The neocloud sector was built on a single bet: that enterprises, researchers, and AI developers would pay a premium for GPU access that hyperscalers could not or would not provide at the required scale, speed, and flexibility. That bet paid off during the GPU scarcity period of 2022 through 2024. H100 clusters that cost $30,000 per month to rent were being booked at $40,000 or more. Waitlists stretched for months. Operators who had the foresight to acquire GPU inventory ahead of the curve built businesses that seemed, for a period, almost impossible to lose.
That period is ending. The normalisation has been gradual rather than sudden. That has made it easier for operators to tell themselves it is temporary.
It is not, and the market data of the past twelve months makes that clear. Operators waiting for GPU scarcity to return as their business model’s salvation are, consequently, waiting for something that is not coming back.
GPU supply has normalised. Nvidia‘s GB200 production has ramped.
AWS, Google Cloud, and Microsoft Azure are offering GPU instances at scale and at price points that directly compete with the neocloud market.
The conditions that made neocloud economics look almost risk-free in 2023 have, consequently, changed on every dimension simultaneously.
The H100 rental market has softened from its peak. Hyperscalers who were slow to build AI-specific infrastructure in 2022 and 2023 have, in turn, caught up aggressively.
AWS, Google Cloud, and Microsoft Azure are all now offering GPU instances at scale and at price points that directly compete with the neocloud market. The question every neocloud operator is now facing is not whether the market will stay as favourable as it was. It will not.
The question is, rather, which operators built something defensible enough to survive the normalisation. Which ones built businesses that were, in turn, only viable during a supply squeeze? The answer is increasingly visible in the operational and financial data that neoclouds are disclosing as they scale.
Revenue concentration, customer retention rates, power cost per kilowatt-hour, and the ratio of contracted to spot revenue are all, specifically, telling the story more clearly than any operator’s investor presentation. The operators who can show strong numbers on all four of those metrics are, in turn, the ones most likely to still be operating independently in 2028.
The answer to that question is not simple. The market does not sort cleanly into winners and losers based on any single variable. It is a combination of decisions across capital structure, infrastructure, customer strategy, and software investment that determines who survives. And most of those decisions were made, or not made, between 2021 and 2024. It sorts based on decisions made over the past three years about infrastructure, customers, capital structure, and software. It is, however, increasingly visible in the operational and financial data that neoclouds are disclosing as they scale. The operators that will survive are those that have moved beyond GPU rental as a business model. Those that have not moved fast enough are, consequently, running out of time.
The GPU Rental Trap
The fundamental problem with GPU rental as a standalone business model is that it provides no structural differentiation. A customer renting H100s from one neocloud gets, in principle, the same compute as one at any other neocloud with the same GPU generation. Differentiation in a commodity market comes from price, reliability, and ancillary services. Price competition is, specifically, a race to the bottom. The operators with the highest cost of capital will lose it first. Reliability is, in turn, increasingly table stakes rather than a differentiator. Ancillary services are where the real strategic value is being built. Only operators who recognised the trap early enough invested in building them.
Operators who treated GPU procurement as the end game are, in turn, the most exposed to the market normalisation underway. They built procurement pipelines, data center capacity, and customer relationships. What they did not build was the software layer or the managed services offering that would make a customer reluctant to leave. When hyperscaler pricing and availability improved, those customers moved. Neoclouds Are Building on Borrowed Time If They Stay Pure Infrastructure made this argument when it was still contrarian. It is, notably, now consensus.
The GPU rental trap has a specific financial manifestation. Operators most deeply in it are those who financed GPU procurement through equipment leasing or asset-backed structures that require consistent revenue streams to service. When rental rates fall, those structures become, in turn, financially stressed. Managing the stress requires either raising new equity at dilutive valuations or restructuring the underlying debt. Leasing the Future: The High-Stakes Gamble Behind Neocloud Growth laid out how this financial structure works and why it creates specific vulnerabilities when market conditions change. Those vulnerabilities are, consequently, becoming more visible as the rental rate compression continues.
What the Survivors Are Building Instead
The neoclouds that are best positioned to survive the current market transition share a set of characteristics that, taken together, define what a defensible neocloud business actually looks like.
