t simply become a cloud itself, Jensen Huang gave the doctrine — “do as much as needed, as little as possible” — and then named the cohort and its provenance directly:
“If we didn’t help CoreWeave exist, they would not exist. If we didn’t support Nscale, they wouldn’t be where they are today. If we didn’t support Nebius, they wouldn’t be what they are today.”
That is not a boast about a portfolio. It is a description of a category that exists because its supplier decided to let it exist — a place to park the risk Nvidia had priced out and the capital intensity the labs had walked away from. So stack up what the neocloud supplied: balance-sheet relief, financing opacity, the Nvidia wrap, turnkey development, speed, and risk absorption. Every one of those was scarce. And here is the mistake the bull case makes — it reads that scarcity as a set of company advantages. It was nothing of the kind.
They were market-structure advantages. They existed because the buyer was private and capital-constrained, the supplier was rationing, the asset was unfinanceable, and the development market was immature. Change those facts and the neocloud does not lose one moat at a time. It loses the environment that made the moats possible. None of it was permanent — and since 2023, every piece has begun to give way at once.
II. What Changed
The gap is closing from every side, and the place you can watch it close first is in how a data center actually gets financed today. Three years ago, GPU infrastructure was financed in the dark, by a handful of private-credit funds willing to underwrite something nobody understood. That has changed with startling speed — and as it changed, it exposed what the neocloud was actually for.
The market learned to finance this in the open
Data-center paper now prints under Rule 144A and Reg S, in ABS and CMBS structures, sold to the same investors who buy any other piece of structured credit; well over fifty billion dollars of it has been issued since 2021. Moody’s, S&P, KBRA, and Fitch have all published dedicated data-center rating criteria. And every one of those methodologies converges on the same principle — the principle that quietly dismantles the neocloud thesis: rate the tenant, not the building.
But the tenant usually can’t carry the debt
Here is the catch, and it is the heart of the matter. The frontier labs — the actual tenants — are not investment grade. Anthropic, in the words of the reporting, is “a startup without a strong credit rating.” OpenAI is no different; it spent much of 2025 trying to finance its own data centers and could not get competitive terms, which is precisely why it pivoted to controlling the hardware inside other people’s buildings instead. And the valuations are beside the point: a three-hundred-billion-dollar markup is equity, not credit. A lender underwriting a fifteen-year lease cannot lean on an equity valuation — it needs someone whose balance sheet will still service the rent if the tenant defaults. The lab’s lease does not get financed on the lab’s credit. It gets financed on someone else’s.
Which is what a credit wrap is — and who actually supplies it
A credit wrap substitutes a creditworthy third party’s balance sheet for the tenant’s, so the debt prices and rates against the strong name. The suppliers who make sense are the ones with a strategic reason to stand behind the lease — the hyperscalers whose own silicon is going inside (Google behind TPU deployments, Amazon behind Trainium), Microsoft, and Nvidia behind its GPUs. The cleanest live example is Anthropic’s: it is putting TPUs into a Louisiana data center leased by Fluidstack, and Google agreed to backstop Fluidstack’s lease payments. Per people involved, the Google guarantee was the thing that secured the financing. The mechanical version of the same move is already in the bond market:
Hut 8 / Fluidstack, River Bend, Louisiana
>
$3.25B senior secured notes · rated BBB– · priced near 6% — on Google novation rights and a springing Alphabet guarantee. As one observer put it: Google credit dressed in a project-finance wrapper. The building is incidental; the tenant is the bond.
So what does the neocloud actually do?
It is the avenue through which the lab reaches the wrap. It does not supply the credit — Google or Amazon or Nvidia does. And increasingly it does not supply anything the lab couldn’t arrange directly. Read CoreWeave’s cost-of-capital story in that light. The compression from roughly SOFR-plus-1,000 in 2023 to an investment-grade SOFR-plus-225 in 2026 — the first IG-rated GPU-backed financing — was not the chips getting better. It was the structure learning to route the deal to a creditworthy backstop, with Nvidia’s own residual-capacity commitment sitting behind the unsold portion as the all-purpose patch. The collateral was never really the silicon. It was someone else’s balance sheet, rented for the occasion, with the neocloud taking a clip in the middle.
And the labs are moving to cut the middle out
The conduit is temporary, because the people who need it are building the capability in-house. Both leading labs are standing up their own data-center and infrastructure-finance teams to arrange these structures — and run the development — themselves:
Anthropic — hired Tim Hughes (CDO, ex-Stack Infrastructure), Brett Rogers and Winnie Leung (both ex-Google data centers); internally targeting 10+ GW; plans to start leasing its own data centers; now hiring an infrastructure-finance lead to own per-site lease and power economics.
>
OpenAI — hired Chris Malone (Head of Data Centers, ex-Meta/Google) and Keith Heyde (ex-Meta); recruiting DC talent since late 2024; on the record that “it’s important for us to also self-build.”
Once a lab can arrange its own wrap and run its own development, the neocloud’s role as intermediary is redundant. And remember what the intermediary was supposedly there to build in the first place: on a gigawatt-scale AI data center the GPUs are roughly forty percent of the cost, networking — also Nvidia — another thirteen, and the remaining forty-odd percent is shell, power, substation, and cooling. That non-GPU half was the neocloud’s real contribution, the turnkey development the lab couldn’t do itself. But standardized triple-net leases, independent-engineer reports as a condition precedent, and a developer base that increasingly looks like ordinary infrastructure have turned site-and-power-and-build from arcane knowledge into a repeatable process.
Both halves of the proposition — the financing and the building — are being pulled in-house and competed away in the same eighteen months. That is the phase change.
