Hyperscalers vs neoclouds: what is the difference?

16 August 2026 · 14 min read

The short answer: a hyperscaler is a general-purpose cloud - Amazon, Microsoft, Google, Meta - that sells hundreds of services, of which AI compute is one, and pays for its data centres out of its own operating cash flow. A neocloud is a specialist that sells almost nothing but accelerated compute, pre-sells it on multi-year contracts, and pays for the hardware with debt raised against that hardware. The two are not competing versions of the same business. They are the same physical asset - a rack of accelerators - held inside two completely different balance sheets.

That single distinction explains nearly everything else: why a neocloud's stock moves on interest rates and dilution while a hyperscaler's moves on advertising and enterprise software, why hyperscalers are simultaneously neoclouds' biggest competitors and biggest customers, and why the six neoclouds on this board sit a median 39.6% below their 52-week highs against a board median of 25.1%.

All board figures below are as at the close on Friday 14 August 2026, the last complete session. Everything else is sourced at the foot of the page.

Hyperscalers vs neoclouds at a glance

Who actually counts as which?

Hyperscalers are the four companies operating global general-purpose clouds at a scale nobody else reaches: Amazon Web Services, Microsoft Azure, Google Cloud and Meta, which builds at hyperscale for itself rather than renting to others. Oracle is usually counted as a fifth on capacity, though its cloud is far smaller and far more concentrated. None of them is on this board, for a reason set out further down.

Neoclouds split into two layers that the market frequently prices as one, and should not:

• The compute layer. CoreWeave (CRWV) and Nebius (NBIS) operate the clusters and sell compute directly to AI labs and enterprises.
• The power-and-campus layer. Applied Digital (APLD), TeraWulf (WULF) and Cipher Mining (CIFR) supply grid-connected sites and buildings, most of them originally developed for bitcoin mining. IREN is attempting both.

A landlord and an operator have different revenue quality, different margins and different exposure to the price of compute. We keep them in one segment because they are financed the same way, which is the variable that has actually driven their prices. The full neocloud explainer is here.

How they function differently: six mechanisms

1. What you are actually renting

A hyperscaler sells an ecosystem. The GPU instance sits alongside identity, storage, databases, queues, monitoring, compliance attestations and a support contract. That breadth is the product, and for anything customer-facing it is difficult to replace.

A neocloud sells the machine. Access is typically bare metal or a thin hypervisor, which removes virtualisation overhead and gives direct control of the interconnect - the thing that determines whether a training run across thousands of accelerators is efficient or merely expensive. There is far less around it, and for a training cluster that is frequently the point.

2. Consumption versus contract

This is the difference that most changes how you read the financials. Hyperscaler cloud revenue is consumption: customers turn capacity on and off, and revenue is recognised as it is used. Neocloud revenue is contracted years ahead, which is why the headline number in every neocloud release is a backlog rather than a run rate.

CoreWeave reported a revenue backlog of approximately $104.2 billion at 30 June 2026, up 246% year on year, with more than $25 billion of net new commitments added in early Q3 on top. Nebius disclosed more than $40 billion of additional contracted revenue from investment-grade customers in July. Neither figure is revenue. It is the promise of revenue, contingent on the capacity being built and energised.

So the number that matters in this sector is not backlog at all. It is contracted power versus active power. CoreWeave ended the quarter with roughly 1.5 GW active against approximately 3.7 GW contracted - about 40% of what it has signed for is switched on. Backlog earns nothing until megawatts are energised.

3. Where the money comes from - the difference that drives the share price

A hyperscaler funds capacity from operations. Amazon, Microsoft, Google and Meta throw off enormous operating cash flow from businesses that have nothing to do with AI, and AI capacity is a claim on it. When capex rises, the pressure lands on free cash flow and the multiple; it does not threaten solvency.

A neocloud has no such business. It buys accelerators with debt secured against those same accelerators, and issues equity alongside. The consequence is that a neocloud's equity is levered to the cost of capital in a way a hyperscaler's is not. A rate move, a spread widening or a discounted raise can move the stock without a single thing changing about demand for the compute.

This is why announcements stopped working. Through 2026 the sector has repeatedly signed large contracts and fallen anyway, because the market's question changed from "can they win the work" - they demonstrably can - to "what does funding it cost, and who pays for it".

4. Whose silicon runs inside

Every hyperscaler now ships its own accelerator: Google's TPU, Amazon's Trainium, Microsoft's Maia and Meta's MTIA, several of them co-designed with Broadcom. They buy enormous quantities of NVIDIA hardware as well, but they have a second source they control, and each generation moves a little more inference onto it.

Neoclouds have no such option. They are NVIDIA shops almost without exception, and in several cases NVIDIA is a shareholder - a 9.3% stake in Nebius, 47.2 million shares in CoreWeave, and discussions reported around a stake in IREN. That is validation and conflict at the same time, and the honest read is that both are true.

