The Neocloud Order: Economic Structure and an Anti-Liberal Reading

Part of the Cloud Economics and Anti-Liberal Policy series — reading the AI build-out as political economy, not just capex. Companion pieces: Universal Nexuspolis (the governing frame this report applies), The Grammar and the Substrate, The New Commons, The Grid Commons, and the balance-sheet mechanics in The AI CapEx Contagion Model and the AI data-center deep dive.

Coverage runs through September 12, 2026. Every load-bearing figure is dated and attributed; vendor/self-reported figures are flagged and distinguished from SEC-filed or regulator-filed data.

TL;DR

The neocloud sector — AI-specialized GPU cloud providers led by CoreWeave, Nebius, Crusoe, Lambda, Nscale, and bitcoin-miner conversions like IREN/TeraWulf/Cipher — is roughly a $48bn annualized-run-rate category (mid-2026, MeasuredAI estimate, analyst-reported not audited) whose growth is real but whose capital structure is genuinely novel and fragile: it rests on GPU-collateralized debt, off-balance-sheet SPVs, and a web of vendor guarantees that let small operators borrow on near-utility terms. Morgan Stanley’s July 2025 report projects “a $800bn+ growth opportunity around data center investment via private credit, led by asset-based finance,” inside a ~$1.5 trillion financing gap through 2028.
Core empirical findings: CoreWeave’s revenue backlog reached ~$104bn by mid-2026 but Microsoft was ~67% of FY2025 revenue; GPU-backed debt went from a $2.3bn private-credit experiment (Aug 2023) to an investment-grade-rated asset class (A3/A-low, March 2026); Nvidia both owns ~11.5% of CoreWeave and is contractually obligated to buy its unsold capacity ($6.3bn, through April 2032); H100 rental rates fell ~64–75% from 2024 peaks; and PJM capacity prices rose ~10x in two years, with data-center load accounting for 63% ($9.33bn) of 2025/26 capacity costs. BIS and IMF have both flagged the buildout’s leverage and concentration as financial-stability concerns.
The anti-liberal thesis — that the most advanced sector of the economy is organized on non-liberal principles (state allocation via export controls, vendor cartelization, sovereign capital, socialized infrastructure risk) while keeping liberal vocabulary — is substantially defensible at the substrate layer (power, land, chips) but fails at the abstraction layer, where entry is genuinely open (300+ new GPU providers in 2025), prices are falling, and hyperscaler dominance is being partly eroded. The honest verdict is a hybrid split by layer: liberal-competitive on the surface everyone touches, dirigiste-cartelized at the base where ownership and power actually reside.

PART ONE: THE ECONOMIC CLIMATE

1. Market structure and size — the category is real but definitionally unstable

“Neocloud” first appeared in analyst reports in late 2024 (per TrendingTopics). Definitions vary; the most useful taxonomy (MeasuredAI, mid-2026) is a three-layer stack: Layer 1, powered-shell landlords — predominantly converted bitcoin miners (TeraWulf, Cipher, Hut 8, Core Scientific, Applied Digital, Galaxy, IREN) that own what is scarce: energized land, interconnection positions, substations; Layer 2, the neoclouds proper (Crusoe, Nebius, Nscale, CoreWeave, Lambda, Fluidstack) that sign 10–15-year leases and fill shells with GPUs; and Layer 3, the spot-market/marketplace resellers.

MeasuredAI estimates category-wide revenue at ~~$12bn/quarter (~~$48bn annualized) as of mid-2026, forecasting ~$300bn by 2030 (~50% CAGR) — an analyst forecast, not independently verified. The elite tier (six named operators) is collectively valued at over $150bn. The long tail is large and growing: over 300 new GPU cloud providers entered in 2025 (Introl). The tier distinction matters enormously for risk: tier-1 operators hold long-dated contracts with named counterparties; the tail sells into a volatile spot market where falling on-demand prices directly erode collateral value.

2. Capital structure — the core of the report

The origin. GPU-backed debt began with CoreWeave’s $2.3bn facility (Aug 3, 2023, led by Magnetar Capital and Blackstone, collateralized by Nvidia H100s — the first time H100 hardware served as collateral, per Reuters).

