The neocloud story has been one of the most compelling infrastructure investment narratives of the past three years. Specialised GPU cloud operators raised capital at extraordinary valuations, signed multi-billion dollar offtake agreements with hyperscalers, and positioned themselves as the essential intermediary between GPU supply and AI demand. CoreWeave’s IPO in March 2026 at a valuation north of $20 billion validated the model publicly. Fluidstack’s reported $1 billion raise at an $18 billion valuation, driven by its $50 billion Anthropic commitment, extended that validation further.
What the headline numbers obscure is a margin structure that is fundamentally more difficult than the revenue trajectory suggests. The neocloud business is not a software business. It is a capital-intensive infrastructure business with hardware depreciation cycles that are shortening, electricity costs that are rising, and customer contracts that carry meaningful renewal risk. The gap between neocloud revenue growth and neocloud profitability is the most important unresolved question in AI infrastructure finance.
The Hardware Depreciation Problem
Neocloud economics rest on a simple premise: buy GPU clusters, rent them to AI customers at a margin above the cost of capital, electricity, and operations. The premise works when GPU generations last long enough to recover the capital investment before the hardware becomes competitively obsolete. It becomes difficult when hardware generations turn over faster than the depreciation schedule assumes.
The Blackwell generation entered volume production in late 2025. The Vera Rubin generation enters production in the second half of 2026. That is an approximately 18-month generation cycle for the hardware that defines neocloud competitive positioning. A neocloud operator who financed Blackwell clusters in early 2025 on a five-year depreciation schedule is now managing hardware that will be two generations behind by 2027. Customers running inference workloads at scale have strong incentives to migrate to newer hardware that delivers more tokens per dollar, which means the neocloud operator either reprices the Blackwell capacity downward to retain tenants or accepts vacancy in clusters that still carry full capital servicing costs.
The rise of inference clouds as a distinct infrastructure tier identified this dynamic early. Inference workloads are uniquely sensitive to hardware efficiency because cost per token is the primary competitive metric and newer GPU generations consistently deliver better cost per token than older ones. A neocloud whose competitive advantage rests on owning the current generation of hardware faces a continuous investment requirement to maintain that advantage, and each refresh cycle requires new capital at valuations that may no longer reflect the same investor appetite as the initial build.
The Contract Structure Carries Risks That Valuations May Not Reflect
Hyperscaler offtake agreements anchor neocloud valuations on terms that transfer more risk to the neocloud operator than the headline contract values suggest. BloombergNEF reported in early 2026 that the majority of neocloud offtake contracts with hyperscalers run on five-year terms, which sounds durable until you consider that five years spans at least two full GPU hardware generations. At the end of a five-year contract, the hyperscaler customer can renegotiate pricing against a market where hardware efficiency has improved substantially, or walk away entirely in favour of a neocloud operating newer hardware.
The neocloud operator whose revenue depends on a single major hyperscaler also carries a customer concentration risk that the diversified revenue base of a mature infrastructure business typically avoids. Fluidstack’s $50 billion Anthropic commitment is the most prominent example. That commitment is transformational in scale, but it ties Fluidstack’s financial performance substantially to Anthropic’s continued growth trajectory, its decision to renew the arrangement, and its willingness to absorb capacity at the contracted terms if its own compute requirements evolve differently than anticipated.
How neoclouds are redefining competition beyond hyperscalers outlined the competitive positioning neoclouds occupy. That positioning is real, but it creates a dependency on the continued willingness of large AI labs to outsource infrastructure rather than build it. As frontier AI labs accumulate the operational expertise to manage their own GPU clusters, the strategic calculus shifts. The lab that builds its own AI factory capability does not need the neocloud to manage its infrastructure. It needs the neocloud only for capacity overflow or workload types it cannot yet serve efficiently internally.
Electricity Costs Are Not Decreasing
The operating cost structure of a neocloud is dominated by two line items: debt service on the hardware and electricity. Hardware costs are fixed at the time of acquisition. Electricity costs are not. Global electricity prices for data center operators have risen 20 to 35 percent since 2022, and the trajectory in key neocloud markets including the United States, United Kingdom, and Western Europe shows no near-term reversal. The neocloud operator who underwrote a GPU cluster investment at 2023 electricity pricing is running that cluster at materially higher operating costs than the original model assumed.
The electricity cost exposure is structural rather than cyclical because AI workload characteristics demand high and continuous GPU utilisation. A neocloud running clusters at 90 percent utilisation around the clock has no off-peak periods during which to benefit from time-of-use pricing, no opportunity to shift consumption to renewable-favourable hours without disrupting training jobs, and no ability to reduce power draw without throttling the revenue-generating output of the facility. Every percentage point increase in electricity price flows directly to the operating cost line with no operational lever to offset it.
The neocloud mindset as it has evolved is increasingly oriented toward operational precision as the margin protection mechanism. Neoclouds that achieve higher effective GPU utilisation, better workload scheduling efficiency, and lower cooling and power overhead per token generated can absorb electricity price increases that less efficient operators cannot. That operational differentiation is real, but it requires investment in software and operational capability that adds cost to the business model and takes time to develop.
Where the Differentiation Has to Come From
The neoclouds that survive the margin compression of the current cycle will be those that build defensible differentiation beyond hardware access. Hardware access is not defensible. Any operator with sufficient capital can procure the same GPUs from the same supply chain. The neocloud whose value proposition rests primarily on owning current-generation hardware will face commoditisation pressure as the hardware generation turns over and competitors emerge with equivalent or newer infrastructure.
The neocloud software stack as the real differentiation layer identified the software and operational capabilities that separate neoclouds with sustainable unit economics from those competing purely on hardware. Workload scheduling optimisation, model-specific inference serving, bare-metal provisioning speed, and networking fabric performance that exceeds what hyperscaler standard configurations provide are capabilities that customers pay for and that are genuinely difficult to replicate. The neocloud that builds those capabilities alongside its hardware investment is building a business. The neocloud that treats those capabilities as secondary is building a trade.
How the Market Is Already Separating Along That Line
The market is beginning to separate along that line. CoreWeave’s IPO narrative leaned heavily on its software and operational capabilities, not just its Nvidia GPU inventory. Fluidstack’s Anthropic relationship reflects a model where custom facility design and operational commitment to a specific customer’s workload requirements creates switching costs that pure hardware rental does not. Those are early signals that the neoclouds most likely to sustain margins are those who have moved furthest from the simple hardware rental model toward something that looks more like a specialised infrastructure partner.
The neocloud margin problem is not a reason to dismiss the sector. It is a reason to be precise about which neoclouds have built businesses with structural defensibility and which have built revenue lines that will compress as hardware cycles turn, electricity costs rise, and hyperscaler customers gain the operational maturity to negotiate more aggressively at contract renewal. The distinction matters enormously at the valuations the sector currently commands.
