Blackstone's Jon Gray Says AI Has "Beef": Nine of the Firm's Top 10 Q2 Gains Were AI-Linked, but Chips, Power and Capital Are Capping Supply
President Gray's annual investor presentation, September 17, 2026
The firm's own portfolio is the evidence
Blackstone President Jon Gray anchored his case for the AI buildout in data from inside the firm. Across roughly 1,400 portfolio companies, GP stakes holdings and borrowers, run-rate spend with Anthropic rose 21-fold in a year, from $25 million in September 2025 to $525 million. Gray said these companies are "getting extraordinary returns on the investment," which is why spend keeps growing. He also said that nine of the ten investments that appreciated most across the firm in the second quarter were AI-related, including holdings in the model companies. The one exception was Aster Care hospitals in India. He expects the third quarter to look "very similar."
Gray offered portfolio examples of the returns. Phoenix Tower invested $4 million in an AI lease-processing system that is now generating $4.5 million a year, an annual return above 100%, and it processes leases five times faster. Enverus is earning an 18x return on what it spends with model companies on code remediation. Tricon can process rental applications 90% faster. Chamberlain, the garage door opener maker, has used AI visualization to build a smart-lock business now at a $40 million to $50 million run rate, which its CEO believes can reach $500 million in five years. Gray said the CEO told him this a few weeks ago.
Early signs in the macro data
Gray argued the gains are starting to show up in aggregate numbers, with a caveat on attribution. US productivity growth, which averaged 1.5% for a decade, has run at 2.6% over the past two and a half years. "It's hard to say how much is AI-driven, but something is happening out there," he said. Hyperscaler revenue per employee is up 65% over three and a half years. EBITDA margins for the S&P 500 and Blackstone's portfolio companies are up 500 and 700 basis points respectively over four years. Earnings growth has accelerated from a 15% average to 32% over the last 12 months, even excluding one-time gains. Gray said he does not know whether that pace will last.
He framed current usage as early. The technology is "mostly just at our desk," and demand for compute should rise as agents move into robotics, autonomous vehicles, national security and space. As an example he cited Waymo, whose mileage is up 250-fold in two and a half years, with serious accident rates 94% lower.
Supply is the binding constraint
The most consequential point for capital allocators is that Gray believes the physical world, not demand, is the limit. Hyperscaler capex at five companies doubled to $820 billion this year from $415 billion, or about 2.5% of US GDP. Blackstone's data center leasing went from 1 gigawatt in 2024 to 2 gigawatts in 2025, and Gray expects at least 6 gigawatts this year. That equals nearly $100 billion of capex, with tenants adding another couple hundred billion dollars of chips.
Even so, he described shortages at every layer. Entitlements are slowing builds as communities push back, and moratoriums have appeared in New York State. Gray called much of the opposition misinformation, arguing data centers can be built with virtually no water use and can lower electricity costs with new power. A turbine order from GE Vernova now means "get in line, 2030, 2031." Hyperscalers have raised capex nine-fold in five years, while chip makers have not even doubled theirs. Gray gave them some credit for caution in a cyclical business, but said the result is a "significant shortage of chips." Capital is the fourth constraint, at roughly $55 billion per gigawatt including power, buildings and chips. Gray said that scale is why Blackstone is "uniquely positioned," and also why supply is not keeping up. He added that the cost of this buildout "is actually, I think, going to constrain the usage of AI over time."
Bubble question: leverage and counterparties differ from past booms
Gray addressed the bubble concern directly, using the 1984 "Where's the beef?" Wendy's ad as his frame. He conceded that a 600% rise in Micron and a 500% rise in SK Hynix looks bubble-like. He noted, though, that SK Hynix trades at a P/E of 4, versus Cisco's 150 times earnings in 2000, and said this reflects lingering skepticism and a fear that memory makers are over-earning.
He drew a distinction with the 19th-century railroads, where roughly 200 companies went bankrupt or insolvent, and with telecom. Those cycles involved highly leveraged supply built ahead of demand. Today, he said, "the demand is running well ahead of supply," and much of the contracted energy and data center capacity is with lowly leveraged counterparties. He did not rule out failures. His historical comparison put the present closer to 1870 than 1900. Over those three decades, US labor productivity doubled, GDP rose fourfold and equities rose sevenfold, though he stressed he is not forecasting a repeat.
Gray also acknowledged that valuations in parts of the market worry him. He pointed to companies with "no revenue with $10 billion values" and to some defense technology names at very high prices. This is partly why Blackstone is concentrated in "that seniormost part of the stack, the compute, where we have the most confidence."
Model company valuations and the firm's deployment
Gray said Blackstone has invested in both Anthropic and OpenAI. Their combined value has risen from $683 billion last September to an estimated $2 trillion to $3 trillion as they move toward going public. Token volumes at Google grew from near zero in 2024 to 480 trillion a month, then to 3.2 quadrillion by May 2026.
Blackstone has deployed accordingly. It has financed GPUs and TPUs, committed capital alongside Nvidia and Broadcom, built a joint venture with Google around Crux, and invested in neoclouds and data center operators including Firmus in Australia and Neysa in India. On power, it has put capital into Sempra Infrastructure, Enverus, Williams, McLean electrical equipment, and European renewables developers Eurowind and Sunotec. Gray said the firm did not invest in the memory makers, which were among the biggest winners.
Disruption risk is real, and software is the clearest casualty
Gray said underestimating technological disruption is "the thing that keeps us up at night the most," and noted it has begun to play out. Professional services, software and information services stocks have de-rated even where underlying businesses were fine, and private equity software deal volume fell 66%. He expects outcomes to diverge, as they did in retail after Amazon, where Kmart, Sears and Toys R Us failed and Walmart, Costco and TJ Maxx thrived. The white-collar survivors, he said, will hold vital systems of record and shift their pricing models "from seats to outcomes." Even so, he expects these companies to carry lower multiples because of the "uncertainty quotient."
He pointed to scarce, non-technology assets as relative winners. Examples include an Indian cricket franchise, the 7Brew coffee chain, beachfront property and the Rome airport.
Risks Gray flagged
The risks he listed were cyber safety, with a major breach at a financial institution or critical infrastructure likely to trigger a political response. He argued that regulation slowing the frontier would not stop diffusion, because current models are already powerful. Others were alternative architectures such as data centers in space and edge computing, which he sees as additive, with terrestrial data centers staying important for now. He also cited geopolitical tension, with 90% of advanced semiconductors made in Taiwan, and excessive valuations.
On the labor side, Gray said overall employment is growing at Blackstone portfolio companies despite AI productivity gains. The number of construction workers at QTS sites has tripled in under two years, and new business applications have doubled in a decade. He also cited AI-assisted drug discovery, with improved Phase 1 success rates, and a McKinsey study showing one trial's duration cut by 40%. He said AI is stronger in chemistry than in biology.