Neocloud sector overview

A neocloud is a cloud-infrastructure company purpose-built around Nvidia GPUs (and, to a lesser extent, alternative accelerators) to deliver high-density, high-performance compute for training and inference. The value proposition versus the incumbent hyperscalers (AWS, Azure, Google Cloud) is threefold: faster access to the latest Nvidia silicon, higher GPU utilisation through software and cluster design optimised for AI, and lower cost per unit of useful AI compute because the stack is stripped of the “bells and whistles” of general-purpose enterprise cloud. In effect, neoclouds exist because demand for AI compute has outstripped what even the hyperscalers can deploy, creating room for specialists to arbitrage GPU access, speed and utilisation.

Synergy Research Group forecasts the neocloud market to approach $400 billion by 2031, growing at better than a 50% compound rate off a small 2024 base, driven by GPUaaS, generative-AI platforms and high-density data-centre capacity. That sits inside a broader AI-infrastructure capital cycle of extraordinary scale: J.P. Morgan estimates hyperscaler capex will reach roughly $697 billion in 2026, nearly double 2025 levels, with Microsoft, Amazon, Alphabet and Meta each spending more than $100 billion individually and capital intensity reaching an historically unprecedented 45–57% of revenue.The demand signals underpinning these forecasts are, for now, unusually concrete. Cloud backlogs are large and growing; enterprise AI adoption is broadening from experimentation into production inference; and every major hyperscaler reports that capacity is being absorbed as fast as it can be deployed. Tellingly, Microsoft’s roughly $80 billion of unfulfilled Azure backlog is described as a function of power availability rather than weak demand. Neoclouds capture the overflow: when Meta signs up to $62 billion of neocloud agreements, or when Nvidia itself contracts with IREN for $3.4 billion of AI cloud capacity, they are routing demand that the hyperscalers physically cannot serve fast enough. Conventional multiples are close to meaningless here: every listed neocloud is GAAP-lossmaking, so P/E and EV/EBITDA on trailing numbers do not apply. The market instead values the group on contracted backlog, forward revenue and power pipeline

Three macro forces determine whether the neocloud thesis compounds or cracks. We treat them as the real “factors” driving the group more than any company-specific execution. (i) The AI capex cycle demand engine and single biggest risk (ii) The cost and availability of capital the rate regime bites here (iii) Power the hard physical ceiling

For an investor expressing the AI-infrastructure theme, the neocloud sector offers the most direct and most volatile exposure available in public markets. Our framework for navigating it: ▪ Select on balance sheet, not growth. Growth is uniformly extraordinary and therefore not a differentiator; survivability under a funding shock is. In a higher-for-longer rate regime, prefer low-leverage, prepayment-funded models (Nebius’s asset-light approach) and be more cautious on maximally debt-funded builds (CoreWeave) despite their superior scale sizing the latter for the volatility it has already shown. ▪ Respect that Nvidia captures the economics. The most reliable way to own the neocloud demand curve may be to own the arms dealer. Neoclouds are the levered, lower-moat middle of the value chain; Nvidia is the scarce input and a beneficiary of every neocloud’s capex. A barbell core exposure to the supplier, satellite exposure to the best-financed neocloud is more robust than concentrating in the middle layer. ▪ Treat contracted power and backlog coverage as the quality screen. Favour names whose contracted power is converting to live capacity on schedule and whose backlog counterparties are creditworthy (hyperscalers, Nvidia itself) rather than thinly capitalised AI startups. The gap between contracted and active GW (Figure 4) is where both the upside and the execution risk sit. ▪ Watch the macro triggers more than the earnings. The 2027 hyperscaler capex guides, the Nvidia Rubin ramp, the rate path, and any visible financing-stress event (a failed data-centre financing, a delayed lease) will move the whole group more than any single company’s quarter. This is a top-down sector wearing bottom-up clothing. Key debates that will decide the sector’s multi-year return: (1) Does hyperscaler/AI-lab demand stay ahead of the wave of new capacity, or does 2027 bring the first air-pocket? (2) Do neoclouds earn an adequate through-cycle return on fast-depreciating, debt-financed GPUs, or does price competition compress it? (3) How much of current demand is genuinely independent versus circular Nvidia-funded recycling? (4) Is contracted power a durable moat or merely a temporary scarcity rent? Our base case is constructive on demand into 2026 but increasingly cautious on the financing and returns questions into 2027 a stance that argues for participation through the highest-quality balance sheets rather than the whole basket.