The Keystone

Executive briefing

An entire capital cycle has been underwritten against a single IPO that has not yet priced. Hyperscaler capex aggregates above three hundred billion dollars in 2025 alone; neocloud project finance, vendor reciprocal commitments, and the institutional credit pool sitting underneath all of it have been written against an OpenAI valuation that holds at or above the November 2025 round of $500bn. The technology works. The cash flows are coming. The dispute is over which entities will be there to collect them, and at what price the assets will change hands when the price-discovery event the cycle has been priced against finally occurs.

The visible artefact is Amazon’s Q4 2024 accelerated-depreciation charge, roughly $920m booked against a portion of its accelerated-compute fleet. The same accounting team that led the 2023 cohort-wide extension of server useful life from five years to six, citing improved telemetry and the longer working life of GPU-class hardware, has now reset on the same fleet running the same workload. Microsoft, Alphabet and Meta have not followed yet. Their forward capex plans still embed the longer convention, and the depreciation charge they would book if they did is a meaningful share of the operating income that justifies the share price.

The argument is structural rather than tactical. Cohort capex of three hundred-plus billion dollars on a single asset class with a useful life the bellwether has just reset is the issue, not any individual disclosure. The cleanest counter-argument — that depreciated training fleets retain commercial value as inference capacity, in the manner of Meta’s “warm shell” repurposing — is a partial offset, not a complete one. It extends effective economic life on a portion of the fleet, but the portion is bounded by inference demand and constrained by the physical fact that yesterday’s training silicon is not always today’s inference silicon at competitive cost.

One stack-level down, the duration mismatch is sharper. CoreWeave’s most recent senior unsecured tranche priced at 9.75 per cent. Its credit agreements include “Cash Trap Event” triggers tied to anchor-tenant contract performance. Its 10-Q discloses payment terms on a major customer extending up to three hundred and sixty days — the structure that converts a customer relationship into working-capital financing, with the neocloud bearing the float. The single anchor tenant in question is OpenAI, and the equity capital that lets OpenAI prepay against multi-year compute commitments is sourced from a circle of counterparties that includes the same compute vendors. When the same dollar appears as revenue on one income statement and as a capital commitment on the same company’s cash-flow statement, between the same counterparties, in the same reporting period, the line-by-line accounting and the consolidated economic substance diverge.

NVIDIA holds equity in CoreWeave. SoftBank closed a $40bn commitment to OpenAI in March 2026, supported by a bridge loan against Vision Fund LP commitments and listed-stock collateral; S&P put SoftBank on negative outlook the same month. Underneath all of it sits the institutional capital layer — BCRED and the four other major listed alternative managers, university endowments with private-equity allocations between a quarter and two-fifths of total assets, NAV-financing vehicles deferring mark-to-market on exposures whose economic reality lags book value. February 2026’s correlated AI-software credit selloff across Blackstone, Apollo, KKR, Ares and Blue Owl is the data point that suggests the published correlations underdescribe the actual exposures.

Two geographies are not competing for these assets. The Gulf programme — HUMAIN in Saudi Arabia, Stargate UAE — has been re-priced by a war. On 3 April 2026 the IRGC published a video designating the Stargate UAE campus as a target conditional on continued American action against Iranian energy infrastructure. The marginal Patriot battery is not available for the marginal data-centre at any sovereign-wealth price. Europe narrows to three viable geographies: France, the Nordics, Iberia. Combined realistic additions reach twelve to sixteen gigawatts by 2030, against an announced European pipeline several times that. The IEA’s Electricity 2026 base case and GE Vernova’s gas-turbine backlog imply a thirty-to-fifty gigawatt global supply gap by 2030.

What is up for collection, when the unwinding occurs, is every asset in the stack whose economic value exceeds the leverage wrapped around it. Apple holds approximately $160bn in cash. Alphabet holds approximately $100bn. Both have sat out the AI capex cycle. Both have a documented history of acquiring infrastructure at the bottom of capex cycles. Four indicators resolve the thesis on a quarterly cadence: whether OpenAI raises equity at or above the November 2025 valuation; whether any anchor-tenant contract is renegotiated or any Cash Trap Event triggered; whether a second hyperscaler follows Amazon on useful-life; whether gas-turbine, transformer and interconnection-queue delivery actually closes the IEA gap. None of these has yet resolved. All of them are reported.