Neocloud Companies and Stocks (2026): Who Runs the AI GPU Cloud?

The word neocloud has become shorthand for the companies renting out GPUs to train and run artificial intelligence. It describes a real business category, but the label covers very different companies: some run full GPU-cloud platforms, some rent developers a single GPU by the hour, and others build the powered data centers where somebody else's GPUs will live.

That distinction is more useful than a generic list of "AI cloud stocks." A company earning cloud-service fees has different funding requirements and margins from a data-center landlord collecting rent—even if both announce the same number of megawatts.

What is a neocloud?

A neocloud is a specialist provider of GPU or accelerator-based computing for AI training, inference and related high-performance workloads. Unlike Amazon Web Services, Microsoft Azure or Google Cloud, it usually concentrates its capital and software on a narrower set of compute products rather than offering hundreds of general cloud services.

The first generation mostly sold access to GPUs: bare-metal servers, virtual machines and dedicated clusters. The next generation adds scheduling, checkpointing, managed training and inference, security, networking and developer tooling. Those software layers are important because customers can otherwise shop among suppliers mainly on price and availability.

Neocloud is not a regulatory or accounting category, so rankings that call data-center landlords, AI inference APIs, Bitcoin miners and cloud operators all "neoclouds" will not always agree.

Neocloud companies to know in 2026

This is a selected map of prominent operators and adjacent companies, not a ranking by quality or a complete census. Listing status and financing examples are stated as of November 13, 2026.

The word private here means not listed on a public stock exchange—not that financial information about the business is impossible to find. Private-company headline financing and public-company reported revenue also cannot be compared as if they were the same metric.

Which neocloud stocks can an investor actually buy?

CoreWeave (CRWV), Nebius (NBIS) and IREN (IREN) are three publicly listed companies providing GPU-oriented AI cloud infrastructure. They differ in maturity, service mix, power ownership and exposure to legacy operations.

Several other listed companies appear on neocloud stock lists but are best treated as adjacent infrastructure or capacity-landlord exposures: Applied Digital (APLD), TeraWulf (WULF) and Cipher (CIFR) provide, develop or convert data-center capacity. Whether they run the GPUs and bill the end AI customer themselves depends on the company and contract. A Bitcoin-mining company planning to convert a site is not equivalent to an AI cloud with deployed GPU clusters and recurring customers.

NVIDIA (NVDA) supplies chips and sometimes invests in cloud operators; it is not a neocloud. The distinction matters when comparing stock prices, revenue multiples and projected growth.

This is an industry map, not a recommendation to buy, sell or short any security.

Where the business models really differ

Not every provider fits only one row: IREN is developing a full AI cloud while owning power-rich sites, and Crusoe combines infrastructure development and GPU services. Classify the revenue and the work performed, not just the marketing label.

Where the funding actually comes from

The largest neoclouds do not usually buy every GPU with venture equity. They combine several sources.

Venture or strategic equity: investors buy shares and accept the business risk. Crusoe announced an initial $3.9B Series F close on September 17, co-led by Atreides Management, Mubadala Capital and Valor Equity Partners. Together AI's $800M Series C, disclosed July 1, included venture and strategic investors such as NVIDIA, General Catalyst and others. Runpod's $100M growth financing was led by Summit Partners.

Convertible notes: creditors lend money that may later convert to shares. Nebius completed $5.75B convertible notes in August. CoreWeave issued $4.2B convertible notes in September. These amounts are debt principal, not customer contracts or new venture rounds.

GPU- and contract-backed loans: lenders secure financing on hardware and contracted service cash flows. In March CoreWeave closed an up-to-$8.5B investment-grade rated DDTL 4.0 facility, which started with about $7.5B available. In May it closed a separate $3.1B publicly syndicated DDTL 5.0 with a lower credit rating. IREN closed a $3.65B facility for its Microsoft contract in June.