The escalation. It scaled fast: a ~$7.5bn facility (May 2024, Blackstone-led with Magnetar); a $2.6bn OpenAI-contract-backed facility (July 31, 2025, priced SOFR+4%, secured on “substantially all assets of CoreWeave Compute Acquisition Co. VII, LLC”); and the landmark $8.5bn DDTL 4.0 (March 31, 2026). Per CoreWeave’s press release, DDTL 4.0 “received ratings of A3 by Moody’s and A (low) by DBRS, respectively, representing the first investment-grade rated financing secured by HPC infrastructure and an associated customer contract”; it was priced SOFR+225bps floating / 5.9% fixed and held in SPV CoreWeave Compute Acquisition Co. VIII, LLC against a $14.2bn Meta contract (GlobalDataCenterHub). CoreWeave then extended the model with the $3.1bn DDTL 5.0 (May 18, 2026, Ba2 Moody’s / BB+ Fitch — the first publicly syndicated HPC-backed loan) and the $2.6bn DDTL 5.5 (Aug 10, 2026), which for the first time financed ~3-year contracts against a ~5-year maturity, meaning lenders were underwriting renewal risk. CoreWeave also issued €2bn of 8.5% bonds due 2032 (June 2026).

The scale. CoreWeave total debt exceeded $21bn at end-2025 (up from under $8bn in 2024). Its Q2 2026 10-Q shows total recourse debt net of discount of ~$31.4bn and additional non-recourse debt including the DDTL 4.0 tranche; the OEM/software financing arrangements alone (11% coupon) exceeded $4bn.

The mechanism. Bankruptcy-remote SPVs into which a defined pool of GPUs and a named offtake contract are ring-fenced. As GlobalDataCenterHub put it of DDTL 4.0: “Moody’s and DBRS rated the collateral not CoreWeave. It is Meta’s credit and the residual value of Blackwell silicon, ring-fenced away from the operating company’s own fragility.” The Meta–Blue Owl “Hyperion” structure runs the identical logic at hyperscale, keeping tens of billions of debt off Meta’s parent balance sheet.

Who lends. Private credit funds — principally Blackstone, Blue Owl, Apollo, PIMCO, BlackRock, plus Magnetar, Macquarie, and Brookfield — originate most data-center debt (Quinn Emanuel). Banks are reluctant because regulatory risk-weights punish fast-depreciating hardware collateral, which created the private-credit opening.

The total scale and its growth. Estimates differ and I flag the conflict. The BIS (Bulletin No. 120) reports that private-credit funds originated over $40 billion in loans to AI-related companies in 2025, versus roughly $3 billion in 2010 — a growth figure. Quinn Emanuel cites outstanding AI-related private-credit loans “surging from near zero to over $200 billion in just a few years” (a stock figure). Morgan Stanley’s July 2025 report “Bridging a $1.5tr Data Center Financing Gap” frames it best: cumulative global data-center capex of ~$2.9tn through 2028, of which hyperscalers cover ~$1.4tn, leaving a ~$1.5tn gap filled by “$800bn+ … via private credit, led by asset-based finance,” ~$200bn of IG bond issuance, and “$150bn of DC ABS/CMBS issuance through 2028.” JPMorgan projects data-center securitization of $30–40bn/yr in 2026–27 (up from ~$27bn in 2025).

The central risk — the amortization–depreciation race. Most neocloud offtake agreements run just 12–18 months (Optio Investment Partners, via Debtwire) while facilities extend ~5 years; rating agencies apply steep haircuts to revenue projected past the contract term. Optio’s David Lindström compared GPU financing to “where oil was before we had a futures market” — there is no established way to hedge a customer walking. If utilization or rental rates fall, the collateral (GPUs) decays on a contested curve while the debt remains; the equity is wiped first, then senior secured lenders face a fast-depreciating asset in a falling spot market. This is why one analyst characterized aggressive versions of the structure as “an equity-like wager dressed up as senior secured debt.”

3. Circular / vendor financing — documented, large, mostly one-directional

The web is real and each leg is a genuine arm’s-length contract:

Nvidia → CoreWeave: owns ~11.5% (47,213,353 Class A shares, per Schedule 13G/A as of Jan 23, 2026, after a $2bn top-up in January 2026 at $87.20/share; up from ~7% in mid-2025). Separately, per CoreWeave’s 8-K (order dated Sept 9, 2025, initial value $6.3bn), “in instances where the Company’s datacenter capacity is not fully utilized by its own customers, NVIDIA is obligated to purchase the residual unsold capacity through April 13, 2032.” This is a signed obligation, not an LOI — Barclays called it a “backstop.”
Nvidia → Nebius: $2bn investment (March 11, 2026, ~8.3–9.3% stake, structured as pre-funded warrants); partnership targets Nebius deploying >5 GW of Nvidia capacity by end-2030.
Nvidia → Nscale, Crusoe, Lambda: participated in Nscale’s Sept 2025 $1.1bn round and Oct 2025 $433m SAFE; Crusoe’s $600m Series D (Dec 2024) and $1.375bn Series E (Oct 2025, ~$10bn valuation); Lambda’s Series D/E. Nvidia separately leases back ~~18,000 GPUs from Lambda (~~$1.5bn) — an offtake, not equity. Exact equity percentages in these private firms are undisclosed.
Nvidia → OpenAI: the “up to $100bn” (Sept 22, 2025) was always a non-binding letter of intent — CFO Colette Kress on Dec 2, 2025: “we still haven’t completed a definitive agreement” — and was superseded by a definitive $30bn equity investment (Feb 27, 2026), with Jensen Huang stating the $100bn was “probably not in the cards” now that OpenAI is going public.
AMD → OpenAI: warrant for up to 160m AMD shares at $0.01 (issued Oct 5, 2025), vesting against a 6 GW purchase commitment; if fully exercised, OpenAI could own ~10% of AMD. AMD stock rose ~24% on the news.
Microsoft: neocloud offtake commitments totaling ~$60bn (Nscale ~$23bn, Nebius ~$19.4bn, CoreWeave >$10bn, IREN $9.7bn, Lambda >$2bn), plus up to $5bn to Anthropic (Nov 2025).

