AI Spend Is Now Paying Off Inside Portfolio Companies

What's New

Blackstone argues the return on AI spending has moved from forecast to evidence inside its own portfolio. Jon Gray made the case at the firm's September 2026 annual meeting. Blackstone tracks AI spend across 1,400 portfolio companies, GP stakes, and borrowers. Their spend on Anthropic reached $525 million annualized. That is a 21-fold increase from September 2025. Gray paired the spend with company-level results, including a 5x speed gain in lease processing at Phoenix Tower. Allocators can now ask managers for measured AI returns at the company level.

Why It Matters

The finding challenges the view that AI capex is a bubble with no downstream payoff. Private equity software owners are on the exposed side. Fear of AI obsolescence compressed software multiples, often for businesses Gray says were fine. PE software deal activity fell 66%. Gray's retail parallel sets the filter. Systems of record that move from per-seat to outcome-based pricing survive, as Walmart and Costco did. The rest face the Sears outcome.

By The Numbers

• $820 billion in capex from the top five hyperscalers, about double the prior year.

• 700 basis points of EBITDA margin expansion at Blackstone portfolio companies over four years. The S&P gained 500.

• 2.6% annualized US labor productivity growth over the past 2.5 years. The 10-year baseline was 1.5%.

• $55 billion to build a 1 GW AI campus, including power, data centers, and chips.

Between The Lines

Gray speaks as a large owner of the infrastructure his argument supports. Blackstone's data center leasing is projected to reach at least 6 GW in 2026. AI-related investments made up 9 of the firm's 10 top-appreciating assets in Q2. The ROI evidence is a curated set of case studies from its own companies.

Gray also hedges unevenly. He says it's hard to know how much of the productivity gain is AI-driven. He then says adoption is "of course" showing up in margins, with no such caveat. The margin gain spans four years. The Anthropic spend jump spans one. The most useful claim for allocators is the supply bottleneck. GE Vernova turbine lead times run to 2030 to 2031, which favors owners of power and grid assets.

The Wrap

A barbell follows from Gray's case: physical AI infrastructure on one side, operators using AI to widen margins on the other. It holds while demand outruns capacity and contracted power and data center capacity stays with low-leverage tenants. It fails if a major cyberattack on financial or power grids triggers heavy regulation. It also fails if chip supply from Taiwan is disrupted. The test runs through 2030 to 2031, the current lead time for heavy-duty turbines.