How Investors Are Financing On-Site Power to Break AI Bottlenecks

Companies building and operating data centers are hungry for ways to secure the power needed to move ahead with massive AI build-outs. That’s prompted big financing pledges and creative dealmaking for on-site power that can cut data centers’ reliance on electric grids.

Investors are getting in on the crunch through several angles, including large investments in specific projects, such as new on-site power plants. Some investors are also taking equity stakes in companies that provide on-site power, or providing outside capital to finance equipment like fuel cells for end customers such as AI cloud providers.

A Blackstone-led group, including KKR and Apollo, this summer agreed to invest $5.3 billion for a 49% stake in five gas-fired power projects pipeline giant Williams is developing to supply on-site power for data centers. Of the $5.3 billion, $4.4 billion will cover the investors’ share of expected project costs, Williams said.

Though on-site power can be pricier than connecting to the grid, availability and speed is an obvious advantage. Williams, which recently expanded into developing and running on-site generation for data centers, said last month it brought its first 200 megawatts online less than 18 months after commercialization. By comparison, Fitch Ratings estimates that building new infrastructure such as grid-connected power plants and high-voltage transmission lines can take five to 10 years.

Many power-generation deals have been financed primarily on a project-by-project basis. But some investors have also started buying stakes in companies that provide on-site power generation, whether that be gas plants, fuel cells or solar projects, giving investors exposure to broader growth rather than a specific project. For instance, Blackstone and Halliburton in May announced a $1 billion equity investment in VoltaGrid, which builds and operates on-site gas-powered systems for data centers.

“With all the private equity and other money that has come into the market, there’s different structures. And when there’s private equity, they want to have an ownership interest in the developer entity,” said Carl Bivens, a real estate partner at Troutman Pepper Locke.

Meanwhile, the up-front costs of bringing power on-site to data centers quickly has created a need for outside capital to provide financing for customers such as smaller AI cloud providers. That’s similar to other areas in AI infrastructure, where some cloud providers have been relying on special purpose vehicles that use outside capital to finance expensive assets like graphics processing units.

On the power side, Bloom Energy announced last fall that Brookfield Asset Management had agreed to provide up to $5 billion to finance future power projects for AI data centers, with Bloom providing the fuel cells. In June, Bloom announced that Brookfield had expanded the amount of potential financing to $25 billion.

Bloom has relied for more than a decade on infrastructure funds and investment firms to form SPVs that finance deployments of its products, but investor interest accelerated more recently, a person familiar with the financings said.

Investors are also financing on-site power projects for Bloom customers, separately from the broader Brookfield framework. In July, project developer Industrial Development Funding and investment firm Oaktree Capital Management announced a $1.7 billion project investment to install Bloom fuel cells at Nebius data centers. Oaktree is contributing minority equity financing, with MUFG leading senior debt financing, while Nebius will pay Bloom for power.

But the rush of financing for on-site power creates a potential timing mismatch across different parts of what’s needed to get an AI data center up and running. Some power assets can operate for decades, while big AI compute contracts are often much shorter and the GPUs inside data centers can become outdated in just a few years.

For developers and the investors financing power infrastructure, longer contracts that lock in predictable cash flow are preferable. Investors are pushing for longer customer commitments to underpin these deals, increasingly favoring 10- to 15-year commitments from large cloud providers with five-year renewal options, Bivens said.

“It provides a longer runway of guaranteed rent for the investors. It allows [institutional investors] to amortize their costs of the behind-the-meter solution or the data center build,” he said.

Customers of the on-site power may want more flexibility, particularly if they’re awaiting grid power. That points to the bigger question of how much data centers will need to rely on on-site power in the long term. For example, more grid power could become available or GPUs could become more energy efficient, and data centers’ on-site power needs might evolve by the time big contracts expire.

Industrial Development Funding, an infrastructure investment firm and the developer behind the Bloom-Nebius project, says on-site power projects can still be valuable even if the data centers’ power needs themselves change. The owners of the project could sell power into broader markets down the line, said IDF CEO Nik Nunes. For instance, if a data center tenant moves on when its initial contract term ends, the power asset could be mostly paid off but could still generate revenue.

"If I have a power block that is significantly amortized in six to eight years, I have a lot of optionality to sell that power into the grid or into other markets,” Nunes said.

Converting to Convertibles

Convertible bonds have been one particularly popular way for tech companies to raise cash in a banner year for financing across the sector, fueled by the AI build-out.

Technology companies have accounted for about 60% of overall U.S. convertible bond issuance so far in 2026, raising roughly $78 billion through Sept. 11, according to S&P Global Market Intelligence. That’s up from around 44% of that market last year. By comparison, tech companies accounted for 50% of equity follow-on offerings, 27% of investment-grade debt and 20% of high-yield issuance so far this year.

The convertible surge comes as financing costs for AI-specific debt have been rising. Investors are swamped with new supply and grappling with uncertainties about how AI investments will pay off. Convertible debt often carries lower coupons than straight corporate bonds since investors get an upside if the stock rises. That can eventually dilute the holdings of existing shareholders, but potentially less so than equity offerings.

“Obviously, there’s the benefits of being not as dilutive as a follow-on. So it has been one of the places of priority,” Chris Fenske, head of capital markets research at S&P Global Market Intelligence, said on a webinar last week. And S&P Global’s total doesn’t count recent convertible issuance like CoreWeave’s $4.2 billion sale last week. But as more companies tap convertibles, that form of financing too could get more expensive—we’ll see whether those firms shift to equity raises next.