The round-tripping case (at strength): The same dollar appears in multiple places. Microsoft’s AI capex flows into CoreWeave (62% of 2024 revenue), which buys Nvidia silicon (booked as Nvidia revenue), which raises Nvidia’s stock (raising its CoreWeave stake). Critics warn this inflates apparent demand and, when demand disappoints, magnifies losses — the IMF flagged it, and the “same dollar in five places” critique is that aggregate “AI demand” figures double-count intra-cohort spend.

The vendor-finance counter-case (at strength): As Noah Smith argues, revenue flows one way in the load-bearing legs — OpenAI pays Nvidia for chips it genuinely needs for its core business; both firms are “just doing what they’re set up to do.” These are scrutinized public companies bound by GAAP and disclosure rules. CoreWeave CEO Michael Intrator called circularity claims “ridiculous,” noting Nvidia’s ~$300m across two rounds for a ~7% stake was too small to prop up operations. Vendor financing built the railroad and telecom eras. Adjudication: it is not fraud and not classic round-tripping (revenue is real and one-directional at the core), but “real” and “additive” differ — a meaningful fraction of headline demand is intra-cohort, and the structure magnifies losses if end-demand from outside the circle disappoints. The number to watch is revenue from counterparties who are not in the loop.

4. Depreciation and useful-life controversy — unresolved, evidence mixed

Michael Burry (Nov 11, 2025) alleged hyperscalers understate depreciation by ~$176bn over 2026–28 by stretching GPU useful life to 5–6 years against a 2–3-year product cycle — “one of the more common frauds of the modern era” — claiming Oracle and Meta profits are overstated ~27% and ~21% by 2028. He disclosed ~$187m notional puts on Nvidia and ~$912m on Palantir. Meta’s own disclosure corroborates the mechanism’s materiality: extending server/network life to 5.5 years cut 2025 depreciation ~$2.9bn.

Counter-evidence: A100s (introduced 2020) still rent and retain residual value six years on; Jensen Huang publicly cited continued A100/H100 utilization. The synthesis (Interesting Engineering, DeepQuarry) is that Burry conflates two timelines: the technological-obsolescence cycle for frontier training (18–36 months, where he is right) versus the economic-utility cycle of a chip cascading from training → inference → small-model serving → rendering (plausibly 5–6 years). Both GAAP and IFRS permit judgment-based useful-life estimates, making the 3–6-year range defensible but also manipulable. The honest position: Burry is directionally right that reported earnings are flattered relative to true economic depreciation, but wrong that older chips are near-worthless. The risk is a future impairment “cliff” if renewal demand for a given generation collapses faster than the depreciation schedule assumes.

5. GPU price trends — falling on-demand, tightening contracted

H100 on-demand rates fell ~~64–75% from 2024 peaks (~~$8/hr, and >$12/hr on hyperscalers in 2023) to ~$1.65–3.50/hr by 2026; spot as low as $1.20/hr; AWS cut P5 (H100) prices ~44% in June 2025. The spread is now enormous — ~$1.38–1.49/hr on marketplaces vs ~$11–12/hr on hyperscaler on-demand (IntuitionLabs). Break-even for ownership is ~14–16 months at 100% utilization; the provider profitability floor is ~$1.65/hr (Introl).

But contracted/reserved capacity tells the opposite story: SemiAnalysis’s one-year H100 rental index rose ~~36% May 2025–May 2026 (~~$1.95 → ~$2.65), and reserved one-year pricing bottomed at $1.70/hr (Oct 2025) before climbing ~40% to $2.35/hr (Mar 2026). The bottleneck moved from silicon to HBM memory (36–52-week lead times; a 3.6m-unit backlog). Generationally, B200 delivers ~2.5x H100 training throughput, so cost-per-result can favor newer chips despite higher hourly rates — meaning the right metric is cost-per-token/training-run, not $/GPU-hour. Implication: falling on-demand prices threaten the spot-market tail’s collateral values; scarce contracted capacity protects the tier-1 operators — a bifurcation the “GPU glut” narrative misses.

