⚡ Energy

Private Equity Committed $45 Billion to AI Data Centers in One Week. The Risk Math Nobody Is Running.

Apollo, Blackstone, and KKR deployed $45 billion in AI infrastructure financing in four days. At the implied $35 billion per gigawatt, scaling to the industry's 20GW target could require $700 billion in PE-financed compute. The question nobody is asking: who holds the bag if demand falters?

By Tomás Reyes

June 16, 2026

Forty-five billion dollars. That is what three private equity firms committed to AI data center infrastructure between Monday and Thursday of last week, across two deals announced 72 hours apart. On June 9, Apollo Global Management and Blackstone launched a $35 billion platform to finance Anthropic's 1-gigawatt compute expansion using Broadcom's custom XPU chips. On June 12, KKR, Nvidia, the Kuwait Investment Authority, and Vistra launched Helix Digital Infrastructure with more than $10 billion in committed capital to build and operate data centers for hyperscalers.

These are not venture bets but infrastructure financings with 10-to-20-year time horizons, structured like toll roads and power plants, backed by institutional capital that expects steady returns over decades. Private equity has spent the last two years circling AI infrastructure the way it circled real estate in 2012: spotting an asset class that generates predictable cash flows, requires enormous upfront capital, and can be leveraged against long-duration revenue contracts. The difference now is scale, and the scale is staggering. U.S. data center investment hit $63.35 billion in 2025, with PE firms supplying more than $45 billion of it, over 70% of the total. Goldman Sachs Research now estimates hyperscalers will spend a combined $5.3 trillion on AI-related capital expenditures from 2025 through 2030, up from $4.5 trillion before Q1 earnings reports, and PE firms are positioning themselves to intermediate a growing share of that capital flow.

The Deals, Side by Side

Two different architectures, two different chip ecosystems, one thesis: hyperscalers cannot build fast enough on their own balance sheets. Broadcom CEO Hock Tan called it a "once-in-a-generation opportunity." Jon Gray, Blackstone's president, said compute demand is "too big to ignore." Adam Selipsky, the former AWS CEO now running Helix, framed it as solving complexity: "Large users of digital infrastructure have an urgent need to reduce complexity and unlock new capacity."

Nobody mentioned the word "risk," which is the interesting part.

The $35 Billion Per Gigawatt Problem

Run the numbers. The AI XPV Platform committed $35 billion to finance Anthropic's initial 1-gigawatt expansion, which works out to $35 billion per gigawatt of compute infrastructure, a figure that includes the chips, networking, physical plant, and power delivery but not the ongoing electricity costs. For comparison, Meta's Richland Parish data center in Louisiana cost roughly $13.5 billion per gigawatt ($27 billion for a planned 2+ GW facility), and that includes three dedicated gas-fired power plants and the transmission lines to feed them.

The gap exists because AI XPV is not just building shells and pouring concrete; it is financing Broadcom XPU deployments at frontier-lab density, which means more silicon per rack, more cooling per chip, and more networking per cluster than a general-purpose hyperscale facility. The $35B/GW figure is a ceiling for first-customer, first-tranche pricing, and scale should compress it considerably. If subsequent tranches come in at $15-25 billion per gigawatt, the 20 GW target by 2028 implies $300-500 billion in total AI XPV financing. Add Helix, add Blackstone's $160 billion development pipeline, add the deals already closed (Blackstone-PPL's $25 billion Pennsylvania JV, KKR-Energy Capital Partners' $50 billion partnership from 2024), and the total PE commitment to AI infrastructure crosses $700 billion before 2030.

That figure is approximately 13% of Goldman's $5.3 trillion hyperscaler capex estimate, which means PE is not supplementing the buildout so much as becoming a structural pillar of it.

Why Hyperscalers Cannot Do This Alone

Goldman Sachs' credit research team identified the binding constraint in a June 2026 report: public debt market saturation. Five companies want to borrow trillions. The hyperscalers already account for a disproportionate share of new investment-grade corporate borrowing. As they issue ever-larger amounts of debt, their weights in widely tracked bond indices climb, triggering portfolio concentration limits at institutional investors, and at some point bond buyers simply stop buying because they already own too much of the same five issuers.

PE sidesteps this constraint entirely because Apollo and Blackstone raise capital from pension funds, sovereign wealth funds, and insurance companies through private credit structures that never touch the public bond market. KKR's infrastructure platform already manages over $100 billion in assets, including more than $70 billion across digital and power infrastructure, and the Kuwait Investment Authority's anchor role in Helix signals that sovereign capital sees AI infrastructure as a generational asset class rather than a speculative cycle.

U.S. data centers consumed 64.4 GW in 2025, demand is projected to triple by 2030, and somebody has to finance the buildout at a pace the public markets cannot sustain.

The Beignet Problem: Who Bears the Risk?

The Meta Richland Parish deal is the Rosetta Stone for understanding PE risk transfer in AI infrastructure. Here is what happened, as documented by Earthjustice: On the same day Louisiana regulators approved Entergy's application to build three billion-dollar gas plants for Meta's data center, Meta restructured the deal. It sold 80% of the project to a Blue Owl Capital joint venture called Beignet Investors, borrowed $27 billion against the new entity, and renegotiated its lease to allow exit after four years instead of the original 15.

