GPU-as-a-Service Revolution or Debt-Fueled Illusion? Inside the Neocloud Model

They promised to democratize AI compute, dethrone the hyperscalers, and make GPUs-as-a-Service the next great infrastructure business. But behind the $42 billion market and the breathless valuations, the neocloud story has a crack running through its foundation.

Somewhere in 2022, a beautiful narrative took shape. The hyperscalers — Amazon, Microsoft, Google — were too big, too slow, too generalized. They were built for the old world of CPU-dominated enterprise workloads. The new world needed something different: purpose-built GPU clusters, tuned for AI training and inference, operated by specialists who understood the hardware deeply and could move faster than a three-year cloud roadmap.

Enter the neoclouds. CoreWeave. Lambda Labs. Crusoe. Nebius. Voltage Park. Backed by Nvidia equity, venture capital, and a market narrative more powerful than almost anything tech had produced since the smartphone — GPU-as-a-Service for the AI era.

The pitch landed. The money followed. The market is now valued at $42.17 billion in 2026, projected to reach $253.79 billion by 2030 at a 56.6% CAGR. Approximately 200 neocloud operators are active globally today. McKinsey dedicated an entire research note to their “evolution and next moves”. CDO Trends predicted neoclouds would grab $20 billion in revenue in 2026, eroding hyperscaler dominance in GenAI.datacenterknowledge+3

The pitch was perfect,the balance sheets are not.

What a Neocloud Actually Is — And Isn’t

Before examining the illusion, it is worth establishing precisely what neoclouds are — because the category is frequently misunderstood as a technology story when it is, in fact, a financial engineering story with a technology wrapper.

A neocloud is a vendor that provides GPU processing-as-a-service. Some were born as AI-native startups. Others migrated from crypto-mining operations when Bitcoin economics soured and GPU economics looked more attractive. Some are pure data center operators who spotted the AI wave early. What unifies them is a single strategic bet: rent Nvidia GPUs at wholesale, rack them in purpose-built data centers optimised for high-density AI workloads, and sell access to those GPUs at a margin.

The architectural differentiation is real. Where hyperscalers offer standardized instance types across general-purpose CPU and GPU workloads, neoclouds offer customisable GPU cluster configurations with NVLink and NVSwitch interconnects, InfiniBand networking, and cooling systems designed specifically for AI training at scale. For AI-native workloads — model training, fine-tuning, high-throughput inference — this specialisation genuinely translates into lower effective cost per training run and higher sustained GPU utilisation compared to hyperscaler equivalents.

The technology is real. The business model is where the complexity begins.

The Debt Machine Running the Revolution

CoreWeave is the bellwether of the neocloud category — the most funded, the most visible, the most recently public. Its numbers are the most transparent case study available for what the neocloud model actually looks like under the hood.

As of March 2026, CoreWeave’s total long-term debt stands at approximately $18.5 billion. The company burned through $8 billion in operating cash over the past four quarters. Its debt-to-equity ratio at the end of 2024 stood at 1,262.8% — debt was over 12.6 times equity. Interest expenses have surged to $311 million in a single quarter, up from $104 million the prior year. Despite posting 60% EBITDA margins on paper, depreciation and interest expenses consumed most of the company’s revenue, resulting in a net loss.

The company’s IPO in March 2025 was a study in investor ambivalence. CoreWeave initially targeted a $2.7 billion raise — and was forced to slash it to $1.5 billion following investor concern about the debt burden. Within weeks of listing, CoreWeave was back in the debt markets seeking another $1.5 billion from JPMorgan.

This is not an anomaly. This is the structural DNA of the neocloud model: you borrow billions to buy GPUs, you depreciate them over 3–5 years, you pray that GPU demand outpaces GPU obsolescence, and you service the debt with “take-or-pay” contracts that guarantee a baseline revenue floor regardless of whether the customer actually uses the compute. It is an infrastructure model dressed in a software growth story — and the market has struggled to price it correctly.

The Customer Concentration Trap

The take-or-pay contract structure — celebrated by analysts as the “silver lining” of the neocloud model — conceals a risk that makes the entire category fragile in ways that headline revenue numbers do not show.