6. Customer concentration and counterparty risk

Per CoreWeave’s S-1, its top two customers were 77% of 2024 revenue (Microsoft 62%, Nvidia); Microsoft was ~67% of FY2025 revenue (up from 62% in FY2024, per the 10-K), and ~72% in 1H2025. Counterparties are increasingly pre-profit AI labs. Nebius signed a $27bn Meta agreement (per its 6-K: $12bn of dedicated 5-year GPU clusters from early 2027 plus $15bn of optional capacity, with Meta obligated to buy unsold capacity). CoreWeave’s OpenAI relationship reached $22.4bn total; Meta commitments up to ~$35.2bn. Of CoreWeave’s ~$103.7bn RPO, only ~41% is expected to convert within the near window (Motley Fool, citing disclosure). A “refinancing wall” of GPU-collateralized loans from CoreWeave, Nebius, Lambda, Crusoe, and Applied Digital comes due 2026–28. The counterparty quality issue is sharp: the largest infrastructure buyers (OpenAI, Anthropic) are venture-backed and pre-profit, so much of the “contracted backlog” is only as good as the labs’ continued ability to raise capital.

7. The physical substrate — power, land, water, grid

Demand forecasts (and their contestedness). LBNL (2024 report): US data centers were 4.4% of electricity in 2023 (176 TWh), projected 6.7–12% by 2028 (325–580 TWh). LBNL’s 2025 Reference Case: 464 TWh (2028), 649 TWh (2030, ~11.8%), with a scenario range to 843 TWh. IEA: global 415 TWh (2024) → ~945 TWh (2030); US +130%. These forecasts are wide and grid planners treat them cautiously — PJM itself now excludes speculative projects via “transmission security agreements.”

Ratepayer cross-subsidy and PUC rulings. PJM’s 2027/28 capacity auction cleared at the FERC price cap of $329.17/MW-day, roughly 10x the $28.92 of 2024/25; the Dec 2025 auction hit $333.44/MW-day while falling 6,625 MW short of reliability requirements for the first time in PJM’s history. Per PJM’s Independent Market Monitor (Monitoring Analytics), data-center load accounted for 63% of 2025/26 capacity costs — $9.33bn of $14.69bn. PJM’s 2026 forecast attributes 94% of 2024–30 peak-load growth to data centers. NRDC projects ~$70/month household increases by 2028 and $100–163bn cumulative excess PJM costs through 2033. State response: 23 states had approved large-load tariffs by May 2026 (EEI); Pennsylvania’s PPL settlement (filed March 13, 2026) adds a data-center tariff and a $275m base-revenue increase, explicitly citing “risk of stranded assets, unrecovered costs, and cross-subsidization from other ratepayers” (PPL’s pipeline is ~20 GW of large loads against a 7.8-GW peak). Some tariffs require 14-year minimum contracts, payment of 85% of contracted transmission / 60% of generation demand, and $1.5m/MW collateral. Notably, PG&E claimed data-center growth helped cut its rates ~11% since 2024 — a contested counter-example (Utility Dive, Feb 2026).

Nuclear/behind-the-meter. Microsoft’s 835 MW Three Mile Island / Crane restart (20-yr PPA, ~$16bn, DOE $1bn loan closed Nov 2025, FERC transmission waiver June 2026, targeting H2 2027); Amazon’s 1.92 GW Susquehanna PPA through 2042 (revised from behind-the-meter after utility opposition); ~9.8 GW committed across 13 hyperscaler nuclear deals, only ~1.92 GW operational as of mid-2026.

Stranded-asset risk falls partly on ratepayers where speculative load inflates capacity procurement that is never refunded if the data centers do not materialize — the explicit fear in the PPL and Sierra Club filings.

8. The state is constitutive, not incidental

Export controls as allocation. Nvidia’s H20 required export licenses (April 9, 2025), triggering a $4.5bn Q1 FY2026 charge for excess inventory; the Q2 outlook flagged an $8.0bn revenue loss. Per Nvidia’s FY2026 10-K, “as of the end of fiscal year 2026, we were effectively foreclosed from competing in China’s data center computing/compute market.” The Biden-era AI Diffusion Rule (Jan 2025) was rescinded (May 13, 2025); H200 sales to China resumed on a case-by-case basis with a 25% export fee plus a ~15% revenue-share expectation (Dec 2025). The net effect: where frontier compute gets built is set by US licensing, not by market price.