Anchored Revenue From Committed Customers
The single most important differentiator between neoclouds that will survive and those that will not is the presence or absence of long-term contracted revenue. A neocloud with a five-year capacity agreement with Meta or Microsoft is in a fundamentally different position from one selling on the spot market. The contracted revenue provides cash flow visibility. The spot market provides none. Contracted revenue provides cash flow visibility. It allows operators to make long-term infrastructure investments, service their debt, and retain the operational talent that makes the infrastructure actually work.
Nebius is the most transparent example of this dynamic. Its Meta agreement and Microsoft relationship provided the contracted revenue base that made its $20-25 billion 2026 capex forecast credible to investors. A similarly aggressive expansion from an operator without anchor contracts would, in turn, trigger very different market reactions. Neoclouds that lack anchor relationships are, consequently, either scrambling to secure them or accepting constrained growth trajectories. Converting spot customers to longer-term commitments is, specifically, a slow process where customers now have more options.
Anchor contracts are, however, not available to every operator. Hyperscalers and large enterprise AI customers have specific requirements around security, compliance, reliability, and technical capability that most neoclouds cannot meet without substantial investment. The operators that have made that investment and secured anchor relationships have, consequently, built a structural moat that purely capacity-focused competitors cannot quickly replicate.
A Differentiated Infrastructure Position
Physical infrastructure differentiation has become, specifically, one of the most consequential variables in neocloud competitive positioning. Operators that have built or secured infrastructure with specific structural advantages hold positions that are, in turn, difficult to replicate on short timescales. Those advantages include geography, power cost, cooling capability, and rack density.
The power cost dimension is, notably, the most important of these. Power is the largest variable cost in neocloud operations and is rising faster than rental rates in most markets. Operators with locked-in low-cost power are structurally advantaged over those buying power at current market rates. Long-term PPAs in renewable-advantaged markets and behind-the-meter generation arrangements at scale are, specifically, the two main routes to that advantage. A neocloud with Nordic hydro power at $0.04 per kilowatt-hour is competing on fundamentally different economics from one paying $0.09 in a primary US market.
Geographic diversification is, also, a meaningful differentiator. Neoclouds built in multiple regions, serving different regulatory environments and customer bases, are less exposed to any single market’s constraints or competitive dynamics. Why the Neocloud Margin Problem Is Getting Harder to Ignore identified geographic concentration as a specific margin risk. Operators that have addressed it through deliberate geographic diversification are, consequently, more resilient than those built around a single market.
A Software and Services Layer That Creates Retention
The most strategically durable neoclouds are those that have built a software and managed services layer on top of their infrastructure. That layer creates retention in ways that raw GPU capacity cannot. A customer who has integrated a neocloud’s job scheduling software faces switching costs that are, in turn, not primarily financial. They are operational. They are operational. Moving to a different infrastructure provider requires re-integrating those tools, retraining teams, and accepting service disruption during the transition. Those switching costs are, specifically, what makes a software-enabled neocloud resilient to price competition in ways that a pure infrastructure provider is not.
CoreWeave has built the most fully-developed version of this model among publicly visible neoclouds. Its Kubernetes-native infrastructure, network fabric, storage architecture, and managed services offering are all designed to make CoreWeave infrastructure feel native to AI development workflows. Switching to a hyperscaler feels, in turn, like a downgrade even when the hyperscaler’s raw GPU pricing is lower. That transition from pure infrastructure to platform is playing out across the leading operators, examined in detail in GPU-as-a-Service Has Grown Up. The Neocloud Business Model Will Never Be the Same.. Operators that have made the transition are, in turn, demonstrating materially better customer retention and lower churn than those that have not.
The Inference Opportunity and Why Not Every Neocloud Can Capture It
Inference is, broadly, understood to be the next major growth vector for AI compute. As AI models move from training to production deployment, the inference compute market will, consequently, grow substantially larger than the training compute market. That is, in principle, good news for neoclouds. In practice, however, the inference opportunity is, specifically, more demanding and more competitive than training hosting, and not every neocloud is positioned to capture it.