CoreWeave’s 86% of revenue comes from just four major clients. A single client — Microsoft — accounts for over 60% of CoreWeave’s total revenue. OpenAI, Meta, and Nvidia dominate the backlog.intellectia+2

Now consider the interdependencies. Microsoft itself has disclosed that approximately 45% of its $625 billion cloud backlog is tied to a single customer: OpenAI. OpenAI has a reported $300 billion compute commitment with Oracle. Oracle is simultaneously one of CoreWeave’s largest competitors and a key Nvidia distribution partner. Microsoft, OpenAI, Oracle, and CoreWeave are all financing their growth on each other’s commitments — in a web of related-party exposure so dense that one financial analyst called it a system where “growth is this concentrated, it does not unwind gently. It turns into a financial tsunami”.newsletter.semianalysis+1

McKinsey named this dynamic explicitly: for neoclouds, take-or-pay contracts are attractive “less for their stand-alone economics than for what they can provide, including an almost-guaranteed baseline level of utilization and a stamp of credibility that makes the neocloud more attractive to investors”. In other words: the contracts are partly a fundraising tool, not just a revenue mechanism.

The Nvidia Umbilical Cord

There is a dependency underneath all the customer dependencies that is rarely discussed plainly: every neocloud is, at its core, a leveraged bet on Nvidia’s continued dominance of the AI compute market.

Neoclouds’ dependence on Nvidia GPUs is sustainable only as long as Nvidia maintains its competitive edge — and this is “precarious if rivals make advances”. AMD’s MI series, Google’s TPUs, Amazon’s Trainium, Microsoft’s Maia 200, and an accelerating wave of custom silicon from startups like Groq, Cerebras, and d-Matrix are all targeting the exact workloads that make Nvidia indispensable today.

The Nvidia dependency runs deeper than hardware. Nvidia itself is a shareholder in CoreWeave. The chip giant benefits from neoclouds buying its GPUs at scale — which means Nvidia has actively supported the neocloud ecosystem financially, creating a circular relationship: Nvidia funds neoclouds, neoclouds buy Nvidia GPUs, Nvidia’s market capitalisation stays elevated, which finances more neocloud equity rounds.

This structure is not fraud. But it is not purely arms-length market pricing either. It is a supply chain with an equity stake in its own demand — and when supply chains have that structure, the correction, when it comes, tends to be correlated and sharp rather than gradual and manageable.

The Commodity Floor: What McKinsey Saw

McKinsey’s November 2025 research on neocloud evolution contains a line that every neocloud investor should read carefully: neoclouds were “born out of GPU scarcity” and now face a “harder test — evolve beyond bare metal or risk repeating Cloud 1.0 history”.

Cloud 1.0 history is not encouraging. The original infrastructure-as-a-service model — raw compute on demand — collapsed into a commodity race to zero as AWS, Azure, and Google competed on price. Margins evaporated. Only scale survived. The neoclouds that fail to build a differentiated software and services layer on top of their GPU infrastructure are destined to replay that exact script.

ABI Research is even more blunt: neoclouds face “the potential danger of being relegated to a merely back-end role if they fail to acquire enterprise customers directly.” Without a push into specific enterprise verticals, neoclouds risk remaining GPU brokers for hyperscalers and chipmakers — trapped in a commodity position, vulnerable to margin pressure as hyperscalers consolidate control.

The challenge is structural. Building enterprise relationships, vertical-specific AI platforms, compliance frameworks, and go-to-market organisations requires a completely different capability set from operating GPU clusters. Most neoclouds are infrastructure companies trying to become software companies while simultaneously managing $14+ billion debt loads. That is an extraordinarily difficult transition to execute at speed.

200 Providers. One Outcome.

There are approximately 200 neocloud operators globally today. There will not be 200 in 2030. The consolidation is not a prediction — it is already underway.

Over 80% of high-performance GPUs are already deployed by a small number of well-capitalised providers. The providers lacking sufficient capital to fund the next hardware refresh cycle — from Hopper to Blackwell to whatever follows — face an existential choice: find a buyer, find a hyperscaler willing to absorb their capacity on lease, or exit the market.

The supply chain crisis compounds the pressure. GPU-ready infrastructure faces component delays across memory, high-speed networking, cooling equipment, and power infrastructure. Power densities for modern AI workloads now exceed 100 kilowatts per rack — a threshold that requires specialised facility design, custom cooling architectures, and power procurement relationships that take years to build and cannot be replicated quickly by under-capitalised operators.

