Renting intelligence: margin, or a margin illusion?

01The shortage that built a cloud tier

The neocloud exists because the hyperscalers could not build fast enough. When frontier-model training and then inference demand outran every capacity plan, a new tier of specialist GPU clouds grew in the gap, and grew violently: the segment cleared $25B of revenue in 2025, up 223% year-on-year in the fourth quarter, with credible forecasts approaching $400B by 2031. Six names have scale; behind them, more than 300 providers entered the H100 rental market in 2025 alone. That last number is the tell. This is a business with almost no barrier to entry at the commodity end, anyone with capital and a Nvidia allocation can stand up GPUs, and the price curve shows it. The demand side of this market, token volumes, agentic workloads, and how the curve evolves to 2030, is sized in our companion note, The Inference Economy; this piece is about who keeps margin while serving it.

02The framework: scarcity decays at different speeds

The organising mechanism for the whole AI Infrastructure sector is simple: value pools wherever scarcity is durable, and every scarcity in this stack decays at a different speed. Power and grid connections decay slowest (a decade). Data-centre shells are next (years). GPUs decay fastest of all, not just because supply catches up, but because the asset itself depreciates: each new silicon generation devalues the last, and each price cut reprices the installed base. The neocloud's tragedy is structural: it sits on the fastest-decaying scarcity in the stack, holding the depreciating asset, funded by debt that outlives the asset's pricing power.

Within the tier, two distinct business models have emerged, and they should be underwritten as different asset classes, because they are.

The market is paying software prices for the second model ahead of public proof. The archetypes re-rated violently through 2025-26: one serving platform went from a $4B valuation to $17.5B in nine months on ~$1B of ARR, and the leading platforms now carry multi-billion marks on revenue around the billion-dollar scale. What the tier has not yet shown publicly is a software gross margin: where numbers have surfaced, they sit in the mid-40s, solid infrastructure economics, not yet the 70-80% the multiples imply. That gap is the central diligence question, and it is answerable rather than rhetorical, because the bull mechanics are real. The best platforms employ the researchers who wrote the defining inference-efficiency kernels, compound that into a genuine cost-per-token lead (customers report multi-fold serving-cost reductions on switching), and are now vertically integrating into owned power and capacity at hundreds-of-megawatts scale to convert the cost lead into margin. That is the AWS trajectory in miniature. Whether it holds turns on whether the serving-software edge stays proprietary faster than open-source stacks absorb it, the moat is the research team's velocity as much as any artifact.

03The economics: depreciation is the whole debate

Every argument about this subsector reduces to one accounting line, so we tested it against market data rather than opinion, in a five-generation study of rental-price decay. Rental prices across five chip generations decay at a consistent ~15-18% per year against launch-era rates; cash-flow life runs to roughly eight or nine years (Azure retired its V100 fleet at 7.5); and about two-thirds of a GPU's lifetime cash flow arrives in its first three years. That evidence cuts both ways. The sceptics' claim that a rented GPU is economically dead in two-to-three years, the basis for a much-cited $176B understated-depreciation estimate, is not supported: old chips still rent at multiples of their running cost. But the books do not get off clean either: straight-line depreciation over five-to-six years charges ~17% of the asset per year while the economics consume ~28% in year one, so early-year reported profits, which is all this young sector has yet shown, are flattered by roughly 1.6x on the depreciation line. Amazon's own move (cutting server lives from six to five years, a $677M nine-month hit) points the direction. The sharpest finding is about vintage rather than life: re-based to the 2023 scarcity peak, the same chips show ~24%/yr decay. The worst outcomes in this tier come not from machines dying young but from capacity bought, leased or debt-financed at panic prices.

