AI 101 — What is a neocloud?

A neocloud is a cloud company built to rent one thing — GPU computing power for AI — instead of the hundreds of unrelated services a hyperscaler sells.

The name is shorthand for "new cloud," and the distinction it draws is narrowness, not size. Amazon, Microsoft and Google run general-purpose clouds, where compute is one item on a menu that also carries databases, email, identity, storage tiers and a hundred other products. The companies the industry now files under neocloud — CoreWeave, Nebius, Lambda, Crusoe, Nscale, IREN — mostly sell GPU-hours, the networking that ties thousands of GPUs into one machine, and the storage close enough to feed them.

This is a different question from how the buildings work — What is an AI data center? covers the halls, the power and the cooling. This piece is about the business stacked on top of the building, where the money and the risk both sit.

Why it matters right now

On November 6, CNBC reported that Microsoft is winding down a 2023 agreement to rent AI capacity from Oracle and consolidating the roster of neoclouds it counts as suppliers, citing two people briefed on the talks. Microsoft says it has not decided which suppliers it will stop using. The direction is clear anyway: a buyer that spent three years signing capacity deals has started doing unit economics.

The market moved first. CoreWeave, Nebius and IREN shares are down roughly half over the past month, CNBC noted in the same report. The financials show why the group is fragile in a way that revenue growth hides. CoreWeave's second quarter, reported in August, brought record revenue of $2.575 billion, up 112% from a year earlier — alongside a depreciation and amortisation charge of $1.393 billion and a net loss of $626 million.

That is the neocloud in one line: fast growth, real gross profit, and an enormous, unavoidable bill for hardware wearing out.

The mental model: three scarce inputs, one spread

A neocloud is a spread business. It needs three things that are all hard to get: chips, power and capital. It buys Nvidia systems (or builds on someone else's), puts them in a hall with enough electricity and cooling to run them flat out, and sells access by the hour or through multi-year reserved contracts. Profit is the gap between the price of a GPU-hour and the cost of owning the GPU — purchase price, electricity, staff, and the interest on whatever debt paid for the silicon.

Everything else about the category follows from how those three inputs are financed. Neoclouds rarely buy chips with cash on hand; they borrow against the chips and against the customer contracts that will pay for them. That works while GPU prices hold and contracts renew. It stops working the moment either assumption slips, which is why the same companies can post 100% growth and a halved share price in the same quarter.

The analogy: a van rental company

Imagine buying 1,000 delivery vans on a loan and renting them out by the day. Good months, the vans are busy and the rent covers the loan with room to spare. Bad months, a newer model makes yours expensive to run, so the day rate you can charge falls — while the loan payment does not. And if one customer rented 60 of those vans, that customer's renewal is basically the business.

Swap the vans for H100s and you have the neocloud. The hardware depreciates on a schedule the company chooses, but the market decides how fast it really loses value, one chip generation at a time. The customer list is short, the contracts are large, and the debt is secured on both.

Scale is the other half of the picture. American Compute's directory counted 536 companies selling GPU capacity outside the hyperscalers as of October 2026, and only 147 of them are neoclouds in the narrow sense — the rest are telecoms, national supercomputing centres, former bitcoin miners and power developers who added GPUs to a business they already had, per its State of the Neoclouds tally. Ten operators hold 87% of the disclosed funding. Everyone else is small, debt-funded, and one customer away from a very bad year.

Why they are cheaper, and why the rate is not the bill

The price gap is real and large. In SAVRN Research's tracker, an H100 hour costs $1.99 to $6.16 at neoclouds and $10.00 to $12.29 at hyperscalers, as of late October, per its GPU price index. The reasons are structural: neoclouds skip the virtualisation layer that eats performance, they do not have to fund a catalogue of unrelated services, they tend to take new silicon early, and they run in fewer places with fewer regions to replicate.

The catch is that an hourly rate is not a bill. Reserved-capacity terms, storage, data egress, interconnect and minimum commitments all move the total, and a cheap rate on older hardware is cheap because it is older hardware. The comparison worth making is the five-year cost of a workload, not the sticker on an H100.