Sovereign AI and state capital. UAE: MGX ($100bn fund, launched March 2024, backed by Mubadala and ADQ with G42), the 5 GW UAE–US AI Campus, and Stargate UAE (1 GW built by G42, operated by OpenAI/Oracle); the US agreed (May 2025) to allow up to ~500,000 advanced Nvidia chips/year. Saudi HUMAIN (wholly-owned PIF subsidiary, launched May 2025; 200,000-GPU Nvidia partnership Nov 2025). Every G20 economy now has a sovereign-AI budget line: France (Mistral), UK (Nscale, Europe’s largest homegrown neocloud, $14.6bn valuation March 2026), India (IndiaAI), Japan, Korea, plus Switzerland’s fully-open Apertus. Crucially, all Gulf/EU compute runs on US-approved silicon — sovereignty conditional on Washington’s licenses. MGX co-led the $40bn Aligned Data Centers acquisition with BlackRock.

Public compute. US: NAIRR pilot (NSF + 13 agencies, 28 industry partners) and a NAIRR Operations Center funded at up to $35m over 5 years. EU: EuroHPC AI Factories, InvestAI, and a Frontier AI Grand Challenge (Feb 2026) whose single winner receives up to 2.5% of EuroHPC capacity. These are orders of magnitude below private capex.

9. The bubble/no-bubble debate — with institutional voices

Institutions are now on record. BIS (March 2026 Quarterly Review, “Financing the AI infrastructure boom”): hyperscaler corporate-bond gross issuance topped $100bn in 2025, mostly long-dated, while CDS spreads rose “especially for hyperscalers with lower credit ratings, reflecting both the volume of supply and uncertainties around the projects’ payoffs.” Moody’s (July 24, 2026): “unprecedented” AI spending threatens credit quality across Microsoft, Amazon, Alphabet, Meta, Oracle and CoreWeave; capex projected $785bn (2026) → ~$1tn (2027); Oracle Baa2 negative (two notches above junk), CoreWeave Ba3; Moody’s counts hyperscaler lease obligations as debt-equivalents. IMF GFSR (Oct 2025 and 2026): stretched valuations, US equities ~55% of the global market, rising NBFI leverage, ~40% of private-credit borrowers with negative free cash flow (up from 25% in 2021), and explicit warnings that concentration in a few compute/model providers is a channel from AI risk to financial-stability risk.

The bear case (at strength): The 1999 telecom/dark-fiber analogy — $500bn of fiber, mostly dark for a decade, Global Crossing bankrupt 2002 — driven by a “traffic doubling every 100 days” narrative and vendor financing (Cisco/Lucent/Nortel), which rhymes with today’s depreciation-flattered earnings, private-credit opacity, circularity, and demand-forecast fragility. Current single-year hyperscaler AI capex ($320bn+ in 2026, ~65–70% AI-related) exceeds the entire cumulative dot-com telecom buildout.

The bull case (at strength): Unlike 1999, the leaders are profitable incumbents; OpenAI reached $1bn revenue within nine months and ~$25bn ARR by Q1 2026 (Modelist, piecing together private figures); contracted capacity is genuinely scarce (SemiAnalysis index rising); well-contracted operators show 25–30%+ target margins; and compute is a real, utilized input, not dark inventory. KKR and Goldman argue the buildout compounds for a decade (Goldman’s baseline: ~$765bn AI capex in 2026 → ~$1.6tn in 2031).

Falsification thresholds are given in Recommendations below.

PART TWO: INTERPRETATION THROUGH NEXUSPOLIS / ANTI-LIBERAL PHILOSOPHY

The thesis under test

That the neocloud order is not a spontaneous market outcome and cannot be fully explained by liberal market theory — that it is produced by (a) state allocation of the key input via export controls, (b) vendor cartelization (Nvidia as simultaneous supplier, shareholder, and demand-backstop), (c) sovereign capital, and (d) socialized infrastructure risk (ratepayer cross-subsidy, public grid buildout, export-control protection) — so that the most advanced sector of the economy is organized on non-liberal principles while retaining liberal vocabulary. I argue it at strength, present the strongest counterargument, and adjudicate without simply confirming it.

Burnham — the managerial thesis and the formal/real distinction

James Burnham (The Managerial Revolution, 1941) predicted the separation of ownership from control and the rise of a class that controls but does not own the means of production, fused with the state. The neocloud maps onto this with unusual precision: as MeasuredAI titled its analysis, neoclouds are “AI’s $150 billion middle layer that owns almost nothing.” Ownership is fractured across Nvidia (equity + demand backstop), private-credit SPVs (which effectively control the ring-fenced assets), converted miners (land/power), and hyperscaler/lab offtakers (the cash flows); control sits with a managerial operator owning almost none of it. This is Burnham’s separation rendered as a capital structure.