The financing stack compounds the problem. The listed wholesale leader added roughly $21.9B of debt in the first half of 2026 alone, at 8.5-9.75% on unsecured notes, against an asset whose rental price has fallen 64% from peak. Interest expense now consumes nearly half of adjusted EBITDA. And the capital keeps coming partly because the silicon vendor recycles it: a $2B direct investment into that same leader, stakes across the tier, over $540B of announced partnership and financing arrangements in 2026, and a discussed $250B lease-payment guarantee for its largest end-customer. The circularity is not a scandal; it is a structure, but it means headline demand signals (backlogs, bookings, round sizes) are partially manufactured by the supply chain that benefits from them, and should be discounted accordingly.

04How to underwrite a name in this subsector

This is the checklist we apply to any neocloud or inference platform that crosses our desk, private allocation or public position. It is designed to separate the contracted-cashflow story from the commodity-rental story, and the software claim from the software reality.

• Contract cover. What share of deployed capacity is under committed, take-or-pay contract, at what weighted duration, and what share is merchant? Merchant exposure is exposure to the $0.99 spot print.
• Customer quality and concentration. Who actually pays, hyperscalers and profitable enterprises, or burn-funded AI startups whose own runway is a venture-market variable? A backlog is only as good as its weakest counterparty. Anything above ~30% single-customer concentration is a structural, not incidental, risk.
• Margin truth. Demand gross margin after depreciation on an economics-matched schedule (declining-balance at the observed ~15-18%/yr rental decay, not straight-line). For platforms: is software priced and reported separately from compute resale, or blended into one number the multiple flatters? And what is the margin by layer, token API vs dedicated endpoints vs cluster rental?
• Price-basis risk. What happens to the model at renewal, when today's contracted $/GPU-hour resets to the then-current curve? A 5-year contract signed at 2024 prices is an asset; its renewal is the business.
• Capital stack vs the cash-flow curve. A GPU delivers two-thirds of its lifetime cash in its first three years; 9%+ debt serviced from the thin end of that curve in later years demands utilisation and pricing hold simultaneously. And check the vintage: capacity bought at scarcity peaks decays ~24%/yr, not 15-18%.
• Vendor circularity. Is the chip vendor in the cap table, in the backlog, or both? Discount demand signals that the supply chain itself is financing.

05Orien's verdict

Signposts we track

• A platform discloses real gross margin. The first credible public print of an inference platform's software-layer margin settles the subsector's central question, in either direction.
• Blackwell-generation pricing. If GB200-class rental holds price materially better than the H100 curve did, the depreciation bear case softens; if it repeats the curve, it confirms.
• A credit event in the long tail. Sub-break-even spot pricing plus 9%+ debt across a 300-provider long tail should eventually produce one; it would reprice the whole tier and reset entry marks.
• Enterprise mix rotation. The platform bull case shows up as lengthening contracts and a rising share of enterprise revenue versus AI-native startup spend. Flagship customer lists still skew to the latter; whether that rotates is the software claim materialising in the revenue mix.
• Renewal prints. The first big wholesale contracts signed in 2023-24 come up for renewal from 2028; the re-contracting price is the truth serum.

06What breaks this call

The software margin proves real. If the leading platforms demonstrate durable 60%+ gross margins as token prices fall, genuine software economics on commodity FLOPs, the layer re-rates as software and our caution will have cost us the best names at pre-proof prices. This is the key upside risk, and the reason our stance is selective rather than avoid.

Scarcity lasts longer than we model. If power constraints, the subject of our forthcoming powered-land deep-dive, keep effective GPU supply tight through 2028+, rental prices stabilise, utilisation stays high, and the levered wholesale model works long enough to deleverage. The neocloud bear case is a supply forecast; supply forecasts have been wrong in both directions.

The vendor keeps underwriting the tier. $540B+ of announced vendor financing can defer the cycle for years. Circularity is a risk that pays out slowly, being right early here has the same P&L as being wrong.

Demand goes vertical again. A step-change in inference demand (agents, video, robotics workloads) could outrun even the current buildout, restoring pricing power across the tier. In that world the commodity is scarce again and everything in it re-rates. The token data in The Inference Economy suggests this is the upside scenario to take most seriously.

07Sources