Why Inference Is Harder Than Training for Most Neoclouds
Training workloads are, relatively speaking, forgiving. They run for hours or days on dedicated clusters, require high throughput but not necessarily low latency, and can tolerate some variability in performance. Inference workloads that serve production applications have completely different requirements. They need consistent, low-latency response times, high availability, and the ability to scale capacity rapidly in response to demand spikes. Those requirements demand infrastructure and operational capability that most neoclouds have not yet built.
The latency requirement is, specifically, the most demanding. Serving an inference query in under 100 milliseconds requires not just fast GPUs. Optimised network paths, local caching, and inference-specific software that reduces the overhead between receiving a query and returning a result are, specifically, all required. Inference at Scale: NeoCloud’s Next Battlefield After Training covered this in depth. Operators that built their infrastructure and software stack with inference requirements in mind from the beginning are, consequently, better positioned to serve this market. Retrofitting inference capability onto training-optimised infrastructure is, in turn, slower, more expensive, and technically harder than building it natively.
The availability requirement is, also, more demanding for inference than training. A training job that experiences a hardware failure can, in most cases, restart from a checkpoint and continue. An inference service that experiences downtime has, in turn, a direct impact on the applications and users depending on it. That requires a different approach to redundancy, failover, and capacity management that most training-focused neoclouds have not had to develop.
The Custom Silicon Dimension
The inference market is, notably, also where custom silicon is having the most immediate competitive impact. Cerebras, Groq, and Tenstorrent have built hardware that delivers lower cost per token than Nvidia GPUs for specific inference workloads. Neoclouds positioned as Nvidia-only infrastructure providers are, consequently, exposed to competition from operators that offer customers a choice. GPU-based inference and custom silicon inference serve different workload profiles, and customers increasingly understand the difference.
Neoclouds that survive the inference transition will be those that treat hardware as a portfolio rather than a single-vendor commitment.
That is, notably, a more operationally demanding model than GPU-only infrastructure. It requires maintaining expertise across multiple hardware architectures and managing different software stacks. Helping customers navigate hardware selection decisions that are, in turn, becoming more complex with each generation cycle is, specifically, a capability that creates retention. That is, notably, a more operationally demanding model than GPU-only infrastructure. It requires maintaining expertise across multiple hardware architectures.
Managing different software stacks is, in turn, a further operational burden. Helping customers navigate hardware selection decisions that are becoming more complex with each generation cycle is, specifically, where the service value lives. That means maintaining Nvidia GPU capacity for workloads where GPU inference is optimal. It also means building relationships with custom silicon vendors for categories where their economics are, specifically, superior. That portfolio approach requires operational complexity that most neoclouds have not yet developed. The operators that develop it first will, in turn, hold a meaningful advantage in the inference market as it scales.
The Power Cost Reckoning
Power is, specifically, the cost variable that most neocloud financial models have underestimated at every stage of the sector’s development. That underestimation was, in the early years, largely inconsequential because GPU rental rates were high enough to absorb it. It is, however, no longer inconsequential. That underestimation was, in the early years, largely inconsequential because GPU rental rates were high enough to absorb it. It is, however, no longer inconsequential. Operators that built on 2022 power rate assumptions are discovering that actual operating costs are running ahead of forecast. In some cases, the gap is material enough to reshape unit economics fundamentally.
The arithmetic is straightforward but unforgiving. And it compounds as data centers scale to the densities that AI training requires. And it gets worse as data centers scale to the densities that AI training requires. A 100-megawatt neocloud deployment at US commercial power rates of $0.09 per kilowatt-hour costs approximately $79 million per year in electricity alone. That is before cooling overhead. At a PUE of 1.4, cooling adds another $22 million. The total electricity-related operating cost for a 100-megawatt facility runs, consequently, to over $100 million annually. At softened 2026 H100 rental rates, that electricity cost represents, consequently, a significantly larger fraction of gross revenue than original business plans assumed.
The neoclouds that recognised this problem early took two types of action. Some pursued geographic arbitrage, moving capacity to markets where power is structurally cheaper. Nordic hydroelectric power and Quebec’s grid both offer, notably, power at rates 40% to 60% lower than US commercial rates in primary markets. Several Asian markets with surplus renewable generation capacity offer similar advantages. Operators who made those geographic commitments early are, in turn, holding structural cost advantages that later entrants cannot easily close. Making similarly expensive geographic bets at a less advantageous time is, specifically, the only path to closing the gap. It is a worse bet today than it was in 2022 or 2023.
