Key takeaways
- Neoclouds specialise in AI compute, often operating GPU clusters inside third-party data centres
- More than $52 billion in selected 2026 financings illustrates the category’s capital intensity, with debt playing a major role
- Older GPUs can retain customers, but their economic life depends on utilisation, pricing, power costs and financing
- Providers are expanding into inference, developer tools and managed services to build lasting customer relationships
- For early-stage startups, neoclouds can become customers, distribution partners, acquirers and competitors
We are in the middle of an extraordinary investment boom in AI infrastructure. Technology companies and investors are pouring capital into the compute capacity needed to train AI models and run the products built on them. Behind the seemingly weightless experience of asking a chatbot a question sits a very physical industry of data centres, power supplies, cooling systems and expensive chips.
That expansion has created an opening for a new generation of cloud providers. Known as neoclouds, these companies specialise in GPU-powered computing for AI. SemiAnalysis now tracks 323 GPU cloud providers, up from 209 in its previous ClusterMAX edition a year ago. That is a 55% expansion in its tracked universe, reflecting broader research coverage as well as the development of the market.
Many of these providers rent space in existing data centres, finance the GPU clusters and sell computing capacity to AI labs, startups and enterprises. Others develop their own facilities or expand into software and managed services. Their emergence is changing who supplies the infrastructure behind AI, and who carries the financial risk.
The opportunity is enormous, but the equipment ages quickly. Can these companies recover their investment before newer chips change the economics? And can they build customer relationships that last beyond a single generation of hardware?
What a neocloud actually does
A neocloud is a specialist cloud provider built around accelerated computing, particularly GPUs used to train and run AI models. Where AWS, Microsoft Azure and Google Cloud offer enormous catalogues of general-purpose services, neoclouds typically start with a narrower proposition: access to high-performance compute, configured and operated for AI workloads.
GPU-as-a-Service, or GPUaaS, is the underlying offering. Customers rent computing capacity rather than buying and operating the equipment themselves. They might reserve an entire cluster for several years, rent capacity by the hour, or buy a managed service that handles much of the infrastructure for them.
In a typical colocation arrangement, the landlord supplies space, power, cooling and physical security. The neocloud installs or finances servers, networking and storage, then operates the service customers use. Some providers also own or develop facilities, so describing them all as data-centre owners misses where the capital and risk actually sit.
The distinction matters because renting the building does not make the business capital-light. Someone still has to pay for the GPUs, connect them into reliable clusters and keep them busy.
Billions are flowing into the category
The scale of financing gives a sense of how quickly this market has developed. A tally of selected 2026 financings across just four providers exceeds $52 billion.
| Provider | Selected 2026 financing | What is included |
|---|---|---|
| CoreWeave | More than $35bn | Chip-backed loans, bonds, convertibles and share sales reported by the Financial Times on 14 October |
| Nebius | $12.86bn | $2bn NVIDIA investment through pre-funded equity warrants, $4.3375bn and $5.75bn convertible-note offerings, and a $775m secured debt facility |
| Nscale | $2.36bn | Initial tranche of its announced pre-IPO convertible financing, excluding a further $1bn NVIDIA tranche expected in November |
| Lambda | $1.926bn | A $926m term loan announced in August and $1bn secured fixed-rate financing announced in October |
| Selected total | More than $52.1bn | Gross financing amounts and facility sizes, rather than venture equity alone |
Sources: Financial Times; Nebius’s NVIDIA transaction, March notes, August notes and secured facility; Nscale; Lambda’s August and October financings. www.ft.com
This is a partial tally, not a market-wide funding total or a comparison of lifetime capital raised. It excludes customer contract values, announced capital expenditure budgets and several other financings. Facilities may be drawn over time, and gross financing can include refinancing rather than entirely new spending power.
Those distinctions are essential. A multibillion-dollar customer contract can help finance a GPU cluster, but it is not equivalent to money raised or revenue already earned.
Increasingly, long-term commitments from creditworthy customers support asset-backed borrowing. Nebius’s July financing, for example, was backed by deployed GPU infrastructure and contracted customer cash flows. This gives providers another way to fund expansion, while making contract quality, delivery obligations and repayment schedules central to the investment case.
A company can have strong demand and still run into trouble if it commits to equipment, power and leases before customer revenue arrives. The timing matters almost as much as the headline contract value.
What happens when the chips get old?
There are three different questions inside the GPU-longevity debate: how long a chip physically works, how long the accounts assume it will remain useful, and how long it earns an attractive return.