Burnham’s Machiavellians (1943) distinction between formal meaning (the ideological description) and real meaning (the actual function) sharpens the reading. The formal description: competitive firms renting commoditized compute at falling prices in an open market. The real function: a mechanism converting sovereign and vendor capital into physical control over scarce power and silicon, with risk socialized onto ratepayers and bondholders and reward captured by the vendor-shareholder — Nvidia sells GPUs at full margin on day one, then backstops the demand. Where it fails to fit: Burnham’s managers were salaried technocrats displacing capitalists; here the dominant actor (Nvidia) is a capitalist owner using managerial operators as pass-through vehicles. The neocloud CEO is an intermediary, not the new ruling class. The ruling formation is the state–vendor–sovereign-fund complex — a fusion Burnham would recognize, but with the owner, not the manager, on top. This is an evidentiary fit for the separation thesis and a partial miss on the class-identity thesis.

Polanyi — fictitious commodities and the double movement

Karl Polanyi (The Great Transformation, 1944) held that land, labour, and money are “fictitious commodities” whose full commodification triggers a protective “double movement.” The neocloud extends the list to compute, electricity, and water as substrates of intelligence. Electricity is the clearest: a regulated public good being pulled into merchant procurement for private AI load. The double movement is empirically visible and fast: the PJM ratepayer revolt ($9.33bn cross-subsidy), 23 states legislating large-load tariffs, six-plus states with construction moratoria, and sovereignty demands (EU/Gulf/India insisting on nationally-controlled compute). The PUC tariffs — 14-year terms, $1.5m/MW collateral, minimum demand charges — are precisely society re-embedding the market to protect the substrate. This is the single strongest vindication of the anti-liberal reading: the substrate is being re-regulated almost as fast as it is commodified, which is exactly what Polanyi predicts a self-regulating market for a fictitious commodity cannot avoid.

Losurdo — the exclusion clause

Domenico Losurdo (Liberalism: A Counter-History, 2011) reads liberalism as always containing an “exclusion clause” — a community of the free defined against those outside it. The compute order’s formalized boundary is export controls. China is “effectively foreclosed” (Nvidia’s own 10-K language); “sovereign AI” status (UAE, Saudi) is admission to the club, granted by US licensing and conditioned on technology alignment — Microsoft’s $1.5bn into G42 was the price of entry. Losurdo’s point that inclusion/exclusion is drawn by power, not universal principle, is illustrated almost too neatly: the identical chip is legal in Abu Dhabi and illegal in Shenzhen by state fiat, not by market. This is a strong evidentiary fit.

Deneen / MacIntyre — liberalism producing its antithesis

Patrick Deneen (Why Liberalism Failed, 2018) argues liberalism fails because it succeeds — its logic of autonomy and market expansion dissolves the practices, institutions, and telos that make freedom meaningful, producing centralization and a managerial state. The neocloud is a candidate instance: a maximally “free market” (open entry, falling prices, private capital) has produced, at its frontier, extreme concentration (Nvidia’s chokehold, hyperscaler capex dominance, sovereign gatekeeping) and dependence on state allocation. Alasdair MacIntyre (After Virtue, 1981) adds the loss of practices oriented to internal goods: compute is organized wholly around external goods (return, share), with no telos beyond scale. The Nexuspolis axiom that “knowledge aims at the Good Life above personal gain” has no institutional home here — NAIRR and EuroHPC are the vestigial gestures toward it, and they are tiny. This reading is interpretive but well-supported by the concentration evidence.

Ostrom — is compute governable as a commons?

Elinor Ostrom (Governing the Commons, 1990) showed common-pool resources can be self-governed under eight design principles. Real attempts exist — NAIRR, EuroHPC AI Factories, DOE/NERSC ACCESS allocations, cooperative compute — and they satisfy Ostrom’s principles at small scale (they are genuine commons). But they fail the scale test decisively: EuroHPC’s flagship offers one winner 2.5% of EuroHPC capacity; NAIRR’s operations center is ~$35m over five years — against ~$785bn of hyperscaler capex in 2026 alone. This is an honest weakness for the commons-first Nexuspolis axiom: frontier training compute is not currently governable as a commons, because the capital intensity exceeds anything a commons has mustered. Inference-tier and research-tier compute plausibly are — and collapsing inference costs may widen that opening — but the frontier is not.

Mazzucato — socialized risk, privatized reward

Mariana Mazzucato (The Entrepreneurial State, 2013) argues the state bears foundational risk while private actors capture returns. The neocloud stack is a textbook case and the thesis’s second-strongest leg. Socialized/collective risk: ratepayer cross-subsidy ($9.33bn/yr in PJM alone); public grid and transmission buildout; the DOE $1bn TMI loan; export-control protection sheltering US/allied operators from Chinese competition; sovereign capital de-risking demand. Privatized reward: Nvidia’s margins, private-credit coupons, neocloud equity. The risk is concentrated at the substrate (ratepayers, bondholders, taxpayers, the grid); the reward is captured at the vendor and financier layer.

