Explainer · Neoclouds

What are neoclouds? NBIS, CRWV and IREN explained

6 August 2026 · 10 min read

A neocloud is a cloud provider built specifically for AI workloads — training, inference and agentic systems — rather than the general-purpose computing that Amazon, Microsoft and Google were designed around. The customer is not renting servers or floor space. They are renting scarce accelerated compute, usually on multi-year contracts, usually because they cannot get it fast enough anywhere else.

Three names dominate the listed universe: Nebius (NBIS), CoreWeave (CRWV) and IREN. Nvidia holds equity in all three. That fact is either the strongest validation in the sector or its central weakness, depending on who you ask — and both cases are worth understanding.

Why neoclouds exist at all

Accelerated compute has four separate bottlenecks, and they are not the same problem:

Power. Grid-connected capacity at scale takes years to permit and build.
Datacentre delivery. Turning megawatts into functioning halls with cooling and networking.
GPU supply. Access to allocation, at the front of a long queue.
Operational software. Keeping enormous clusters actually utilised rather than idle.

A company controlling one of these can be valuable. A company controlling several is a different business — and should be valued differently. That distinction is the single most useful lens for reading this sector, because the market frequently applies one multiple to companies solving very different problems.

CoreWeave and Nebius sit closer to the cloud layer. Applied Digital, TeraWulf and Cipher Mining sit closer to the power-and-campus layer. IREN is trying to bridge both.

The accounting reason hyperscalers use them

This is the part most explainers skip, and it explains far more of the demand than "they ran out of GPUs".

When a hyperscaler builds its own capacity, the spending lands on the balance sheet as capital expenditure and weighs on free cash flow. When it rents that same capacity from a neocloud, the cost is recognised as operating expense spread across the contract's life.

Meta is the clearest example. Analysts expect roughly $136 billion in cash from operations in 2026 against stated capex guidance of $125-145 billion — potentially free-cash-flow negative before you count anything else. Meta also holds up to $62.2 billion in neocloud agreements, with CoreWeave and Nebius contracts extending into 2031-2032. Built in-house, that would be balance-sheet capex. Rented, it averages under $10 billion a year in opex.

Neoclouds are not only a capacity solution. They are a balance-sheet solution, and that demand driver persists even if GPU supply loosens.

The three names, and how they actually differ

CoreWeave (CRWV)

The purest listed neocloud — the business was built around selling AI cloud capacity to labs, hyperscalers and enterprises. It has secured more than 3.5 GW of contracted power and is targeting roughly 1.7 GW active by the end of 2026. Its OpenAI relationship alone is reported at around $22.4 billion in contracts, and 2026 capex guidance runs to about $35 billion.

The model is capital-intensive and balance-sheet-owned: CoreWeave buys the GPUs, funds them with debt, and carries the depreciation. Customer concentration and refinancing costs are the two things to watch.

Nebius (NBIS)

Relisted on Nasdaq in 2024 after spinning out of Yandex's cloud infrastructure division, Nebius is building a vertically integrated AI cloud rather than a general-purpose wrapper. It has also secured around 3.5 GW of contracted power and targets 800 MW to 1 GW connected.

The meaningful difference is financing. Nebius has pivoted toward an asset-light model, in contrast with the debt- and dilution-heavy balance-sheet ownership used by CoreWeave and the miner-conversion names. In a sector where the cost of capital is the binding constraint, that structural choice matters more than any quarter's revenue.

IREN

The clearest miner-to-AI conversion story, and arguably the most directly validated. IREN signed a five-year, $3.4 billion AI cloud contract with Nvidia — roughly $680 million a year for 60 MW at its Childress, Texas campus — alongside a broader infrastructure partnership.

Its genuinely scarce asset is what bitcoin mining accidentally accumulated: grid-connected power at scale, already permitted. That is years of lead time somebody else does not have to wait through.

Nvidia's position, and the circular financing debate

Nvidia holds equity in all three: a 9.3% stake in Nebius, 47.2 million shares in CoreWeave, and discussions around a potential $2.1 billion stake in IREN vesting toward 600,000 GPUs. It has also been reported in talks over a $250 billion financing guarantee tied to OpenAI's data centre project.

