The Keystone — Sources

External companion to the long-read. Citation chain for every load-bearing claim, organised by section.

Verbatim quotes

Introduction

• Nadella’s “warm shells” remark and Microsoft’s ~$80bn fiscal 2026 AI capital budget. Bg2 Pod, November 2025.
• Close to half of planned 2026 US data-centre builds delayed or cancelled for lack of electrical backbone; transformer lead times stretched from roughly two and a half years (2022) to past four; US imports of high-power transformers from China up roughly fivefold in three years. Bloomberg, April 2026.
• H100 pricing: roughly $40,000 at the 2023–24 allocation peak, down to roughly $14,000 on the secondary market in early 2026; GB200 delivering roughly fifteen times the inference performance per watt.
• Amazon’s accelerated-depreciation charge (~$920m, six-year to five-year useful life) versus Meta’s extended useful life (three years to five and a half, ~$3bn of additional operating income), both disclosed in Q1 2025 filings.

Section 1 — The inversion, and the accounting

• NVIDIA’s fiscal 2026 year-end inventory of $21.4bn, more than double the prior year and the largest in the company’s history. NVIDIA fiscal 2026 filings.
• Secondary-market H100 clearing rates down roughly 60% from their mid-2024 peak; refurbished-certified units from $42,000–$45,000 to $14,000–$16,000.
• NVIDIA’s Vera Rubin platform, scheduled for the second half of 2026, drop-in compatible with existing Blackwell racks, delivering approximately five times the inference performance; industry commentary that “Vera Rubin obsoletes current AI iron six months ahead of launch,” January 2026.
• Transformer lead times of eighteen to thirty-six months from qualified manufacturers; US grid interconnection queues of two to four years; new combined-cycle gas turbines ordered in early 2026 not delivering electricity before 2029.
• Amazon’s 2023 extension of server useful life from five to six years (~$3.1bn added to 2024 operating income), reversed in February 2025 back to five years for its AI-relevant fleet, cited explicitly in subsequent quarterly filings.
• Estimate of unrecognised depreciation exposure: NVIDIA Data Center revenue of roughly $315bn across 2024–2025, hyperscalers representing “slightly over 50%” per NVIDIA disclosure, implying a $250–300bn hardware base at the system level and $15–25bn of additional annual depreciation from a one-year reduction in assumed useful life.

Section 2 — The arithmetic of a stranded gigawatt

• Illustrative 100-megawatt cluster model: ~$3.7bn all-in capital cost (roughly 60% silicon, 40% power and shell), five-year take-or-pay contract, 10% discount rate. On-schedule day-one NPV of roughly +$0.25bn; a six-month delay taking day-one NPV to roughly −$1.70bn; a twelve-month delay to roughly −$2.52bn.
• Refurbished-certified H100 units at the three-year mark trading at 30–40% of peak retail; raw uncertified units at 20–25% of peak.

Section 3 — The supply-side arithmetic

• Combined gas-turbine manufacturing capacity of GE Vernova, Siemens Energy and Mitsubishi Power reaching roughly 50–60 gigawatts per year globally by late decade, across all gas-fired generation, not only data centres. GE Vernova’s late-2025 earnings call confirmed heavy-duty turbine slots sold out through 2029; NextEra has cited 2032 for newly ordered equipment.
• Global commercial aircraft backlog of roughly 14,000 firm orders between Boeing and Airbus, competing for the same turbine-blade foundry capacity as data-centre gas generation; F-35 engine hot-section shortages grounding combat-ready US aircraft through much of 2025.
• Iranian strikes on Qatar’s Ras Laffan LNG export terminal, 1–2 March and 18–19 March 2026: the first prompting QatarEnergy force majeure on its entire output and a 50% intraday spike in European wholesale gas; the second structurally damaging two production trains and roughly 17% of Qatar’s export capacity (3–4% of global LNG supply), with a three-to-five-year repair timeline.
• IEA base-case forecast implying a 30–50 gigawatt structural gap by 2030 between available power supply and the announced AI data-centre pipeline — comparable to Germany’s (~56GW) or the UK’s (~33GW) average national electricity load.

