Jon Gray Is Right About the Century. The Exit Is a Decade Problem.

Blackstone's data-centre book went from $55bn to $200bn in 25 months on a 19th-century analogy. In the same month, the two channels it needs to sell out of — sovereign buyers and public listings — both closed. The operating case is winning. The exit case is what the equity has already marked and the credit has not.

Jon Gray has a story he tells with pictures. In September he stood in front of more than a hundred chief investment officers and put up an AI rendering of himself as a 19th-century farmer, pitchfork and all, then the same man in a suit in 1900. Between those two images, he said, the world replaced wood, horses and candles with steel, railroads and electricity. We are in one of those moments. It is a good argument, and it is the argument on which Blackstone has put $200bn into data centres.

Here is what the picture leaves out. A technology can be a once-in-a-century event and still ruin the people who finance it. The railroads were exactly that transformation, and the subscribers who funded the later waves of them lost most of their capital. The network survived; the money did not. What separates those outcomes is the exit — the price at which the first owner hands the asset to the second. In October, both of the doors Blackstone needs for that handover moved at once, and almost nobody repriced it.

What $200bn actually is

The number itself is not the story; the speed is. In June 2024 Blackstone's data-centre exposure, including debt and projects still under construction, stood at $55bn. By July 2026 it was $200bn. That is roughly $6bn of new exposure a month, sustained for 2 years, into a business the public market spent the same period marking down.

Put it against the rest of the firm and the concentration becomes obvious. Blackstone's entire infrastructure book, including debt, is $218bn. Its entire real estate book, the business Gray grew up in, is $608bn. Data centres alone are therefore something close to 92% of infrastructure — which means the “infrastructure” allocation that institutional clients thought they were buying for its low correlation and inflation linkage is, in practice, a single levered bet on one industry's capital cycle.

The origin story matters because it explains the confidence. Blackstone bought QTS Realty Trust in 2021 for $10bn and underwrote its land bank — thousands of acres with power already contracted — as roughly a decade of runway. ChatGPT shipped in November 2022. The land ran out 18 months after the deal closed. When you underwrite 10 years of runway and it runs out in 18 months, the lesson you learn is that the constraint is demand. It was, for about 3 years.

The book is not one asset. It is one counterparty.

Read the deal list and a pattern emerges that the word “diversified” conceals. There is the real estate: QTS, AirTrunk in Asia-Pacific, Firmus and Rowan. There is the credit: an $8.5bn May loan to CoreWeave anchored by Blackstone and secured on chips and contracts with Meta, and a role in the $95bn raised to date to fund chips for Anthropic alongside Apollo, Broadcom and the banks. There is the equity: $5bn into Crux AI, the cloud joint venture with Google, or $25bn counting debt. There is the $500bn pledge with Apollo, BlackRock, Brookfield, Goldman and KKR to fund Nvidia's customers. And there are the labs, through BXPE.

Six instruments, four asset classes, three continents. But trace the cash flow to its source and almost all of it terminates in the same handful of balance sheets: Microsoft, Meta, Alphabet, Oracle, Amazon. A hyperscaler — the short industry term for the five or six firms that rent compute at national scale — is the tenant, the offtaker, the credit support and, in the case of Google, the joint-venture partner. The diversification is across instruments on one credit. That is a perfectly respectable structure. It is also not what an allocator means when they say they want infrastructure exposure.

Jeffrey Hooke of Johns Hopkins put the specific risk in one line to Bloomberg: “The dependence of data centers on hyperscalers and contracts is as yet untested.” He was blunter about the wrappers: “Sticking these into retail and other special funds could prove to be a challenge if things don't go according to plan.” The retail point is not incidental. Blackstone raised $2bn for BXDC, a trust that buys new data centres, and the lab exposure sits in BXPE, a vehicle for wealthy individuals. Vehicles with periodic redemption features are being filled with assets that have no observable secondary market. That is a maturity mismatch wearing a real-asset costume.