Ibn Khaldun — asabiyyah and the cycle (flagged as a stretch)

Ibn Khaldun’s Muqaddimah describes how asabiyyah (group feeling) drives a dynasty’s rise, and how luxury/complacency drive decline over ~three generations. The boom’s intense group-feeling — the shared conviction binding Nvidia, labs, neoclouds, sovereigns, and lenders into one project — resembles the cohesive ascendant phase; the “luxury” phase (ever-larger buildouts on ever-thinner marginal demand) resembles the onset of decline. I flag this explicitly as metaphor, not model: Ibn Khaldun’s cycle is generational, tied to political legitimacy and desert-to-city sociology, not a five-year hardware amortization schedule. It is a suggestive frame for the sentiment cycle (the asabiyyah of a bubble), not evidence.

”Do everything with nothing”

The reader’s thesis — that agentic software makes the operating labor of the compute factory vanish — is half-true, and the other half is decisive. Weightlessness at the abstraction layer (agents, autonomous ops, software-defined everything) is real. But it is purchased by extreme heaviness at the substrate: gigawatts of power, copper and concrete, HBM in 36–52-week backlog, cooling water, 14-year grid contracts. The “nothing” of the software layer sits atop the “everything” of the physical layer — and the ownership question lives entirely at the substrate. Whoever owns the power, the interconnection-queue position, and the silicon owns the industry, which is exactly why the converted bitcoin miners (energized land) and Nvidia (chips) capture the structural rents while the “weightless” neocloud operator “owns almost nothing.” Do-everything-with-nothing is true for the operator and false for the polity: the labor vanishes, the mass does not, and the mass is where power resides.

The counterargument, stated at strength

The good-faith liberal rebuttal is serious. (1) Entry is genuinely open — 300+ new GPU providers in 2025; not a closed cartel. (2) Prices are falling — H100 rates down 64–75%; competition delivering consumer surplus, not monopoly rent. (3) Hyperscaler dominance is being eroded — neoclouds grew from nothing to ~$48bn run-rate and won contracts from the incumbents’ own customers (Microsoft’s decision not to exercise a ~$12bn CoreWeave option in favor of Nebius; Meta building its own capacity). (4) Circularity is ordinary vendor finance — one-directional core revenue, scrutinized public companies, a pattern that built railroads and telecom. (5) Export controls are national security, not market-rigging — and they reduce US firms’ addressable market rather than protect them (Nvidia’s $4.5bn China charge). (6) The state’s role in power/grid is the normal regulation of a utility, long predating AI.

Adjudication — a split by layer

The honest verdict is a hybrid, split by layer. At the abstraction layer (renting GPU-hours, software, the spot market), the liberal description is largely accurate: open entry, falling prices, real competition, ordinary vendor finance. At the substrate layer (chips, power, land, sovereign capital), the anti-liberal description is largely accurate: allocation is by state fiat (export controls), the key input is a single-vendor chokepoint that also backstops demand, capital is substantially sovereign, and risk is socialized. The two coexist because the liberal competition on top is contained within a non-liberal allocation of the bottom. The neocloud order is therefore neither a clean refutation nor a vindication of liberalism, but specific evidence for the Deneen/Burnham hypothesis: a liberal-competitive surface that has produced, and now depends on, a non-liberal, managed, state-and-vendor-cartelized substrate. The liberal vocabulary persists because it is true of the layer everyone interacts with — and misleading about the layer where power actually sits. That locates precisely where the thesis holds (the substrate) and where it fails (the surface), which is the finding most faithful to both the evidence and the Nexuspolis project’s own commitments.

Recommendations / Falsification Thresholds

On the bubble question, watch:

Confirms bear: SemiAnalysis contracted H100/H200 rental index turns sustainably negative YoY; a tier-1 neocloud misses debt service or a major offtake is renegotiated down; CoreWeave RPO conversion falls below ~35%; a GPU-backed ABS/loan tranche is downgraded below investment grade; hyperscaler capex guidance cut >20% in a quarter; the 2026–28 refinancing wall reprices materially wider.
Confirms bull: Contracted-capacity utilization stays >90% through the Blackwell→Rubin transition; lab inference revenue (not just training) scales demonstrably; neocloud gross margins hold >25%; the refinancing wall clears at flat-or-tighter spreads.

On the anti-liberal thesis, watch:

Strengthens thesis: Export-control tiers harden into permanent allocation; sovereign funds’ share of frontier-compute capital rises; ratepayer cross-subsidy grows despite tariffs; Nvidia’s equity-plus-backstop model spreads to more operators.
Weakens thesis: A credible non-Nvidia stack (AMD, Google TPU, a Chinese frontier chip) breaks the single-vendor chokepoint; large-load tariffs succeed in shifting ~100% of incremental cost onto data centers (de-socializing risk); a commons-scale public compute program reaches frontier scale.

