Microsoft ↔ OpenAI
Generated 2026-06-06. Compiled from primary sources (Microsoft/OpenAI blogs, SEC 8-K ex99 accession 000119312525256310, accounting analyses) + flagged knowledge fills.
The arrangement
- Cumulative investment ~$13B (from ~$1B in 2019 through 2023). A large portion delivered as Azure credits, not cash — so OpenAI "spends" Microsoft's investment back into Microsoft's own cloud. [exact cash/credit split never itemized — med confidence]
- Oct 28 2025 restructuring (definitive): Microsoft holds ~27% of the new OpenAI Group PBC, reported value ~$135B. IP rights through ~2032; reciprocal revenue-share (reported cap ~$38B); Azure right-of-first-refusal removed (OpenAI free to use Oracle/CoreWeave/Google).
- OpenAI's incremental $250B Azure commitment — a quarter-trillion-dollar purchase pledge to the cloud of its own 27% owner.
- Apr 27 2026 "next phase" deal supersedes the Oct 2025 terms; the equity + Azure-commitment architecture persists.
The round-trips (why this is circular)
- Credits round-trip: Microsoft funds OpenAI partly in Azure credits → OpenAI consumes Azure → Microsoft books it as Azure revenue. Microsoft both funds and is paid by the same dollar.
- Commitment round-trip: OpenAI raises cash (SoftBank, Nvidia, sovereigns) → pledges $250B to Azure → Azure buys Nvidia GPUs (Nvidia also funds OpenAI). The dollar circles supplier→customer→supplier.
- Equity-method drag (the honest counterweight): OpenAI's net loss flows onto Microsoft's income statement. Q1 FY26 hit: $3.1B / $0.41 EPS, up 6× from $523M / $0.07 a year earlier. OpenAI's burn is now a visible, accelerating cost to Microsoft GAAP earnings — the circularity cuts both ways.
Scale anchors
| Item | Value | Period |
|---|---|---|
| MSFT stake in OpenAI | ~27% (~$135B) | Oct 2025 |
| OpenAI → Azure commitment (incremental) | $250B | Oct 2025 |
| MSFT FY25 capex (P&E, excl. leases) | $64.55B (FY24 $44.48B) | FY2025 |
| MSFT commercial RPO | $631B | Q2 FY26 |
| OpenAI equity-method loss to MSFT | $3.1B/qtr | Q1 FY26 |
| OpenAI revenue run-rate vs commitments | ~$13–20B vs ~$1.4T | 2025–26 |
The last row is the crux: OpenAI has pledged on the order of $1.4 trillion in compute (Azure $250B + Oracle ~$300B + CoreWeave ~$22B + Broadcom/AMD/Nvidia/SoftBank) against a revenue run-rate ~70× smaller. Every node in the web has booked a slice of that $1.4T as backlog/RPO/expected revenue.
The asset behind the backlog ages out (added 2026-06-11)
That $1.4T of "expected revenue" is also $1.4T of compute someone must build with debt- and lease-financed GPUs — and those GPUs economically age in ~2–3 years while the financing runs 5–19 (Microsoft itself depreciates GPUs over ~6yr; on honest lives that overstates profit by ~$2.9B/yr on one tranche). So even the "real assets" side of this web carries the duration-mismatch defect (fin-ai-depreciation-debttrap, depreciation_trap D1–D4): the backlog is promised against compute that depreciates faster than the loans that bought it amortize. The equity-method drag above (OpenAI's loss flowing onto MSFT GAAP at $3.1B/qtr) is the visible cost; the deferred one is the depreciation that a long assumed useful life is borrowing from a future writedown.
← Research index · structured data: fin-microsoft-openai.json · fin-microsoft-openai.md