Energy

Blackstone Takes 49% of Five AI Power Plants

Williams sold 49% of five dedicated power projects to Blackstone, Apollo and KKR for $5.34 billion. Insurance money now underwrites the electricity behind AI - and that puts a floor under the price of compute that cheaper chips cannot cut through.

EnergyBy Servola Tech Desk2026-07-146 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Blackstone Takes 49% of Five AI Power Plants

Key takeaways

On 13 July 2026 Williams Companies sold a 49 percent non-controlling stake in five dedicated power projects to Blackstone Credit and Insurance, with Apollo and KKR insurance vehicles alongside, for $5.34 billion - about 4.9 billion euros - made up of $4.4 billion of growth capital and around $0.9 billion of additional consideration.
This is insurance capital, not venture capital, and insurance capital requires a contracted long-dated yield, which means the electricity behind AI compute now carries a required return and a buyout window running from year seven to year fourteen.
That return does not disappear when chips get cheaper. It puts a floor under the cost of a compute hour that no improvement in silicon can cut through, because the power contract underneath it still has a coupon to pay.
The portfolio is 2.6 gigawatts announced against a backlog of more than 6 gigawatts, and Williams keeps 51 percent with operational control - the model is dedicated generation for large loads, which is the answer European operators reach for when the grid queue says 2032.

Who wrote the cheque

Read the names on the announcement and the story changes. Blackstone Credit and Insurance led the investment. Alongside it: insurance vehicles and accounts managed by Apollo and by KKR. This is not growth equity chasing a multiple on an AI story. It is the capital that stands behind life insurance policies and annuities, and it has one defining requirement, which is a predictable, contracted, long-dated stream of payments.

What it bought, on 13 July 2026, is 49 percent of a joint venture holding five power projects owned by Williams Companies: Socrates, Apollo, Aquila, Socrates the Younger and Neo. The total is $5.34 billion, roughly 4.9 billion euros, of which $4.4 billion is growth capital and around $0.9 billion is additional consideration. Williams retains 51 percent and, importantly, commercial and operational control. The stake is non-controlling. The structure includes a buyout option, valued at Blackstone's outstanding investment balance, exercisable between years seven and fourteen.

These are not merchant plants selling into a market. They are dedicated generation built to serve large loads directly, which in 2026 means data centres. Williams describes a backlog of more than 6 gigawatts, of which 2.6 gigawatts is announced.

What a required yield does to your compute bill

Here is the part that does not appear in any coverage of this deal, and it is the reason an operator should care about a Tulsa pipeline company's joint venture.

When insurance money funds a power plant, the plant must produce a contracted return over decades. That return is not a hope about future demand. It is an obligation, matched against liabilities that an insurer owes to policyholders, and it is why the structure has a defined buyout window rather than an exit into a hot market. The consequence is that the electricity flowing into that data centre now has a fixed cost of capital attached to it for the life of the contract.

Now trace it forward to what you actually buy. The price of an hour of AI compute is roughly the cost of the chip, amortised, plus the cost of the power to run it, plus the operator's margin. Everyone in the industry is betting the first term falls, and it probably will: competition in accelerators is real, and the second-source pressure on Nvidia is now visible in silicon. But the second term has just been contractually fixed by people who do not accept variability. Cheaper chips will not reduce a power payment that was underwritten to yield.

That is a floor. It means the widely held assumption that inference gets structurally cheaper forever needs an asterisk: it gets cheaper until it hits the cost of the electricity, and the electricity has now been financed by counterparties who are paid to make sure that number does not move.

Bring your own power is becoming the price of a large load

Why does a hyperscale operator build its own plant rather than plug into the grid? Because the grid has a queue, and the queue has a date on it that nobody likes. That is the thread connecting this deal to everything happening in Europe right now: connection dates measured in years, transmission that has not been built, and system operators who cannot promise capacity to a 500 megawatt load on the timeline the load requires.

Dedicated generation is the workaround, and the workaround is being institutionalised. When Blackstone, Apollo and KKR put insurance money into five of these projects at once, they are not making a bet on one campus. They are declaring the category financeable at scale, which means the next fifty projects have a template, a comparable, and a source of capital that did not exist for them two years ago.