Common misconceptions

"Neocloud means small cloud." It means narrow. CoreWeave reported a revenue backlog of $104 billion at the end of June, a number larger than most AI commitments disclosed by companies ten times its age.

"They compete with AWS." Mostly they supply it. Hyperscalers rent from neoclouds when demand arrives faster than they can build, which is why Microsoft was CoreWeave's biggest customer in the year it listed — 62% of 2024 revenue, by CoreWeave's own IPO filing. That is a channel, not a rivalry. It is also the concentration risk in a single sentence.

"The GPU is the asset." The contract is the asset. A GPU is a depreciating box whose rental price falls when the next generation ships.

"Depreciation is an accounting detail." It is the product going obsolete, priced in. When a neocloud's depreciation line grows faster than revenue, the company is not getting more efficient — it is buying more boxes to stand still.

Where to learn more

CoreWeave's quarterly filings are the clearest public window into the model, because it is large enough to have to disclose customer concentration and depreciation. American Compute publishes the best count of who exists and how they are financed. For the technical layer underneath, our explainers on what HBM does inside a GPU and what AI inference actually is pick up where this one stops.

If you rent GPU capacity, what would make you leave your provider? Write to us via our contact page.

AI 101 — What is an AI data center?

An AI data center is a building full of specialized computers — mostly GPUs — designed from the power feed to the cooling pipes to train and run AI models at massive scale. From the outside it looks like any windowless warehouse. Inside, the difference is density: racks of chips packed so tight, and drawing so much power, that ordinary air conditioning and ordinary grid connections no longer do the job. Why it matters right now. These buildings have become the physical front line of AI.

AI 101 — What is quantization?

Quantization is storing an AI model's numbers at lower precision — fewer bits each — so the same model fits in less memory and usually runs faster, at the cost of a small, often invisible loss in accuracy. That is the whole idea. Everything else is about how much accuracy you lose, and how cleverly you can spend the bits you keep. Why it matters right now The arithmetic is what forced it.

AI 101 — What is tokens per second (tok/s)?

Tokens per second is how fast a language model writes: the number of word-fragments it produces each second once it has started answering. It is the spec that has quietly moved to the top of every model release. Context windows used to be the number vendors fought over; now the second line of the announcement is a speed claim, usually with a hardware name attached and a batch size buried in the fine print.

AI 101 — What is a KV cache?

A KV cache is the working memory a language model keeps of the text it has already read, so it can write the next word without redoing the work for every word that came before. That is the short answer. The reason it deserves an explainer is that this cache — not the model's weights — is usually what decides how many people one GPU can serve at once, and therefore what long conversations and agent loops actually cost. Why it matters right now Every agent that runs for twenty minutes and…

AI 101 — What is a GPU, and why does AI need one?

A GPU is a computer chip built to do thousands of small arithmetic operations at the same time — which turns out to be the exact shape of the math inside a modern AI model. The name stands for graphics processing unit, because the chip was designed to draw video games frames. AI is the job it turned out to be extraordinary at. If you follow AI news, the GPU is the noun that never leaves the sentence: who is buying them, who cannot get enough of them, why one chip costs more than a car, why…

AI 101 — What is a chiplet?

A chiplet is a piece of a processor that is manufactured as its own small chip, then packaged alongside other small chips so that the finished part behaves like one big chip. Instead of printing one enormous processor, you print several modest ones and wire them together inside a single package. Every flagship AI accelerator shipping today is built this way. It matters right now because the constraint on AI hardware has moved.

AI 101 — What is prompt engineering?

Prompt engineering is the practice of writing instructions for an AI model so it gives you what you actually wanted — from a one-line question in a chat box to the multi-paragraph specifications companies embed in their products. "Prompt" is just the text you send the model; "engineering" is the unglamorous work of making that text specific enough to work. Why it matters right now.

AI 101 — What is AI inference?

Every answer an AI gives you is inference: running a trained model on new input to produce an output. Training is how a model learns; inference is how it works. If training is teaching, inference is doing. Why it matters right now. Training makes the headlines — a new model, a bigger cluster, a record run — but training happens once per model, while inference happens every time anyone asks a question.