Others pursued long-term power purchase agreements, locking in fixed pricing before market rates moved higher. A neocloud that signed a ten-year PPA at $0.04 per kilowatt-hour in 2022 operates on power cost economics its 2026 competitors cannot access. Those agreements required capital commitment and credit relationships that not every operator could access. The operators who made them have, consequently, acquired a structural advantage through procurement discipline as much as through technical capability.
Operators that did neither are still buying power at current market rates on short-term or spot contracts. Their cost structure is, in turn, progressively less competitive as GPU rental rate compression continues. Power cost is not a problem that can be engineered away with operational improvements alone. It requires either geographic repositioning or long-term procurement relationships, and both require time and capital that mid-tier operators under financial pressure may not have.
The Capital Cycle Trap
The GPU generation refresh cycle creates, specifically, a capital trap that many neoclouds are only now beginning to fully understand. GPU generations advance every eighteen to twenty-four months. Each new generation offers materially better performance per dollar for AI training and inference workloads. The operators that can afford to refresh their GPU fleet with each new generation maintain competitive performance credentials. Those that cannot are, consequently, offering customers progressively older hardware at prices that are, in turn, harder to justify as newer alternatives become available.
The capital required for a meaningful fleet refresh at neocloud scale is substantial. Replacing 10,000 H100s with B200s or GB200s at current GPU pricing requires several billion dollars of capital, whether financed through equity, debt, or equipment leasing. Operators still servicing debt from their initial H100 fleet face a specific problem. Another round of GPU financing on top creates a leverage profile that is, notably, difficult to sustain. If revenue growth slows or rental rates compress further, that structure breaks.
Operators best positioned to navigate the refresh cycle are those with the lowest cost of capital and the strongest contracted revenue base. They also, notably, have the most diversified customer relationships. That diversification reduces the risk that a single large customer’s reduced spending creates a revenue hole during the transition period. Lenders also need confidence in their ability to deploy new hardware effectively. That operational track record is, in turn, as important as the financial metrics. Those criteria map, in turn, to exactly the same operators winning on anchor contracts and power cost. Those advantages compound, making the competitive gap between the strongest and weakest operators wider with each generation cycle.
Geographic Strategy as Competitive Moat
The geographic dimension of neocloud competitive strategy has, notably, received less attention than the technology and financial dimensions, but it is, in turn, becoming one of the most durable sources of competitive advantage available to operators who position correctly.
The Sovereign AI Opportunity
National governments across Europe, the Middle East, and Asia are, specifically, investing in sovereign AI infrastructure at a scale that was not anticipated even two years ago. National AI strategies from France, Germany, the UK, Japan, South Korea, and Singapore are driving demand for compute capacity located within national borders. Sovereign wealth fund commitments to domestic AI infrastructure add further weight. The EU AI Act is, additionally, creating requirements that favour in-country compute for certain application categories.
Neoclouds built in multiple sovereign markets are accessing, consequently, a demand pool that US-based hyperscalers cannot fully serve. Regulatory relationships, data residency capabilities, and in-country operational presence are, specifically, barriers that hyperscaler scale does not automatically overcome. This dynamic is examined across different national contexts in Neoclouds as Strategic Levers in National Digital Sovereignty. Operators positioned to capture sovereign AI demand are, notably, building a customer base that is less price-sensitive and more relationship-dependent than the commercial AI market.
The Latency Arbitrage in Edge Markets
Inference at scale is, also, creating a geographic demand pattern that favours operators with distributed infrastructure over those concentrated in a small number of large campuses. Applications serving users with latency-sensitive requirements need inference infrastructure close enough to deliver sub-100-millisecond response times. Conversational AI, real-time recommendation, and AI-assisted search all, specifically, fall into this category. That requirement is, specifically, not achievable from a single large campus serving a continental market. The physics of network latency make it impossible. That requirement is, specifically, not achievable from a single large campus serving a continental market. The physics of network latency make it impossible. That requirement cannot be met from a single large campus serving a continent. It requires, rather, distributed inference nodes in multiple metropolitan markets.