Section 4 — The financing layer

• CoreWeave’s adjusted debt rising from roughly $11bn (end of 2024) to roughly $30bn (April 2026), including $5.75bn raised in two weeks of April 2026 at 9.75% interest. NVIDIA’s $2bn direct equity investment in January 2026 roughly doubling its stake in exchange for preferential next-generation GPU access.
• CoreWeave’s 2025 free cash flow deficit of $7.2bn, funded with debt; year-end current ratio of 0.46; December 2025 amendment loosening minimum-liquidity and debt-service-coverage covenants five months after the original terms were agreed. S&P’s B+ rating affirmed with a positive outlook contingent on remediating internal-control weaknesses by end-2026.
• CoreWeave quarterly filings disclosing invoice payment terms of up to 360 days for a significant customer, and a “Cash Trap Event” triggered by three consecutive months of non-payment by a large customer.
• Oracle’s $300bn five-year cloud commitment from OpenAI representing roughly 57% of Oracle’s $553bn backlog; roughly 90% of Oracle’s 10 gigawatts of committed capacity “funded through partners” per Oracle’s own quarterly disclosure, with equipment either prepaid by the customer or supplied directly by the customer.

Section 5 — Money in a circle, and the keystone

• The reciprocal-commitment loop: NVIDIA’s pledge of up to $100bn to OpenAI against hardware deployment; OpenAI’s $300bn cloud agreement with Oracle; Oracle’s roughly $40bn commitment to NVIDIA chips; AMD’s warrant grant to OpenAI for 10% of AMD at a penny a share, contingent on 6 gigawatts of deployed capacity; Microsoft’s $13bn-plus stake in OpenAI against OpenAI’s $250bn cloud-spend commitment to Microsoft; NVIDIA’s 7% stake in CoreWeave and $6.3bn capacity purchase commitment.
• OpenAI revenue trajectory: from effectively zero to $13bn in recognised revenue in under four years; 2026 operating losses projected at $14bn against $13bn of recognised revenue.
• SoftBank’s funding of its $40bn OpenAI commitment: liquidating its entire NVIDIA position, selling $4.8bn of T-Mobile US holdings, drawing $11.5bn of margin loans against Arm, and arranging a $40bn twelve-month unsecured bridge loan from JPMorgan, Goldman Sachs and three Japanese banks. S&P moved SoftBank’s outlook to negative in early March 2026 citing OpenAI concentration.
• Aramco’s December 2019 IPO as the scale precedent: originally targeting a 5% float at a $2tn valuation to raise $100bn, ultimately a domestic-only 1.5% float at a $1.7tn valuation raising $29.4bn, backed by subsidised Saudi retail participation.

Section 6 — The capital backdrop

• PIMCO’s reduction of carrying values on a private-credit book in late 2025, and the February 2026 software-credit selloff that prompted the five largest alternative managers to publicly defend their underwriting.
• Stargate UAE: 1-gigawatt first phase of a planned 5-gigawatt campus. American-Israeli strikes on Iran, 28 February 2026; Iranian drone retaliation against AWS facilities in the UAE and Bahrain; the Islamic Revolutionary Guard Corps’ 3 April video naming the Stargate UAE campus as a conditional target.
• European capacity estimate of roughly 12–16 gigawatts realistically addable by 2030 across France, the Nordics and Iberia, against a European share of the announced AI pipeline that is several multiples of that figure.

Section 7 — The asymmetric bet, and what to watch

• Amazon’s roughly $90bn of cash equivalents; Alphabet’s roughly $100bn of cash and its historical pattern of acquiring infrastructure at the bottom of capex cycles (YouTube, DoubleClick, Motorola); Apple’s roughly $160bn of cash, short-term investments and marketable securities, and its absence from the AI capex cycle to date.
• Four falsifiable indicators: whether OpenAI sustains equity-raising pace to prepay Oracle, Microsoft and AWS commitments on schedule and prices at or above its November 2025 valuation; whether any anchor-tenant contract collateralising a neocloud loan is renegotiated or triggers a Cash Trap Event; whether a second hyperscaler follows Amazon in shortening server useful lives; whether gas-turbine production, transformer manufacturing and interconnection-queue clearance deliver against the announced pipeline.