$7.5B Ellison Exit Plus a $2.8B Restructuring Tab — Oracle's AI Cloud Beat Can't Outrun Its Capex

Within the same news cycle in mid-September 2026, two Oracle headlines landed that point to the same conclusion. The Financial Times reported that Larry Ellison adopted a Rule 10b5-1 trading plan in June to sell up to 50 million shares — about $7.5 billion at the then-prevailing price — through October 24, 2026. A day later, Oracle confirmed that it had added another $700 million to the cost estimate of its 2026 Restructuring Plan, pushing the plan ceiling to $2.8 billion. Neither disclosure came with a fundamental trigger in the headline sense; together they reveal how uncomfortable the AI-capex math already is.

Ellison's June 2026 plan authorizes the sale of up to 50 million Oracle shares — roughly 4.3% of his reported 1.16 billion share, 40.6% stake — by October 24, 2026. At the closing price of $150.28 on September 11, the cap on the plan represents about $7.5 billion of proceeds, and Ellison has now executed sales large enough to push the plan into public view. A 10b5-1 plan is technically pre-scheduled, but the size is unusually heavy for a founder who has used such plans sparingly and is signing it at the moment Oracle's shares sit about 56% below the September 2025 peak of $345.72.

Oracle's \"2026 Restructuring Plan\" was originally announced at up to $2.1 billion in cost. In March 2026 the Wall Street Journal reported the company was adding $500 million to cover severance tied to AI-driven head-count actions. On September 11, Seeking Alpha reported a further $700 million increase, taking the plan ceiling to $2.8 billion. The Q1 FY2027 10-Q still classifies the period's $94 million restructuring expense primarily under this same plan.

FY2026 was the year the AI revenue beat and the FCF line broke at the same time. Oracle generated $67.4 billion in revenue (+17%), $32.0 billion in operating cash flow, and $55.7 billion in capital expenditures (+162% year over year). Free cash flow printed at negative $23.7 billion — a swing of roughly $23 billion from the negative $0.4 billion of FY2025. Q1 FY2027, reported September 10, 2026, did not repair the gap: capex of $28.5 billion (versus $8.5 billion in Q1 FY2026) and free cash flow of negative $5 billion, even as cloud-infrastructure revenue doubled.

Oracle, and the10-K discloses an additional $260 billion of off-balance-sheet data-center lease commitments plus $13.3 billion of unconditional power-purchase obligations.

Underneath the capex, Oracle and the decision to keep funding $50 billion of annual capex. Strip Stargate out and the multiple compresses immediately.

That makes OpenAI's execution the single most important non-Oracle variable. A pullback in Stargate, a rephasing of the five-year ramp, or a credit event at any of the OpenAI financing partners (SoftBank, MGX, the U.S. government-backed cohort) would invalidate the RPO trajectory and force a fresh look at whether $55–$60 billion of annual capex is the right run-rate.

Every dollar of Oracle GPUs fabricated at TSMC (TSM ADR), deployed inside Oracle] data centers that draw power from U.S. utility grids, sold to AI customers like OpenAI, xAI, and Microsoft workloads that increasingly run on Amazon Web Services alternatives. The question is who carries the risk when [Oracle](orcl]'s FCF stays negative and Stargate wobbles.

Days-to-quarters: the next catalyst is Ellison's October 24 plan expiry, with quarterly Form 4 filings in the interim showing how aggressively he is executing. Watch Q2 FY2027 (reported December 2026) for capex run-rate and any change to the $45–$50B calendar-2026 raise pacing. The 8-K filed September 10, 2026 also included the Series D preferred-stock mechanics that will convert into common by January 2029, layering additional dilution on top of the ATM program.

One-to-three years: the structural question is whether AI cloud revenue compounds at the FY2027 guide (≥$90B) while capex normalizes back to the FY2024–FY2025 range of roughly $20–$25B. If capex stays at $55–$60B and FCF stays negative for two more years, [Oracle](orcl]'s net debt approaches $200B and the equity story shifts from AI growth to AI balance-sheet risk. If OpenAI's Stargate scales on schedule and lease commitments convert into multi-year recurring revenue, FCF inflects in FY2028.