The neocloud market in 2030 will almost certainly be a three-to-five player oligopoly dominated by the survivors who either raised enough capital to continuously refresh their GPU fleets, built a differentiated software layer that created genuine enterprise stickiness, or secured long-term hyperscaler capacity agreements that provided a revenue floor sufficient to service their infrastructure debt. Everyone else is a consolidation target.

India: The GPU Cloud Paradox

India’s position in the neocloud story is uniquely contradictory — simultaneously one of the market’s highest-growth opportunities and a cautionary case study of the risks that come with building on imported dependency.

India’s GPU cloud and GPUaaS market is projected to grow at a 63% CAGR from 2026 to 2032 — among the fastest rates globally. The IndiaAI Mission has deployed over 80,000 GPUs to national compute pools. Data center spending in India is projected to jump 20.5% in 2026. The demand signal is unambiguous.

The structural vulnerability is equally unambiguous. India’s GPU cloud ecosystem is almost entirely dependent on imported hardware — Nvidia H100s and Blackwell chips procured at global market prices, subject to US export controls, trade policy shifts, and Nvidia’s own allocation decisions. The government is running trials for indigenous GPU production — but “trials planned by end-2025” that have not yet scaled to commercial availability do not solve the 2026 infrastructure problem.

DPDPA enforcement creates an additional dimension: Indian enterprises increasingly need AI compute that can operate within data residency constraints, but most neocloud providers are global operators whose data governance frameworks were not designed around India’s regulatory specifics. The GCC sector — which processes sensitive financial, healthcare, and legal data for global multinationals — faces a specific version of this problem: it needs sovereign AI compute at neocloud speeds and prices, but the neocloud market has not yet built the compliance infrastructure that GCC procurement requires.​

The Illusion Defined

The neocloud illusion is not that the technology is fake. The GPU clusters are real. The infrastructure is real. The performance advantages for specific AI workloads are real. The 63% India CAGR is real.

The illusion is the business model story being told around it.

Neoclouds were pitched as the infrastructure layer of the AI revolution — the picks-and-shovels play that wins regardless of which AI model ultimately dominates. What they actually are is highly leveraged, customer-concentrated, hardware-dependent infrastructure businesses operating with thin margin windows, 18-month technology obsolescence cycles, and an exit thesis that depends entirely on the enterprise AI adoption curve moving faster than their debt maturity schedules.

When enterprise AI adoption is at 25% full production deployment globally, and 73% of AI deployments are failing to hit projected ROI, the take-or-pay contract structures that currently provide the neocloud revenue floor are functioning as intended. But take-or-pay contracts get renegotiated. Hyperscalers do build their own infrastructure. OpenAI has moved material compute to Oracle away from Microsoft. Meta is bringing more infrastructure in-house. Every one of these shifts is a crack in the revenue assumption that the debt structures were built on.benpouladian+3

What Survives the Reckoning

This is not an argument that neoclouds will fail uniformly. It is an argument for precision about which ones will survive and why.

The neoclouds that emerge from the consolidation wave will share specific characteristics:

• Vertical depth over horizontal breadth — providers that build sovereign AI compute for specific regulated industries (BFSI, healthcare, government) will outlast generic GPUaaS brokers
• Software differentiation — managed MLOps platforms, inference optimisation layers, and enterprise-grade observability tools create switching costs that raw GPU rental never can.
• Diversified customer bases — any neocloud with more than 50% revenue concentration in one or two clients is a single contract renegotiation away from a liquidity event
• Energy-first infrastructure design — the providers who locked in long-term power purchase agreements and purpose-built cooling systems before the energy bottleneck tightened are structurally advantaged over those still paying spot market power prices
• Capital structure discipline — the neoclouds that matched debt maturity to GPU depreciation cycles rather than taking short-duration debt to fund long-duration assets will be the ones still operating when the consolidation runs its course

The neocloud revolution is real. But revolution and sustainable business model are not the same thing — and in the infrastructure layer of AI, confusing the two is how the next generation of stranded assets gets built.

Final Thoughts

“Every infrastructure revolution produces both the companies that build the future and the carcasses of the companies that over-financed it. The neocloud market in 2026 contains both. The hard part — for enterprises, investors, and the Indian ecosystem betting on it — is learning to tell them apart before the debt maturities do it for you.”