III. What’s Left for the Neocloud
Lay the whole thing out and the pattern is hard to miss. Five of the six advantages the neocloud was built on were facts about the market of a moment. Only one — operational competence — is a fact about a business.
Operational competence — actually running the densest, newest silicon at scale and deploying it fast — is real, and it is worth real money. It is simply not a moat. It is the margin of a very good contractor, not the equity story of a platform, and it competes toward cost-plus the instant the supply constraint loosens. Everything else on the list was on loan from the market structure of 2023.
The Nvidia wrap is the sharpest case, because it is the most important asset the neocloud has and the one most clearly not its own: lent revocably by a supplier who has said on the record that he provides exactly as much as needed and not a dollar more, and whose actual incentive is not to protect any one intermediary’s margin but to maximize the absorption of GPUs across the whole ecosystem. But do not mistake this for a story about Nvidia alone. A credit wrap is what every sub-investment-grade tenant needs, regardless of how large its valuation has grown — and the labs will be signing a great many of these leases. The wrap is structural, not incidental; and the party that supplies it — Google, Amazon, Microsoft, Nvidia — is the party with the leverage, not the neocloud routing the paperwork between them.
A moat your supplier owns and rents to you is not a moat. It is a lease, and you do not control the renewal.
IV. What Does the Future Look Like?
The wrong question is whether neoclouds survive. The right question is where the rent goes — because the compute is not going anywhere; it is going to be enormous; the only thing in motion is on whose balance sheet it sits and who collects the margin. And the rent is leaving the middle in four directions at once.
It is moving upstream, to Nvidia. The wrap exists because Nvidia chose to let the financialization layer exist; the long-term memory and foundry commitments, the near-monopoly on training silicon, and the allocation power all sit with the supplier, and the supplier can withdraw the support that built the category as easily as it extended it.
It is moving sideways, to custom silicon. The neocloud is, by construction, almost entirely a bet on Nvidia — and the architecture is fragmenting underneath it. Anthropic has committed to up to a million Google TPUs and multiple gigawatts of Amazon Trainium; OpenAI and Broadcom have announced ten gigawatts of custom accelerators; Meta has its MTIA roadmap and Microsoft its Maia. Broadcom is carrying a backlog that implies a hundred-billion-dollar AI run rate within a couple of years, and custom-ASIC server shipments are projected to outgrow merchant GPUs for the first time. None of that flows through a Nvidia-only intermediary. The hyperscalers and labs that can design their own silicon have every reason to route their inference around it.
It is moving downstream, to power and land — and this is the big one. Roughly twenty-six hundred gigawatts of generation sit stuck in interconnection queues; median wait times have stretched toward half a decade. Power, not chips, is now the binding constraint, and the value is accruing to whoever controls megawatts. Independent power producers and behind-the-meter dealmakers have become the new long-duration asset class of the build-out: a restarted nuclear unit under a twenty-year hyperscaler PPA, an eighteen-billion-dollar grid-connected nuclear deal, frackers and gas operators pivoting their capacity to data-center load. The underwriting itself has shifted from sticks and bricks to megawatts. As long as power is the limiting factor, the people who own power are well-placed — and the neocloud, by default, owns none of it.
And it is moving back to the demand itself, as the labs and hyperscalers vertically integrate. They are building around the neocloud, not through it. Microsoft alone paused multiple gigawatts of capacity it would otherwise have rented from intermediaries once it decided to build its own. Stargate, Prometheus and Hyperion, Colossus — the largest buyers are becoming the largest builders. The bull case was always the surge gap, the window in which hyperscaler demand outruns hyperscaler self-supply. That window is real today. It narrows from here.
Squeezed from four sides at once, the neocloud has exactly one escape, and the better ones have already found it: stop being a pure intermediary and become one of the control points. Become a power developer. Become a builder. Become a vertically integrated infrastructure platform that owns something the labs and the supplier cannot route around. This is precisely what the survivors are doing — Crusoe, which started in flare-gas mitigation and now operates a gigawatt-plus flagship with billions in dedicated generation behind it; Nebius, driving toward gigawatts of contracted power and signing its own behind-the-meter supply; Nscale, securing Norwegian hydro; CoreWeave itself moving into self-build, targeting many gigawatts of owned capacity by the end of the decade, bankrolled by a fresh Nvidia equity check.
But look closely at what that migration is. It is the neocloud ceasing to be a neocloud — climbing out of the closing gap and onto one of the banks. The pure GPU-as-a-service intermediary — the thing the thirty-billion-dollar debt stack was actually built on — is the single most exposed link in the chain. The bear case was never “everyone called a neocloud dies.” It is narrower and harder to dodge: the middleman, as a middleman, does not get to keep the rent.
So return to where we started. The demand is real and the numbers are staggering, and nothing here is a bet against AI infrastructure; the world will spend more on it than almost anyone forecasting it has the nerve to say out loud. The question was only ever where that compute lives. The neocloud was a bridge thrown across a specific gap in 2023 — private capital-constrained buyers, unfinanceable chips, scarce allocation, an immature development market. Every plank of that bridge is being pulled up. The buyers go public. The chips get comparables. The allocation stays with the supplier. The development standardizes. The power becomes the prize.
The compute ends up where it was always going to end up: with the people who own the demand — the labs, going public and integrating vertically — and the people who own the land and the power — the developers, the IPPs, the hyperscalers building to suit. The neocloud stood between them and was paid, handsomely and for a while, to absorb what neither side wanted to hold. That was a real service. It was just never a destination.
You do not build a durable franchise by standing in the middle of a gap that is closing. Bridges are valuable right up until both banks decide they’d rather just touch.