Gray answered the right question with the wrong chart

At the September address Gray did something intellectually honest and, I think, revealing. He reached for the 1980s Wendy's advert — where's the beef? — and asked the question his own investors are asking: is this just circular financing, money lending itself money until someone loses it all? His answer was the AI uptake, the return on investment, a cartoon patty labelled “ROI from AI” inside a bun made of chips, data centres and power.

The demand half of that answer is not in dispute, and the receipts are strong. Blackstone will lease about six gigawatts of capacity this year, on its own estimate — roughly six times its 2024 volume, and a gigawatt being enough to power some 750,000 homes. BloombergNEF expects data centres to consume about 20% of US electricity in 2035, against 5.9% today. PricewaterhouseCoopers puts cumulative global data-centre capital spending at $31.6trn by 2050, with the US taking close to half. If you believe those numbers, the build is not a bubble. It is behind schedule.

But look at what the question was and what the answer addressed. Where's the beef is a solvency question: does the cash flow exist to service the paper? The ROI chart answers demand. In a business you own forever those are the same question. In a private-equity fund they are not even the same decade. The fund's return is a function of three things — entry price, operating improvement, and the multiple someone else pays at exit. Gray has now published an overwhelming case for the second and no case at all for the third. That is not a criticism of his analysis. It is a description of what the analysis was for.

Both exits closed in the same 4 weeks

Here is the empirical core of this piece, and it is a calendar fact rather than an opinion. Private-markets exits run through two doors: a sale to a strategic or sovereign buyer, or a listing. In October, both doors moved.

Door one: Norway's $2.1trn sovereign wealth fund, the single largest pool of patient capital on earth, said it is steering clear of unlisted data-centre investments. That fund's entire institutional purpose is to hold illiquid long-duration assets through cycles. When it declines a category, it is not making a tactical call; it is telling you the category does not meet its definition of infrastructure.

Door two: Firmus Grid, a Blackstone-backed data-centre company, scrapped a planned listing in Australia after failing to attract global investors, and AirTrunk delayed a Singapore offering. Two vehicles, two markets, 1 month, both withdrawals rather than discounts. A pulled IPO is a very specific piece of information. It is not a price signal — nobody found out what the clearing level was. It is the absence of a price signal, which is worse, because it means the marks on the book are now the only numbers anyone has.

Sunaina Sinha Haldea of Raymond James framed the requirement precisely: investors in these vehicles “will need to price technology obsolescence and contract quality far more carefully than in conventional infrastructure,” and “while Blackstone has taken leadership, it has to ensure it deploys capital and cashes out at right times to maximize returns, else the strategy can unravel.” Cash out at the right time. That is the whole risk, stated as a housekeeping item.

The reason the buyer list is short is structural, not cyclical. A $200bn single-theme book has maybe a dozen potential purchasers globally, and each of them is already being asked to take the same trade by Apollo, Brookfield and KKR simultaneously. Add the concentration limits that most allocators now apply to AI as a single exposure, and the honest description is a market with many sellers arriving and very few buyers able to write the ticket.

The asset lives less than the debt

There is a physical reason the exit is hard, and it is the most under-discussed thing in the whole trade. Conventional infrastructure is underwritten on an asset that outlives its financing — a toll road, a substation, a water utility. A data centre is a shell with two lives. The building can last 30 years. The compute inside it is depreciated over five to 6 years on the hyperscalers' own schedules, and the hall itself is built around a power density and a cooling design specific to a generation of silicon.

Rack density is the tell. Halls built through the 2010s were designed around single-digit kilowatts per rack. The current GPU generation wants an order of magnitude more, which is why the industry is talking about liquid cooling at all. A hall built to the old specification cannot be re-tenanted to the new one without being rebuilt, which means many of these assets are effectively single-tenant for their economic life rather than multi-let. Obsolescence is not a tail risk here. It is the base case, arriving on a schedule you can roughly date from the commissioning year.

