Where's the Beef? Blackstone's Jon Gray on the Trillion-Dollar AI CapEx and the Payoff

When the room filled for the first time since September 17, 2025, Jon Gray opened with self-mockery about how to become a finance fitness influencer, then laid out four questions to guide the morning: what changed in the past year, what the distant past teaches, where we stand now, and where we are headed. He narrowed the lens to two letters on purpose — AI — not because war in Ukraine, Middle East commodity pressure, sticky inflation and fiscal deficits do not matter, but because intelligence at scale would shape portfolio value and the economy most. That choice explained why every chart that followed was read through an AI filter. The consistency was underscored by keeping the same dress code as a year earlier, a small signal of a steady thesis.

What Twelve Months Changed: From Tokens to Revenues

Scale came first through tokens : Google's monthly token flow was 480 trillion a year earlier and pressed to 3.2 quadrillion by May 2026, higher still by September from a near-zero base in 2024. On the model side, OpenAI and Anthropic grew roughly a hundredfold by July 2025 from almost nothing and were running at a 105 billion dollar pace by July 2026, a ramp rarely seen in economic history. Blackstone's own window — about 1,400 companies across portfolio, GP stakes and borrowers — showed annualized spend on Anthropic rising twenty-one times from 25 million dollars in September 2025 to 525 million dollars, because users were seeing strong returns. Valuations followed: the pair worth 683 billion dollars a year earlier was priced at two to three trillion dollars heading toward public lists. The takeaway was not hype but paid use that earns back.

The infrastructure bill grew in the shadow of that usage. Five large hyperscalers spent 415 billion dollars last year and are tracking toward 820 billion dollars this year, roughly two and a half percent of United States GDP, a scale that still surprises even insiders. On data center leasing Blackstone sees the world's largest platform scale from one gigawatt in 2024 to two in 2025 and at least six this year, implying about 100 billion dollars of build cost plus a couple hundred billion more in chips from tenants. The missed corner was also named: Micron and SK Hynix led memory with share gains near five to six times, and even with that surge SK Hynix priced near four times earnings, far from Cisco at one hundred fifty times in 2000, a sign skepticism still travels with the trade.

Capital that was actually deployed in the past year filled the gap: GPU and TPU financing with Nvidia and Broadcom , neocloud bets, a joint venture with Google called Crux, Firmus in Australia, Neysa in India and a set of data-center operators worldwide. The list mattered because it turned an abstract forecast into committed capacity — the need for compute was not a slide, it was a set of signed builds. Recalling that a similar infrastructure call a year earlier proved correct gave the new, larger number more weight rather than less.

Power Is the Other Infrastructure

The second leg was power , framed a year earlier as not plastics but power after a line from The Graduate, and still central now. Utility capital spending reached 800 billion dollars over the past five years and is expected to almost double in the next five, with places like Canada talking hundreds of billions to expand and upgrade the grid. Turbine, cooling and electrical equipment makers lifted 44 to 81 percent in a single year, turning a boring corner of the market into a growth engine. Blackstone listed its own positions: a large credit for Sempra Infrastructure on a major LNG project, analytics firm Enverus , Williams building power plants alongside data centers, MacLean electrical gear, and in Europe renewables Eurowind and Sunotec . The point was simple — without electrons the chips and the halls do not run.

When the portfolio itself was scored, the return showed up directly: in the second quarter nine of the ten investments that appreciated most were tied to AI, with India's Aster Care hospitals the lone exception, and a similar mix was expected in the third quarter. That made the story internal rather than imported, and it explained why bets on energy transition and infrastructure stayed in place — demand was not only model curiosity but also a delayed grid build that must happen anyway.

From 1870 to 1900: A Rehearsal

A deliberate jump to the distant past asked how the same person would have lived on September 17, 1870 versus 1900. In 1870 a wood house, a horse and a candle; thirty years later a steel building, a rail car, a few automobiles and Edison light, more likely in a city. Numbers matched the image: annual labor productivity doubled, gross domestic product quadrupled, real manufacturing multiplied by six and the equity market by seven between 1870 and 1900. The parallel was that today feels closer to 1870, at the edge of an industrial-type shift, a reminder of how quickly technology can re-price economies and assets when the base changes.

Every boom carries a bust and the era's backbone was railroads — roughly two hundred carriers went bankrupt or insolvent. The difference was stressed: then supply was built ahead of demand on high leverage, much like the telecom overbuild; now demand leads supply , and much of what is signed — energy and halls — is with lightly leveraged counterparties. That does not remove risk, but it changes the character of the cycle and why capital is still being put to work with that gap in view. Caution and conviction were held together rather than traded off.

