Companies that build and operate data centers urgently need stable power to push forward with large-scale AI compute expansion. This has spawned huge financing commitments and innovative deal structures, with on-site power systems able to reduce data centers' reliance on the public grid.

Investors are entering this power-shortage contest from several angles: making large investments in specific projects to build new on-site power plants, and, in other cases, acquiring equity stakes in on-site power service providers or providing external capital to finance equipment such as fuel cells, serving end customers like AI cloud providers.

This summer, a consortium led by Blackstone Group, with KKR and Apollo participating, agreed to contribute US$5.3 billion to acquire 49% of a portfolio of five gas-fired power projects from Williams, projects dedicated to supplying on-site power to data centers. Williams said that of the US$5.3 billion, US$4.4 billion will be used to cover the investors' corresponding share of project construction costs.

Although on-site power usually costs more than connecting to the public grid, availability and speed of construction are significant advantages. Williams, which recently entered the data center on-site generation business, said its first 200-megawatt project took less than 18 months from commercialization to grid connection and operation. Fitch Ratings noted that, by contrast, new grid-connected power plants and high-voltage transmission lines, the traditional grid infrastructure, often take 5 to 10 years.

Many generation projects are financed on a single-project, standalone basis. But some investors have already begun acquiring equity in on-site generation companies, with project types spanning gas-fired plants, fuel cells and photovoltaic plants, in order to capture broader growth returns rather than being limited to a single project.

For example, Blackstone and Halliburton announced in May a US$1 billion equity investment in VoltaGrid. The company specializes in building and operating on-site gas-fired generation systems for data centers.

Carl Bivens, a real estate partner at law firm Troutman Pepper Locke, said: "A large amount of private equity money has poured into the market, creating diverse deal structures. Private capital tends to hold ownership of the project development entity."

At the same time, the upfront investment to quickly build on-site power facilities is high, and customers such as small and mid-sized AI cloud providers need external capital to raise funds. This mirrors the logic elsewhere in AI infrastructure: some cloud providers set up special purpose vehicles (SPVs) and use external funding to purchase compute hardware such as GPUs. In the power sector, Brookfield Asset Management committed last autumn to provide up to US$5 billion in financing for Bloom Energy's subsequent AI data center power projects, with Bloom supplying the fuel cell equipment. In June this year, Brookfield raised the financing amount to US$25 billion.

A person familiar with the financing said Bloom Energy has relied for more than a decade on infrastructure funds and investment institutions setting up SPVs to fund product deployment; only recently has investor enthusiasm heated up substantially. Beyond Brookfield's broader framework, investors are also separately building on-site power plant projects for Bloom's customers. In July, project developer Industrial Development Funding and Oaktree Capital announced US$1.7 billion in project funding to deploy Bloom fuel cells at a Nebius data center. Oaktree provided minority equity financing, Mitsubishi UFJ Financial Group led senior debt financing, and Nebius pays Bloom for electricity.

But the on-site power financing boom brings the risk of cycle mismatches across the various stages of AI data center construction. Power assets can operate stably for decades, but large AI compute procurement contracts have much shorter terms, and GPUs in data centers face technological obsolescence within a few years.

Bivens said power infrastructure developers and investors prefer long-term contracts to lock in stable cash flow; investors are pushing customers to sign longer cooperation agreements and increasingly favor the 10-15 year commitments with 5-year renewal options offered by large cloud providers. "This can bring investors more stable returns over a longer cycle. It makes it easier for institutional investors to amortize the cost of building this behind-the-meter power solution or the data center."

But on-site power customers prefer to retain flexibility, especially in scenarios where they are waiting for the grid to expand supply. This raises a core question: over the long term, how much do data centers really need to rely on on-site power? For example, if future grid supply capacity improves, or GPU energy efficiency advances substantially, by the time large contracts expire, data centers' need for on-site power may already have changed.

Industrial Development Funding, the developer of the Bloom and Nebius project and an infrastructure investment institution, said that even if data center power demand changes, on-site power plant projects still have value. Its CEO, Nick Nunez, said the project owner could in the future sell surplus power into the public electricity market. For example, if a data center tenant does not renew after the initial contract expires, the loan on the power asset will most likely have been largely repaid, and it can still generate revenue continuously. "If a generator set completes most of its amortization within 6 to 8 years, we have many options and can sell power to the grid or other markets."

Convertible bond financing boom

AI compute expansion has driven a big year for technology sector financing, and convertible bonds have become a popular fundraising tool for tech companies. S&P Global Market Intelligence data shows that as of September 11, 2026, tech companies' convertible bond issuance accounted for about 60% of total US market issuance, raising about US$78 billion. Last year that proportion was only 44%. By comparison, tech companies accounted for 50% of equity follow-on offerings this year, 27% of investment-grade bonds and 20% of high-yield bonds.

This convertible bond issuance boom comes against the backdrop of ever-changing AI-specific debt financing. With a large supply of new bonds in the market, investors also find it difficult to judge whether AI investments will deliver returns. Convertible bonds usually carry lower coupon rates than ordinary corporate bonds, because investors can enjoy gains when the share price rises. They still dilute existing shareholders' equity, but generally to a lesser extent than directly issuing new shares.

Chris Fenske, head of capital markets research at S&P Global Market Intelligence, said in a webinar last week: "Clearly, convertible bonds have a smaller equity dilution effect than issuing new shares, so they have become the preferred financing channel." S&P's tally does not yet include recent issuances such as CoreWeave's US$4.2 billion convertible bond last week. But as more and more companies use convertible bonds, the cost of this type of financing will also rise, and we will watch whether these companies subsequently turn to equity financing.