Neoclouds that have recognised this and begun building distributed edge inference capacity are, consequently, positioning for a demand pattern that will grow significantly. As AI applications move from research and back-office use cases into consumer-facing and real-time applications, that demand will, in turn, accelerate further. Building that distributed infrastructure requires a fundamentally different operational model from running a single large training campus. Operators that develop the capability to manage distributed infrastructure at scale are, in turn, building a moat that single-campus operators cannot quickly replicate.
The Consolidation Question
The neocloud sector is, in all likelihood, heading toward consolidation. A market where hyperscalers are competing aggressively and capital requirements for meaningful scale are increasing cannot sustain the current number of operators.
History suggests infrastructure markets consolidate around four or five major players, with a long tail of specialists serving specific niches.
Those specialists survive by being genuinely excellent in their chosen niche. Being the best liquid-cooled inference provider in Europe is, in turn, a viable long-term position. So is being the most reliable bare metal GPU platform for regulated industries.
Being a mid-sized GPU rental platform with no distinguishing characteristics is, however, not a viable position in a consolidating market. Those too small to compete effectively but too large to pivot quickly are, consequently, the most exposed to the consolidation pressure that is building.
Who Gets Acquired and Who Does the Acquiring
The consolidation dynamic has a specific logic. The operators most likely to be acquired are those with valuable physical assets and committed customer relationships. Geographic positions that a larger operator cannot build organically on the required timeline are, specifically, equally attractive. A neocloud with a 500-megawatt campus in a power-advantaged European market and a strong customer pipeline is, specifically, more attractive as an acquisition target. The same capacity in a saturated US market without differentiated power access commands, in turn, a materially lower valuation.
The acquirers are likely to come from three directions. Hyperscalers looking to accelerate their AI infrastructure buildout in specific markets without the development timeline of greenfield construction. Infrastructure investment firms with long capital horizons that can hold assets through the current market cycle to the longer-term AI infrastructure demand recovery. And strategic buyers from adjacent sectors, utilities, telecoms operators, and industrial companies, that see AI infrastructure as a natural adjacency to existing infrastructure positions.
Operators positioning themselves as attractive acquisition targets are, in turn, preserving optionality that pure infrastructure builders have already given up. Asset quality, customer relationships, and operational track records are what strategic buyers actually pay for. The Neocloud Sector Is About to Find Out Which Business Models Actually Work framed this as an inevitability rather than a possibility. Market conditions of 2026 are, consequently, making that framing look, specifically, prescient.
The Timeline Pressure
The consolidation timeline is, notably, compressing. Normalising GPU rental rates, rising power costs, and hyperscaler competition are, in turn, creating financial pressure on mid-tier neoclouds. The capital intensity of staying current with GPU generations adds further weight. There is no obvious self-help solution. Operators facing GPU generation refresh decisions are doing so in a more uncertain market. The return on that capital is less certain than it was when they made their previous generation commitment.
Operators who move decisively will have more options than those who wait for market conditions to force their hand.
That means securing anchor contracts, differentiated infrastructure positions, and software layers that create defensibility.
Or engaging with potential acquirers from a position of strength rather than distress.
By 2028, the neocloud market will look materially different from the one of 2024. The operators competing independently at scale will be fewer, better capitalised, and more differentiated.
The ones that did not get there will have been acquired, merged, or wound down. Those still competing independently at scale in 2028 will, ultimately, be the ones who understood in 2026 that the market had changed and acted accordingly.
Those who did not will be a footnote in the consolidation data. A very short one. That data is, in turn, already starting to accumulate. The decisions that determine the outcome are, specifically, being made right now in 2026, not in 2028 when the consequences become visible. Capital allocation, customer strategy, infrastructure positioning, and software investment are all live variables in 2026. Operators treating them as live variables are the ones that will still be operating independently when the market stabilises. The window to make the right calls is, in turn, not indefinitely open. Markets of this type do not give operators unlimited time to find their footing after conditions change. The operators who move with conviction will, consequently, have more options than those who equivocate.