A functioning GPU can become less competitive before it breaks. If newer hardware produces substantially more useful output per unit of electricity, customers may demand a discount to use the older generation. The operator still has power bills, rent and debt payments. At a power-constrained site, retaining old equipment also carries an opportunity cost: newer hardware could generate more output from the available power.
But older chips can retain commercial value. In a September investor presentation filed with the SEC, CoreWeave disclosed an A100 renewal extending into 2029. That architecture launched in 2020. The customer uses the capacity for scientific datasets and model fine-tuning, demonstrating that previous-generation hardware can remain useful well beyond its time at the frontier.
That does not establish that every physical GPU operates profitably for nine years. Purchase price, deployment date, utilisation, electricity costs and renewal terms still determine the economics. It does challenge the assumption that a new chip generation immediately renders the previous one worthless.
The question I would ask is: how much of the investment is recovered during the first customer contract, and what has to remain true after it expires?
A fleet that has already recovered its purchase cost has different options from one that needs premium rental prices for several more years to service its debt. Similarly, a provider with established customers for older hardware has a different risk profile from one assuming that buyers will appear when its initial contracts end.
Growing AI demand can coexist with poor returns for individual operators. The industry can need more compute while some companies buy the wrong equipment, overpay for it or finance it on terms they cannot sustain.
Why neoclouds are moving up the stack
This explains why ambitious neoclouds want to sell more than GPU hours. Managed training, model deployment, inference optimisation, security and developer tools can make their services more valuable and harder to replace.
Customers care about getting a job done. A lower hourly GPU price is little help if the cluster is unreliable or their engineers spend days troubleshooting it. The relevant measure is the total cost of completing the workload, including engineering time, failed runs and delays.
SemiAnalysis’s ClusterMAX assessment reflects that distinction. It evaluates managed clusters across compute, networking, storage, orchestration, monitoring, support and security. Access to the same chips does not mean providers deliver the same service.
Nebius provides a concrete example of the move into software. Its agreement to acquire Eigen AI was designed to strengthen Token Factory, its managed inference platform. Its Tavily acquisition announcement extended the proposition into real-time search for AI agents.
These services bring a provider closer to the developer’s workflow. A customer buying raw capacity can compare competing offers relatively easily. A customer relying on deployment tools, monitoring, model optimisation and operational support has more to consider before switching.
Moving up the stack also creates new competition. Neoclouds encounter software specialists as well as the hyperscalers’ established platforms. Building a broader product catalogue is only useful if customers choose those products and the additional revenue justifies the cost of developing and supporting them.
My expectation is that the category will become more differentiated. Some providers will compete through scale, power access and financing. Others will build strong managed platforms or specialise in particular workloads, regions and customer requirements. The same business model will not suit all 323 providers in the tracked universe.
Why early-stage investors should pay attention
For those of us investing at pre-seed and seed, the opportunity extends well beyond financing GPU fleets. Neoclouds are potential customers, distribution partners and acquirers for startups that improve inference efficiency, secure workloads, manage deployments or help developers operate across different environments.
There is a competitive risk too. As providers expand their software offerings, some standalone products will become bundled features. Founders need to understand whether a cloud partnership expands their distribution or puts them directly in the path of the provider’s product roadmap.
For application founders, supplier choice also deserves more attention than a comparison of GPU prices. Reliability, support, data portability and the ability to move workloads become important when a provider sits underneath a production product. A cheap service is expensive if switching away from it becomes an emergency.
I expect the neocloud market to produce large integrated platforms, focused specialists, acquisitions and failures. Access to power and affordable capital will matter alongside engineering and customer relationships. A rapidly expanding provider list does not guarantee hundreds of durable, independent businesses.
The test is whether a neocloud can remain valuable to customers through several generations of hardware. Today’s GPUs create the capacity to serve them. The business has to earn the right to keep serving them when those GPUs are replaced.
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.
- The neocloud boom and the race to outlive the GPUs - October 14, 2026
- The AI opportunity hiding between capability and adoption - October 13, 2026
- Weekly Firgun Newsletter – October 9th 2026 - October 9, 2026
Sources
- ClusterMAX 3.0: The Industry Standard GPU Cloud Rating System Returns
- CoreWeave targets ‘unlimited’ insurance money to fund AI build-out
- Investing Beyond Data Center Platforms
- CoreWeave September 2026 investor presentation
- Nebius agrees to acquire Eigen AI, strengthening Nebius Token Factory as a frontier inference platform
- Nebius selected to establish Israel’s national supercomputer