The gas plants have 30-year lifespans, which means if Meta walks in 2033, Entergy will not have recouped its construction costs, and Louisiana ratepayers will cover the shortfall through higher electricity bills. The Wall Street Journal called it "Frankenstein financing," Earthjustice petitioned the Louisiana Public Service Commission to investigate, and the PSC declined without explanation.

This is the template for how risk migrates through PE-financed AI infrastructure. PE structures systematically separate the entity that uses compute from the entity that owns the infrastructure and the entity that supplies the power. Each layer has its own exit clause, its own financial guarantee (or lack thereof), and its own risk transfer mechanism. When demand holds, everyone profits, but when demand falters, the question becomes: which layer absorbs the loss, and in Richland Parish the answer turned out to be the people who live there.

The Duration Mismatch

The core tension: data center buildings last decades, but the chips inside them become obsolete in 3-5 years, creating a structural mismatch between the asset and its purpose. PE firms are financing 20-year assets whose primary revenue driver (frontier AI model training and inference) could shift dramatically as model architectures evolve, inference efficiency improves, or demand plateaus. If model efficiency gains reduce raw compute demand per unit of output, a 2028 data center could be oversized by 2032, and the building would still stand, the power plant would still run, but the customer paying $35 billion per gigawatt may not need a gigawatt anymore.

This is not hypothetical speculation; it is the demand curve PE is explicitly betting on. Broadcom's Q2 FY2026 results showed AI semiconductor revenue of $10.8 billion, up 143% year-over-year, which is extraordinary growth. But 143% does not last. Growth rates at that level persist for the duration of an infrastructure buildout cycle, and then they flatten, which is precisely when PE return targets start to bind.

Three Firms, One Chokepoint

Blackstone has $150 billion in data center assets and a $160 billion development pipeline, KKR's infrastructure arm manages over $100 billion with $70 billion concentrated in digital and power, and Apollo is leading the largest single AI infrastructure financing in history. Three firms now collectively control or finance a meaningful fraction of the physical infrastructure that frontier AI models depend on to train and run, a concentration of leverage that has no precedent in the history of computing infrastructure.

Goldman Sachs projects private credit assets under management could exceed $3 trillion by 2030, and in that scenario PE firms do not merely finance AI infrastructure but become the landlords of it, extracting returns proportional to their tenants' dependency for decades to come.

Limitations

The $35B/GW implied cost is derived from a single first-tranche deal (AI XPV financing Anthropic's 1 GW expansion) and likely overstates the per-gigawatt cost at scale. Subsequent tranches targeting the remaining 19 GW will almost certainly achieve lower unit costs, but those terms have not been disclosed. The $300-500 billion total AI XPV financing estimate uses an assumed $15-25B/GW range for later tranches, which is an informed estimate, not a contractual figure.

PE returns and fee structures for these specific deals are not public. The 15-20% IRR target used for comparison is an industry benchmark for infrastructure PE, not a disclosed figure from Apollo, Blackstone, or KKR. Helix Digital Infrastructure has not published a GW target, so its capacity cannot be directly compared to AI XPV.

The Goldman Sachs $5.3 trillion hyperscaler capex estimate is a forward projection, not committed capital, and is subject to revision as AI spending plans evolve.

The Strongest Counterargument

PE is solving a real problem, and the alternative is worse. Speed matters. Without private capital, the AI infrastructure buildout slows at precisely the moment Goldman Sachs has identified public bond market saturation as a binding constraint on hyperscaler financing. If Anthropic, OpenAI, and Google cannot raise debt fast enough to build the data centers they need, inference costs remain high, AI productivity gains stall, and the economic benefits of frontier AI accrue more slowly to the broader economy. The premium PE extracts, whether it is 200 or 500 basis points above public debt rates, is the cost of speed. Meta's Richland Parish deal, for all its structural complexity, put a 2 GW facility under construction in months rather than years, and the Earthjustice complaint highlights real ratepayer risk, but the data center also brought thousands of construction jobs and hundreds of permanent jobs to a rural Louisiana parish with a median household income of $31,000. Capital has a cost, and so does waiting.

The Bottom Line

If you manage a pension fund or endowment: understand that PE-financed AI infrastructure is now an asset class, and you are likely already exposed to it through your alternatives allocation. The duration mismatch between 3-5 year chip cycles and 20-30 year physical assets is the risk to price. Ask your PE manager what happens to returns if AI inference efficiency improves 10x by 2032 and raw compute demand plateaus.

If you live in a state where utilities are building power plants for data centers: your electricity bill is now correlated with AI demand, so monitor your public utility commission filings closely. The Earthjustice petition template is publicly available and worth studying. Ask whether the data center customer has an early exit clause, and who pays if they use it.

If you work in AI: the economic structure of your industry just changed, because three PE firms now control an increasing share of the physical infrastructure your models run on. When the buildout phase ends and the rent-extraction phase begins, inference costs will have a floor set not by chip prices or electricity rates but by PE return targets, and that floor does not come down.

Related

• 69 U.S. Counties Have Now Banned or Restricted Data Centers
• Your Electricity Bill Is About to Subsidize AI
• Alphabet's AI Capex Now Exceeds Its Free Cash Flow