For anyone holding chip stocks this is the practical consequence: hyperscaler demand is split between NVIDIA and in-house silicon, while neocloud demand is not. Growth in neocloud capacity is close to a pure read on NVIDIA. Growth in hyperscaler capex is a read on NVIDIA and on the custom-ASIC supply chain, and those two do not always move together.

5. How long the hardware is assumed to last

This one rarely makes the headlines and moves reported earnings more than most things that do. AWS, Microsoft and Google all write servers off over six years. CoreWeave also uses six, having extended from five at the start of 2023. Nebius uses four.

The same asset, the same revenue, and a materially different reported profit. A shorter life means higher depreciation, lower reported earnings now, and less exposure if second-hand accelerator values fall faster than assumed. A longer life flatters current earnings and takes the risk that the fleet outlasts both the shortage that priced it and the architecture that replaced it. Neither is wrong. But comparing two operators' margins without checking this number is comparing two different accounting policies, not two businesses.

6. Who the customers are

No single customer is material to AWS or Azure. Neocloud revenue is the opposite: Microsoft alone accounted for about 67% of CoreWeave's 2025 revenue, up from 62% the year before, and no other customer reached 10%. Concentration has fallen as OpenAI, Meta, Anthropic and others have signed, but it remains the structural risk of the model. One counterparty deciding to build rather than rent is a different kind of event for a neocloud than for a hyperscaler.

Are hyperscalers competitors or customers of neoclouds?

Both, simultaneously, and the reason is accounting as much as capacity.

When a hyperscaler builds its own data centre, the spending is capital expenditure: it lands on the balance sheet and weighs on free cash flow in the year it is spent. When it rents the identical capacity from a neocloud, it becomes operating expense spread across the life of the contract. In a year when the four largest buyers are guiding to capex somewhere between $660 billion and $725 billion depending on which tracker you use - the spread is mostly Microsoft's June fiscal year and whether Oracle is counted - the difference between those two treatments is not a rounding error. It is a deliberate financing choice.

Which is why the customer lists read the way they do. Meta committed a further $21 billion to CoreWeave in April 2026, taking its total commitments to roughly $35 billion and extending into the 2030s. Microsoft was CoreWeave's largest customer by a wide margin. A neocloud is not only a way to get compute sooner. It is a way to get compute off the balance sheet, and that demand persists even if GPU supply loosens.

The tension is obvious and unresolved: the same companies renting capacity are building their own, designing their own silicon to fill it, and are the neoclouds' most creditworthy counterparties. A contract that de-risks a neocloud today is signed by the party best placed to stop renewing it.

Why no hyperscaler appears on this board

ChipSentiment tracks 34 US-listed names across AI silicon, optics, memory, neoclouds, equipment and power. Microsoft, Amazon, Alphabet and Meta are not among them, and that is deliberate rather than an omission.

They are the demand side. Their share prices are set by advertising, retail, enterprise software and consumer products; AI capex is an input to those businesses, not the business itself. A sentiment score built from price would be measuring the advertising cycle and reporting it as an AI-infrastructure reading. Neoclouds are here because the whole of their equity story is the buildout - they behave like leveraged infrastructure, and the score reads them as such.

Two consequences worth being explicit about. This board cannot tell you anything about hyperscaler sentiment, because it does not measure it. And it will register a change in hyperscaler behaviour only second-hand, through the suppliers and renters that sit downstream of it.

What the board's own numbers say about the neoclouds

Our market sentiment score is built from six weighted, price-only components - momentum, relative strength, trend, volume, volatility and drawdown - and nothing else. It has no view on business models. Which makes it useful here, because whatever it shows about the neocloud group is a property of how those shares actually trade rather than of anything we believe about them.

The drawdown component is bounded at -4.8 points, and it is the one place in the formula where the neocloud group separates itself from everything else on the board.

Of the five names across all 34 whose drawdown component sits within 0.3 of that floor, four are neoclouds. IREN, TeraWulf, Cipher Mining and Applied Digital are all at -4.5 or worse; the only non-neocloud beside them is Arm. Put the other way: four of the six neoclouds are pinned at the drawdown floor, against one of the other 28 names on the board.

That is the leverage showing up in the price series rather than in the filings. These are the names that fell furthest when the funding question replaced the demand question, and they have not come back. Nebius is the exception in both directions - 7.4% off its high after a 34% session on Q2 results, and the only neocloud whose drawdown component is positive.

One caveat we would rather state than have inferred. A deep drawdown is a measurement, not a verdict. It says these shares are a long way below where they traded, which is equally consistent with a repricing that has further to run and with a group that has already taken its punishment. The score does not forecast, and neither should a reader of it.

1% of the market value, 24% of the conversation

Two more measurements from this board that say something about how the two models are held.

The six neoclouds are worth $167 billion combined. The whole 34-name board is worth $16.6 trillion. So the neocloud group is 1.0% of the board's market value - and 3.1% of NVIDIA on its own.

