Blackstone has spent the past two years buying the parts of the artificial intelligence economy that sit behind the chip, and the total now stands at $200 billion. The figure describes a wager that reaches from the chips installed in a server rack to the plants that supply the electricity to run them.
The size of the portfolio was reported on October 15 by Gelonghui, which said the firm’s data center holdings have grown into one of the largest positions of its kind. Blackstone has paired that real estate with a cloud computing business built alongside Google, has taken stakes in AI laboratories including OpenAI, and has lent money secured against Broadcom and Nvidia chips. It has also put capital into power plants and the equipment used to cool data centers, the two inputs that determine how quickly a campus can be built and how much of it can be operated at once.
Jon Gray, the firm’s president, said AI investment is the most important thing Blackstone is working on now. The remark was brief and, for a firm of its size, unusually direct about where the next decade of returns is expected to come from.
The approach differs from that of many peers. Several private equity firms have grown cautious about AI transactions as the sums involved have grown larger and payback periods have stretched further out, according to people familiar with the matter. Valuations for companies connected to data centers have risen sharply, financing has become more expensive, and the gap between what a project costs to build and what it earns in its first years has widened. Blackstone’s position is among the few that has not loosened.
Analysts said the portfolio is best understood as a bet on scarcity rather than on any single technology. Electricity, land near transmission lines, cooling capacity and the chips themselves are all limited, and each is controlled or financed by a different set of owners. A firm that owns pieces of several of those layers can profit regardless of which model developer wins, and can price its exposure as demand shifts.
Lending against semiconductors as collateral carries its own risks. Chip values depend on generation and on whether they remain in demand for training or inference work, and a new family of parts can reset the price of the previous one. Broadcom and Nvidia supply most of the accelerators being installed in new facilities, which means a concentrated set of borrowers is posting collateral whose value moves with a concentrated set of suppliers.
The push into power generation addresses the constraint that has slowed projects most often. Grid connections in the largest markets now take years to secure, and operators have responded by contracting directly with plants or building their own capacity. That work is capital intensive, regulated in ways that vary by state and country, and slower to produce returns than leasing an existing building.
Cooling has moved from a facilities problem to an engineering one. The racks being installed for the newest accelerators draw far more power per square foot than the equipment the industry was designed around, which has pushed operators toward liquid systems and made the suppliers of those systems a scarce asset.
The strategy also makes Blackstone a counterparty to companies whose prospects are tied to the same handful of assumptions. Its stakes in AI laboratories depend on those firms continuing to raise capital and spend it, and its lending depends on the borrowers finding demand for the computing capacity they are building. A slowdown in model development would not end the need for data centers, but it could change which of them are worth finishing.
The capital behind a position of this size comes from sources with long horizons. Insurance portfolios, pension funds and sovereign investors have been the largest suppliers of money to data center projects, and they have accepted returns that arrive years after construction begins because the assets are expected to hold value for decades. A firm that raises that money and deploys it across several layers of the same industry can keep the structure private while each individual project is financed separately.
The cloud business built with Google extends the same logic into services. Operators who own buildings and power can sell computing by the hour rather than only by the lease, which captures more of the value created by the demand and gives the firm a way to use capacity that a tenant has not yet committed to fill.
What would change the thesis is not a slowdown in model development but a change in how models are served. If the cost of running inference falls faster than the volume of it grows, the price tenants pay for computing could decline even as usage rises, and a portfolio built on scarcity would earn less than its construction budget assumes.
Mr. Gray has not said how the portfolio is financed, how much of it is held on the balance sheet of any single fund, or what proportion of the $200 billion has been deployed rather than committed. Those distinctions matter to investors trying to judge the risk, and the firm has not drawn them in public. What is clear is that the firm has placed its largest wager on the premise that computing capacity will remain scarce for long enough to earn a return on everything it is building to supply it.