European operators should read that as a forecast rather than as American news. The same grid constraint exists here in a harder form, because the queues are longer, the permitting is slower, and the political tolerance for a data centre consuming the connection capacity of a small city is lower. The structure that solves it will arrive in Europe with the same logic and the same investors, and it will land into a market where dedicated generation also has to satisfy heat reuse rules, efficiency reporting and local objections. It will be more expensive here, not less.

What this changes for an operator who buys compute

If you buy GPU capacity rather than build it, three things follow.

Treat long-term compute pricing with more scepticism than the vendor roadmap invites. A quotation that assumes falling unit costs across a three-year term is assuming away the component of the cost stack that has just been locked in by people whose job is to lock things in. Ask a provider what proportion of their power is contracted, at what price, and for how long. A provider who cannot answer that is not pricing their own input.

Second, note who your counterparty actually is. A neocloud renting you capacity in a campus financed this way is passing through an obligation, not absorbing it. That is fine, and it is knowable, but it changes what happens to your price if their utilisation falls: the coupon does not fall with it.

Third, if you are siting anything in Europe that needs meaningful power, get the connection conversation started earlier than feels reasonable, and treat the connection date as the project's real critical path. The reason American operators are buying power plants is that they discovered the alternative was waiting. That discovery is arriving here.

The quiet repricing

The AI infrastructure story has been told as a chip story for three years. It is becoming an energy story, and specifically a financing story, which is where the durable costs get set. Chips are a purchase, and purchases get cheaper. Power under a twenty-year contract with an insurance counterparty is a liability, and liabilities do not.

Nobody announced a price increase this week. What happened is more consequential and much less visible: the cost base under the thing you rent by the hour was quietly moved from a market that fluctuates into a contract that does not. Price your next three years accordingly.

Frequently asked questions

This is a US deal. Why should a European operator care?

Because you buy compute in a global market and a large share of it is trained and served in exactly these campuses, so the cost structure being set in Oklahoma and Ohio ends up in the price you pay in Frankfurt. The second reason is that the constraint driving the deal - a grid that cannot connect a large load on the timeline the load needs - is worse in Europe, so the same solution and the same financiers will appear here, into a stricter regulatory environment.

Does dedicated generation for data centres make electricity more expensive for everyone else?

It is genuinely contested, and the honest answer is that it depends on whether the new generation is additional. If a campus builds and pays for capacity that would not otherwise exist, it is not competing with households for existing supply. If it takes grid capacity or scarce equipment and connection resources that others were queuing for, it is. That distinction is exactly what European regulators will end up legislating about, and an operator siting a large load should expect to have to prove which of the two it is doing.

Should we assume compute prices stop falling?

No. Assume they stop falling as fast, and that the falls come from silicon and utilisation rather than from energy. The useful planning position is to model the power component of your compute cost as flat to rising in real terms over your planning horizon, and to treat any vendor forecast that shows it falling as a claim requiring evidence about their actual power contracts.

The interesting number in this deal is not $5.34 billion. It is the buyout window that runs from year seven to year fourteen, because that is a statement about how long the electricity behind AI compute has been promised to earn. Chips are the part of the AI bill that gets cheaper. Power is the part somebody has now underwritten, and underwritten costs do not fall. Build your compute budget on that asymmetry rather than on the assumption that everything in this industry deflates.

BlackstoneWilliams Companiesbehind-the-meterdata centre powerprivate creditAI infrastructureApolloKKR

Samsung's Quarter Set to Beat Nvidia on Profit

Samsung is expected to post roughly 85-86 trillion KRW in Q2 operating profit, edging past Nvidia to become the world's most profitable tech company. Here is what owners and operators should read into it.

3 min read

Energy

Japan's Energy Giants Just Bet on Drilling for Heat

Quaise raised a 134 million dollar first close, backed by Japan's JERA and Idemitsu, to build the first commercial superhot geothermal plant. Here is what it means for operators.

3 min read

Digital Sovereignty

The EU Council Revived the Rule That Scans Private Messages

The EU Council revived the lapsed rule that lets providers scan private messages with AI. For owners, confidential communication is now a governance choice.

2 min read

← All articles