Now set that against attention. We count posts on X carrying each ticker's cashtag ourselves, one row per complete New York day, retweets excluded. On 14 August the board drew 22,939 posts. The six neoclouds took 5,407 of them - 23.6%.

One per cent of the market value, roughly a quarter of the conversation. Nebius alone drew 2,975 posts against NVIDIA's 4,637, while being about 79 times smaller.

Posts on X carrying each ticker's cashtag, counted by this site over 27 August 2026. Retweets excluded. See all 34 names →

What that gap does not mean is worth stating as plainly as the number. We tested it on our own data: across all 34 names on a full session, the rank correlation between post volume and the day's price move was -0.00. Post counts track turnover and controversy, not direction - a collapse and a rally generate the same kind of number. High-beta, heavily-shorted, retail-held equities generate conversation. That is what the 23.6% is measuring.

The scale asymmetry nobody prices properly

Here is the comparison that puts the whole relationship in proportion, and it is straightforward arithmetic on two public figures.

Combined 2026 capex guidance across the largest cloud buyers: $660 billion to $725 billion. Combined market value of every neocloud on this board: $167 billion.

The hyperscalers plan to spend, in one year, roughly four times the entire market capitalisation of the listed neocloud sector.

Two things follow, and they point in opposite directions - which is precisely why the sector is volatile. A very small share of that budget redirected toward rented capacity is transformational for companies this size. And a small share withdrawn is equally transformational, in the other direction, for companies whose debt was raised against the assumption it would arrive. The neoclouds are a leveraged derivative of a spending decision taken inside four companies that are not on this board.

What this means if you are building an AI company

The practical differences, stated without a recommendation attached:

• Price. Published on-demand rates for the same accelerator typically run substantially below hyperscaler list prices on the specialist clouds - comparisons in 2026 have put the gap at roughly two to three times on H100-class instances. Two caveats matter: those are list prices, and hyperscalers discount heavily on committed multi-year spend, so the realised gap for a large buyer is narrower than the sticker gap.
• Access to new hardware. Specialists generally deploy new NVIDIA architectures sooner, because the newest accelerators are their entire product rather than one instance family among hundreds.
• Performance shape. Bare metal and direct interconnect control matter for large distributed training runs and matter much less for single-node inference. The advantage is real and it is workload-specific.
• What you give up. The managed services, the compliance surface, the regional footprint and the integrations. For an application stack that is a large loss; for a training cluster it is often irrelevant.
• Counterparty risk runs the other way too. A neocloud's balance sheet is now something a customer signing a multi-year commitment has to form a view on. That is not a question anyone asks about AWS.
• Which is why most mature teams run both. The common 2026 pattern is training and batch inference on a specialist cloud, with the product, data and serving stack on a hyperscaler. Choosing one exclusively is usually a decision about procurement, not about engineering.

How to tell the two apart in the numbers

1. Read the funding line before the revenue line. If capacity is funded by debt secured on the hardware, the cost of capital is the business. If it is funded from operations, it is not.
2. Contracted power versus active power. The single most informative number in the neocloud sector, and the one least often quoted. CoreWeave: roughly 1.5 GW of 3.7 GW.
3. Check the depreciation life before comparing margins. Six years versus four is a different reported profit on identical economics.
4. Customer concentration, and who the customer is. A backlog signed by an investment-grade buyer that is also building its own capacity is a different asset from one signed by a startup.
5. Backlog is not revenue. It is a promise contingent on delivery, and delivery is a construction and grid-connection problem.

See live prices and market sentiment across all six neoclouds →

None of the above is a recommendation and it does not account for your circumstances. Board figures are as at the close on 14 August 2026 and refresh with each rebuild; company figures are drawn from the filings and reporting linked below and were accurate when published.

Frequently asked questions

Sources

• CoreWeave - second quarter 2026 results - revenue, $104.2bn backlog, 1.5 GW active against 3.7 GW contracted, 2026 guidance
• CNBC - Meta commits an additional $21 billion with CoreWeave - hyperscaler-to-neocloud commitments and contract duration
• CoreWeave 10-K - customer concentration - Microsoft at 67% of 2025 revenue, up from 62%; no other customer above 10%
• Futurum Group - AI capex 2026 - per-company 2026 capex and the Microsoft fiscal-year caveat; lower end of the range
• ValueAdd - AI spending tracker 2026 - upper end of the combined 2026 capex range, updated 11 August 2026
• SiliconANGLE - resetting GPU depreciation - six-year server useful life at AWS, Microsoft and Google
• Bizety - GPU depreciation: CoreWeave vs Nebius - CoreWeave's six-year schedule against Nebius's four
• Tom's Hardware - the custom AI ASIC state of play - hyperscaler in-house accelerator programmes and Broadcom's role
• Nutanix - neoclouds versus hyperscalers - service-scope and bare-metal architecture differences
• VESSL - GPU cloud pricing compared - published on-demand H100 list rates; vendor-published, treated as list pricing only

Figures are taken from the public filings and the reporting linked above.

Stocks mentioned

More from ChipSentiment

← All articles