Now do the arithmetic that no mark-to-model exercise performs. Suppose the book is financed at something like 70% loan-to-value, an assumption rather than a Blackstone disclosure. That is $140bn of debt against $200bn of gross value, leaving a $60bn equity cushion, and the debt is sized against the residual value of the asset rather than against the contracted rent. Now let the residual come in a quarter below carrying — not a catastrophe, just a hall that does not re-tenant at the same rate in year eight. That is $50bn of value, against a $60bn cushion. The cash flow can be perfect for 8 years and the equity is still almost entirely gone, because the equity is the residual claim and the residual is the only thing in doubt.

Rising rates make this worse in a way that is easy to state and hard to internalise. The 2yr par yield has gone from 3.47% on 2 January to 4.77% on 19 October, roughly 130bps, with the 30yr up about 60bps over the same stretch. On $140bn of floating or refinancing exposure, 130bps is close to $1.8bn a year of additional carry — before a single tenant has moved out. The discount rate applied to a 2033 exit has moved against the trade this year, quietly, while the demand story kept improving.

What equity has marked and credit has not

This is where the piece becomes tradeable rather than merely interesting. Two markets are looking at the same asset base and reaching incompatible conclusions.

The equity market has already taken a view. Blackstone's stock is down almost 50% from its 2024 peak, through a period in which the operating story — demand, leasing, capacity — kept strengthening. A halving of the equity against a better operating story is the market saying something specific: it is not the cash flow that worries us, it is the ending. That is a rational reading of a company whose growth now depends on the one thing it cannot demonstrate — a completed, value-realising exit.

The credit market is saying something else. High yield option-adjusted spread — the extra yield over Treasuries you are paid for taking a corporate's default risk — sits at 3.31%, and investment grade at 0.85%. The high yield level is tighter than 4 days in five over the past 3 years. Investment grade is 12bps off its tightest print in the same window, and sat higher for half of it. Credit is pricing a contracted, serviced, performing asset. It is pricing the lease, not the exit.

Both of those readings are internally coherent, which is why the gap persists. Equity is a perpetual claim and must price terminal value. Credit is a dated claim and only needs the coupon to arrive. A 5yr loan on a data centre does not care whether the building is re-tenantable in 2034. A share of equity cares about nothing else. The mismatch is not a mistake by either market. It is a difference in horizons, and it means the risk is being transferred from the equity holder to whoever holds the residual when the music stops.

The uncomfortable corollary: if the equity is right about the exit, the credit is mispriced. Not because default is imminent — it is not — but because recovery given default depends on the same residual value the equity is now discounting. A loan against an obsolete hall is a loan against scrap and land, and scrap-and-land recoveries are not 70 cents.

The railway rhyme, told correctly

Back to the pitchfork. Gray's analogy is not wrong; it is unfinished in the one place that matters to an investor. In the 1840s Britain raised capital for railways on a scale that dwarfed anything before it, on exactly the argument Gray is making — that the world was about to be remade. It was. The network that resulted carried the economy for a century. And a large share of the subscribers who funded the second and third waves of it ended up with permanently impaired capital, because they bought the right technology at the wrong point in the financing cycle and needed a buyer to agree with them later.

The fibre version is closer in structure. Between 1999 and 2001 the world laid enormous quantities of long-haul fibre, most of it never lit, financed by carriers that went bankrupt. Global Crossing and WorldCom did not survive. And then the fibre was bought out of bankruptcy for cents on the dollar and became the backbone of everything that happened in the following 15 years. The asset was right. The first owner was wrong. The second owner's return was determined entirely by the price he paid, and he paid almost nothing because there were no other bidders.

That is the shape of the risk here, and it is why “once in a century” is simultaneously true and useless as an underwriting statement. It is a claim about the technology, deployed to support a claim about the security. Those are different propositions with different evidence requirements, and only the first has been evidenced.

So the position is not that Blackstone is wrong. It is that the firm has built the correct asset base at the moment when the mechanism for realising it has narrowed to almost nothing, and that the operating story is so good it is currently doing the work of an exit story it cannot supply. Watch the exits, not the demand. One completed sale or listing at or above carrying value, in this market, at this size, would settle the question in the bulls' favour far more cheaply than another gigawatt of leasing.