The bear case is straightforward. Nvidia sells the GPUs, takes equity in the buyer, sometimes receives a share of the cloud revenue those GPUs generate, and in places backstops the financing. Critics argue this makes demand look more independent than it is — the supplier is helping fund its own order book.

The counter-argument is arithmetic. Nvidia's roughly $2 billion into CoreWeave is around 5.7% of a single year's $35 billion capex. The other 94% comes from lenders and investors with no stake in Nvidia's revenue. If the entire sector rested on vendor financing, that ratio would look very different.

The honest read sits between them. The equity stakes are real and the conflict is real; they are also small relative to the capital actually flowing, and the contracted backlog comes from customers with no Nvidia relationship. Anyone presenting either side as settled is arguing rather than analysing.

Why the contracts came and the stocks fell anyway

The most instructive thing about mid-2026 is the disconnect. IREN announced a $3.4 billion Nvidia contract and up to $2.1 billion in investment — and the stock fell 37% over the following month. Nebius dropped 13% in a session. Core Scientific, TeraWulf and Applied Digital all slid, with Applied Digital sliding despite a 61% fiscal Q3 revenue beat.

Announcements stopped moving these stocks because the market has shifted the question. It is no longer "can they win contracts" — they demonstrably can. It is now:

How is this funded, and at what cost? Debt-funded GPU fleets are exposed to rates and refinancing.
How much dilution arrives before the capacity does? Equity raises to fund buildouts transfer value from existing holders.
How fast does contracted power become active power? Backlog earns nothing until megawatts are energised. Most contracted capacity across the sector is not yet online.
What is the real useful life of an accelerator? Depreciation schedules drive reported earnings, and assumptions vary between operators.

That is a maturing sector, not a failing one. But it means announcement-chasing has stopped working, which is precisely why watching where attention concentrates has become more useful than watching press releases.

What to actually watch

Contracted power versus active power. The gap between the two is the single most important number in this sector.
Contract duration and counterparty credit. A five-year contract with a creditworthy buyer is a different asset from announced megawatts.
Financing structure. Asset-light versus balance-sheet-owned determines who survives a rate move.
Depreciation assumptions. Buried in the filings, and they move earnings more than most headlines do.
Customer concentration. Large contracts cut both ways.

See live prices and sentiment across the neocloud group →

None of the above is a recommendation and it does not account for your circumstances. Figures are drawn from the reporting linked below and were accurate when published.

Frequently asked questions

What is a neocloud?

A cloud provider built specifically for AI workloads — training, inference and agentic systems — rather than general-purpose computing. Customers rent scarce accelerated compute, usually on multi-year contracts. The listed names include Nebius, CoreWeave and IREN.

What is the difference between a neocloud and AWS or Azure?

Specialisation. Traditional hyperscalers were built for general computing and offer AI capacity as one service among many. Neoclouds are built entirely around GPU-heavy workloads, which lets them move faster on capacity and often achieve higher utilisation. Hyperscalers are also neocloud customers — Meta holds up to $62.2 billion in such agreements.

Does Nvidia own part of CoreWeave, Nebius and IREN?

Yes. Nvidia holds a 9.3% stake in Nebius and 47.2 million shares in CoreWeave, and has been in discussions over a potential $2.1 billion stake in IREN vesting toward 600,000 GPUs. It also signed a five-year, $3.4 billion AI cloud contract with IREN.

What is circular financing in AI?

The concern that Nvidia sells GPUs, takes equity in the buyers, sometimes receives a share of the resulting cloud revenue, and in places backstops financing — making demand appear more independent than it is. The counter-argument is scale: Nvidia's roughly $2 billion in CoreWeave is about 5.7% of a single year's $35 billion capex, with the rest from unrelated lenders and investors.

Why did neocloud stocks fall despite winning contracts?

The market's question changed. Winning contracts is no longer in doubt; funding them is. Debt costs, equity dilution, the lag between contracted and active power, and depreciation assumptions on accelerators now drive the stocks more than announcements do.

Which neocloud stocks are listed?

Nebius (NBIS), CoreWeave (CRWV) and IREN are the primary names. Applied Digital (APLD), TeraWulf (WULF), Cipher Mining (CIFR) and Core Scientific sit closer to the power-and-campus layer, generally converting sites originally developed for bitcoin mining.