Concrete next steps for a researcher/builder: Operate where you have leverage. For a builder, the abstraction layer is genuinely competitive and cheap — build there and assume compute prices are subsidized today. For a political-economic project (Nexuspolis), the substrate layer — power, interconnection priority, chips — is where ownership and therefore politics reside; any intervention aimed at the compute order must target the substrate (public power, public interconnection priority, commons-scale compute), because abstraction-layer interventions will be washed out by falling prices.

Caveats

Recency limit: Coverage runs through September 12, 2026. Fast-moving items (the Nvidia–OpenAI deal structure, PJM auctions, export-control rules) changed multiple times within 2025–26 and may change again.
Vendor/self-reported vs verified: Backlog, ARR, “sold-out,” and category-size figures (MeasuredAI’s $48bn/$300bn; CoreWeave’s $104bn backlog) are company- or analyst-reported and not independently audited. SEC-filed and regulator-filed figures (revenue, net loss, concentration %, debt, the $6.3bn backstop, the $8.5bn facility ratings, PJM auction clearing prices, Nvidia’s H20 charge, LBNL/IEA energy data, BIS/Moody’s/IMF assessments) are higher-confidence.
Conflicting/single-sourced claims flagged in text: The AI private-credit stock figure (“over $200bn,” Quinn Emanuel) conflicts with the BIS flow figure (“over $40bn originated in 2025 vs ~$3bn in 2010”) — these measure different things and should not be summed. The $176bn depreciation figure is Burry’s own estimate. Exact Nvidia equity percentages in private neoclouds (Nscale, Crusoe, Lambda) are undisclosed.
Inference vs speculation: Part One is evidentiary. Part Two is explicitly interpretive — the Burnham, Polanyi, Losurdo, and Mazzucato readings are well-grounded in the Part One evidence; the Ibn Khaldun reading is metaphor, not model; the split-by-layer adjudication is my reasoned judgment, not settled fact.

Strongest objections to this report’s own conclusions

Layers may be converging, not fixed. If abstraction-layer competition keeps eroding hyperscaler margins, the substrate chokepoints may loosen too (cheaper power tech, alternative chips), making the order more liberal over time — in which case the split-by-layer verdict is a snapshot, not a structure.
The Nvidia backstop may be pro-competitive, not cartel behavior. A chip vendor rationally de-risking its ecosystem lowers entry barriers for small operators (they borrow at IG-like pricing because of it), which is the opposite of cartelization.
The Polanyian “double movement” may be misidentified. Utility regulation long predates AI; the PUC tariffs may be ordinary cost-allocation, not a novel protective countermovement against a newly-commodified substrate.
Commons pessimism may be premature. Collapsing inference costs could make commons-scale inference achievable even if training remains out of reach — partly rescuing the Ostrom axiom.
US-centric distortion. The evidence leans heavily on US/PJM data; the picture differs materially in the EU (public AI Factories), the Gulf (pure sovereign capital), and China (state-directed, export-constrained), so generalizing a single global “neocloud order” risks overreach.

Sources

Morgan Stanley, “Bridging a $1.5tr Data Center Financing Gap” (Jul 2025).
Bank for International Settlements, BIS Bulletin No. 120 (AI-related private credit); IMF Global Financial Stability Report (data-center leverage and concentration).
CoreWeave: Q2 2026 Form 10-Q, FY2025 disclosures, and press releases for DDTL 4.0 / 5.0 / 5.5 and the €2bn 2032 notes.
Reuters (CoreWeave $2.3bn H100-collateralized facility, Aug 2023); Moody’s / DBRS / Fitch ratings on DDTL tranches.
MeasuredAI neocloud taxonomy and category revenue estimates (mid-2026); TrendingTopics (“neocloud” coinage); Introl (300+ new GPU providers, 2025).
Quinn Emanuel (AI-related private-credit stock estimates); GlobalDataCenterHub (DDTL 4.0 / Meta–Blue Owl “Hyperion” structure).
PJM capacity auction results (2025/26 capacity costs; data-center load share).
Nvidia Q3 FY2026 filing (Nov 19, 2025) on gross margins; CoreWeave ownership stake and capacity backstop disclosures.
Theory: James Burnham, The Managerial Revolution; Karl Polanyi, The Great Transformation; Domenico Losurdo, Liberalism: A Counter-History; Patrick Deneen, Why Liberalism Failed; Alasdair MacIntyre, After Virtue; Elinor Ostrom, Governing the Commons; Mariana Mazzucato, The Entrepreneurial State; Ibn Khaldun, Muqaddimah.

Part of the Univrs research ecosystem:

Research: metalearn.org · ardeshir.io · sepahsalar.org · univrs.io
Creative: imagine.univrs.io
Code: github.com/univrs · github.com/ardeshir