An eighty-four year leap landed on a school day in Highland Park, Illinois, the Super Bowl of that year and two commercials from the same game. One was the famous Apple Macintosh spot on January 24, 1984, saying 1984 will not be like 1984 — democratize computing from COBOL and MS-DOS to a mouse click, a forerunner for putting intelligence within reach of the smallest village today. The other was eighty-one-year-old Clara Peller with the line Where's the beef? A big puffy bun is nice, but where is the meat? That framed the central doubt for trillions in spending — is it circular finance or real return — and the answer was sought not in slogans but in operating cases.

Evidence on the Ground: Small Bets, Large Paybacks

Operating cases arrived as company stories. Phoenix Tower processed leases five times faster after a four million dollar AI build, earning about four and a half million a year for well over a one hundred percent return. In India IGI improved diamond grading, Tricon in rental housing closed applications ninety percent faster and lifted the customer experience, and Enverus in software saw an eighteen-times return on what it spends with model providers by fixing code faster. Single gains looked modest; together they explained why spending keeps rising when payback is visible.

Some stories created new businesses, not just efficiency. The analog garage-door firm Chamberlain used AI visualization to launch MyQ Secure View with face access at the door, a line already at a forty to fifty million dollar run rate and guided toward five hundred million within five years. Inside Blackstone the same pattern held: legal and compliance marketing reviews fifty percent more efficient, and a CIO portfolio intelligence agent compressing schedules by ninety-nine percent. Modest individually, together they trace a pattern where the agent moves from the desk into the business model itself.

Aggregated at the macro level the pattern sharpens. Annual United States productivity, averaging one and a half percent for a decade, ran at two point six percent over the past two and a half years, with the AI share still debated but the step up recorded. Hyperscaler revenue per employee rose sixty-five percent over three and a half years, EBITDA margins widened by five hundred basis points for the S and P and seven hundred for the portfolio over four years, and earnings growth that averaged fifteen percent reached thirty-two percent over the past twelve months even after stripping one-offs, with chief executives still guiding positively for the year ahead.

As agents leave the desk for the physical world, scale widens. Waymo miles under autonomous driving grew two hundred fifty times over two and a half years while serious crash rates sat ninety-four percent lower, feeding both use and trust. On the jobs side QTS data centers saw construction crews triple in under two years, employment across the portfolio kept growing despite AI efficiency gains, and new business applications doubled over a decade as an army of agents made it easier to start a company. In science, a Nature study showed phase one trial success moving higher on AI-found molecules and a McKinsey example cut trial length by forty percent, echoed by portfolio firm Advarra , with health framed as the most exciting frontier.

Three Beliefs, Four Bottlenecks

The forward view was reduced to three beliefs: use cases will spread, productivity gains will follow across medicine, law and the physical economy, and demand for intelligence will compound exponentially. The constraint is physical — chips, power and data centers . Bottlenecks were listed plainly: community entitlements and misinformation that can halt a near waterless hall, with a moratorium example in New York State ; a turbine queue at GE Vernova stretched to 2030 or 2031; empty shelves as a metaphor for chip scarcity, with hyperscaler build rising nine times in five years while chip makers barely doubled as a cyclical industry stayed cautious; and the ticket for a single gigawatt AI factory at fifty-five billion dollars, which is why only platforms that can marshal capital at scale can meet the queue. Without that build, use itself is throttled.

Closing brought risk and dispersion together. Cyber safety around financial or critical infrastructure could draw a political response that slows the frontier without stopping diffusion; data centers in orbit and on-device edge computing will come yet the mix still needs nuclear, gas and renewables much like power itself; ninety percent of advanced semis built in Taiwan keeps geopolitical tension in frame; and frothy private marks — ten billion dollar tags with little revenue, rich defense tech multiples — demand care, hence focus on the seniormost part of the stack at the compute layer. The retail split against Amazon — Kmart, Sears, Toys R Us fading while Walmart, Costco, TJ Maxx thrived on vital systems of record and a pivot from seats to outcomes — was offered as a model for white-collar dispersion, with compressed software multiples and a sixty-six percent fall in private software deals already hinting at it. Scarce assets from an Indian cricket team with ten slots for one point four billion fans to 7Brew coffee's Blondie, beachfront land and Rome airport essential infrastructure rounded out non-tech winners. The final answer to Where's the beef returned to the bun of chips, halls and power in trillions and the beef of return on investment that keeps rising — the same north star from Pete and Steve's founding forty years ago, to deliver for clients, now restated as why the enormous build can still constrain use rather than be constrained by it.