Customer prepayments: buyers agree to pay cash in advance to secure future service. This may reduce equity dilution and interest expense, but it also creates a duty to deliver capacity. CoreWeave projected about $9B in near-term advance payments and deposits after Q3. Nebius expects prepayments and loans secured against contracts to finance 75–85% of its planned $40–45B 2027 capex. Neither is automatically cash already received or revenue already earned.

A detailed breakdown, including credit ratings and borrowing rates, is in How GPU financing works. For individual company history use VCT's investment feed and the company profiles linked above.

Who are the customers—and what is the hidden concentration risk?

The same technology companies appear on both sides of many deals. Microsoft and Meta purchase large compute capacity; NVIDIA supplies accelerators and invests in infrastructure companies; banks finance hardware against the expected customer payments.

That structure can accelerate construction when buyers have strong credit quality. It can also mean that one customer holds several forms of influence over the supplier. Losing a renewal or missing a deployment can affect revenue, debt covenants and future borrowing capacity at once. A hardware supplier taking an equity stake does not automatically make the underlying compute revenue circular or fictitious; the useful question is which payments and obligations represent genuine third-party demand, and when they settle.

For Nebius, delayed capacity at a Microsoft-linked New Jersey site pushed some planned delivery into 2027. For CoreWeave, a $132.5B reported revenue backlog signals contracted future service, but it must be fulfilled before it becomes recognized revenue. IREN's November results likewise distinguish $9B of cash and committed funding sources from another $6B of target financing in its remaining FY27 capex plan.

Those aren't three interchangeable kinds of "funding."

How much margin can a neocloud keep?

Cheap GPU rentals can win early customers; sustainable economics require more.

McKinsey's November 2025 research warned that bare-metal GPU rentals have weak margin protection. In its illustrative B200 model, economics become fragile if GPU prices fall or utilization drops below roughly 80%. That is a model sensitivity, not a measured break-even threshold for every provider. Operators can have different hardware purchase terms, software revenue, utilization, power prices and financing costs.

Every cloud operator faces a sequence of questions: Did the GPUs arrive? Is the site powered? Are customers using the machines? Are the bills being collected? What is left after operations, equipment depreciation, interest and eventual GPU replacement?

A provider may look profitable on an adjusted EBITDA basis while recording a GAAP loss or consuming substantial cash. Conversely, a long-term low-margin hyperscaler contract may still be attractive if it provides predictable base utilization and supports better financing terms.

What could determine the winners in 2027?

Supply versus demand: today a shortfall in available compute can support higher prices. That can reverse if competitors bring large volumes of similar GPU capacity online.

A software advantage: better inference runtimes, scheduling, cluster reliability, security or domain-specific tools can make a provider harder to replace. McKinsey's thesis is that remaining a commodity bare-metal lessor offers little differentiation.

Power and deployment: contracted megawatts, grid-connected megawatts and operating GPU clusters are different quantities. Delays affect the dates when customers can be billed.

Capital discipline: compare cash already received, loan capacity that remains undrawn, obligations to customers and the timing of refinancing. The biggest capex plan is not necessarily the best business.

The second life of GPUs: after an initial contract ends, an older accelerator might serve a lower-priced inference workload—or sit idle. Residual value is an assumption to test, not a guarantee.

Consolidation: some firms may be bought by infrastructure funds, hyperscalers or large cloud operators; others may specialize in sovereign compute or narrow workloads. These are possible paths, not announced transactions.

How to use this map

For a founder choosing where to run AI workloads, start with GPU availability, price, networking, data placement, support and ease of scaling. For a venture investor, start with the company's differentiated software, customer durability, cash financing and renewal economics. For a public-markets reader, first decide whether the business is a GPU cloud, infrastructure landlord or diversified chip supplier.

The most revealing question in neocloud investing is no longer simply "how many GPUs does it have?" It is "who pays for them, when do they start earning, and what happens when the first contract ends?"

For the company-level details read CoreWeave's Q3 debt and backlog analysis, Nebius's capex and delivery analysis, and the CoreWeave–Nebius financing comparison. The 2026 AI infrastructure funding market map covers neighboring software, chips, networking and power companies.