The rise of artificial intelligence is doing more than just boosting electricity demand—it's fundamentally reshaping the entire business model of data centers. As securing power from public grids becomes increasingly difficult and slow, developers are now pushing their involvement upstream, moving directly into energy development. This involves acquiring power generation assets, negotiating fuel supply contracts, and taking on the ongoing responsibility for operations and maintenance. This significant shift is being driven by the unprecedented scale of data centers currently in the planning pipeline.

Traditional data centers typically require electricity in the tens of megawatts. In contrast, the new large-scale AI computing campuses being planned today demand power at the gigawatt level, a consumption load comparable to an entire city. US utility companies are struggling to keep pace with this explosive growth rate from data centers. Grid construction projects face strict regulations, require massive capital investment, and involve lengthy construction timelines. According to data from the global commercial real estate services firm Jones Lang LaSalle, the average wait time for grid interconnection in the US is approximately four years. The 2026 JLL Global Data Center Outlook highlights the severity of the issue, noting that a single 50-megawatt grid connection application in Northern Virginia could face a seven-year wait. A 2025 industry survey further corroborates this, with 44% of data center companies reporting grid interconnection wait times of four years or more.

The availability of power has now become the single most critical factor in site selection. Michael Thomas, founder and CEO of Cleanview, a Colorado-based research firm focused on data center development, noted that hyperscale cloud providers previously prioritized land, fiber optics, water resources, and proximity to transmission lines, with the availability of sufficient power or a clear path to power delivery being a secondary concern. However, he says, "The situation has completely changed over the past two to three years. Now, the only question that matters is: does this land have power?"

As a direct result, a large number of developers who are unwilling or unable to wait for grid expansion are choosing to build their own dedicated power generation facilities on-site. Others are forming alliances with utility companies, power generators, and infrastructure investors to develop new power capacity specifically tied to their projects. While these approaches can significantly shorten the time to power delivery, they also shift the various risks traditionally borne by utility companies—including permitting, fuel supply, equipment procurement, financing, and operations and maintenance—onto the data center developers and their partners.

A deep dive reveals that grid capacity constraints are forging a new type of partnership. Power supply has become the primary constraint in choosing a site for hyperscale computing projects. Previously, hyperscale cloud providers relied mainly on long-term Power Purchase Agreements (PPAs) to absorb their growing electricity needs, and these agreements also provided the financial backing for new wind and solar projects. In 2024, Google signed clean energy PPAs totaling over 1.5 gigawatts, while Microsoft contracted a staggering 19 gigawatts of renewable energy across 16 countries. Meta announced in 2025 that it had procured more than 15 gigawatts of clean renewable energy. While PPAs are effective for financing new generation projects, they do not solve the physical challenge of delivering that power to a specific data center. They cannot build substations, alleviate transmission bottlenecks, or shorten the grid connection queue. Over the past year, the industry has rapidly shifted from external PPAs to on-site self-generation, a trend that is only expected to accelerate.

For instance, xAI operates two data centers, "Colossus 1" and "Colossus 2," on the outskirts of Memphis, supported by nearly 1.5 gigawatts of behind-the-meter gas turbine generation. According to Cleanview's statistics, as of June this year, there are 59 data center projects in the US planning a combined total of about 90 gigawatts of behind-the-meter power capacity. The 2026 State of the Data Center Report from the Data Center Industry Association (AFCOM) shows that 25% of surveyed companies already have on-site generation, up from 19% the previous year. Furthermore, 23% plan to implement on-site generation within the next 12 months, and another 17% are currently evaluating such options.

This move to behind-the-meter generation transfers risk and, in doing so, creates new collaboration models. Data center operators adopting behind-the-meter generation must now take on responsibilities and risks that were once the domain of utility companies. This includes securing and maintaining environmental permits, negotiating fuel contracts, procuring equipment, ensuring equipment reliability, and addressing opposition from local communities. These constraints have already caused project delays. For example, the "Stargate" project in New Mexico has seen its progress slow down, with an August announcement stating that the natural gas pipeline project intended to supply its 2.5-gigawatt on-site power plant is now delayed until 2027. In another case, a Microsoft-affiliated project in New Jersey faced environmental审批 hurdles for its on-site gas-fired power generation plans in August. Fuel price volatility, environmental compliance, and power equipment maintenance are routine for utilities and power companies but represent entirely new territory for most data center developers.

Kemal Hawa, a partner at the law firm Kirkland & Ellis and a data center development advisor, emphasizes the need for caution: "You must fully consider the realistic operational issues. Many of these risks are not within a company's control." As developers become more involved in power supply, they are also creating new partnership models with utilities, power producers, and infrastructure investors. In this model, the partner handles the financing, construction, and operation of the power plant. In turn, long-term purchase commitments from data center customers reduce the risk of a newly built plant having no customers, which helps secure project financing. Developers may also take on some of the downside risk of project delays, scaling back, or cancellation through minimum payment clauses, security deposits, and termination fees.

Mercer, an industry observer, notes the emergence of three primary models for this transition. The first is complete self-supply. For example, xAI is building a 1.2-gigawatt natural gas plant to support its computing expansion in the Memphis area, where Colossus 1 and 2 are already operational and small gas turbines have been deployed on-site. Thomas describes this as a typical "Musk-style" approach: "SpaceX builds all its parts, Tesla builds all its parts. This is more of a corporate culture preference, albeit an extreme one."

The second model combines the public grid with on-site generation. Thomas illustrates this with an example where a utility tells a developer, "We can supply you with power 90% of the year." Since the project requires 100% power reliability, the developer builds its own gas-fired plant to fill the gap. The third model involves bringing in a non-utility, specialized power provider. Hawa notes that many power suppliers, landowners with power infrastructure capabilities, and data center operators are now forming strategic alliances to offer complete behind-the-meter power generation solutions. For instance, Volta Grid is deploying natural gas microgrids for clients like Vantage and Oracle, having secured $5 billion in debt and credit financing in November 2025 with a plan to reach a total power capacity of over 4.3 gigawatts by 2028. Meta, on the other hand, has partnered with the natural gas midstream company Williams to build dedicated natural gas power projects for its four data center parks in Ohio, under supply agreements lasting 10 to 12.5 years.

As data center parks continue to expand in scale, their energy consumption and capital requirements skyrocket. The total investment for a current under-construction hyperscale data center can reach $5 billion or more. To secure this capital, developers must prove to investors that they have the professional expertise to handle various operational risks, making the exploration of different partnership models a practical and essential path forward. A reliable power supply will remain the number one operational risk, making collaboration with utilities, independent power producers, and energy permitting specialists crucial for project success.

Looking ahead, some developers view behind-the-meter generation as a transitional solution until they can connect to the public grid, while others see it as a permanent part of their power architecture. Rehlko's Mercer notes, "Initially, everyone treated it as a transition, but the reality has changed. Now many companies are building plants to operate in island mode directly." However, in the long term, data centers with their own power plants could also become deeply integrated with the public grid. Parks equipped with on-site generation, energy storage, and flexible load capabilities could connect to the grid later, enhancing overall grid resilience. For example, Josh Parker, NVIDIA's head of sustainability, said in an interview that the company has developed technology that allows data centers to adjust their power consumption in real-time based on grid conditions.

This brings up a significant policy question: should these new power sources and flexible loads remain isolated and independent, or should they connect to the grid to create value for the entire power system? Grid upgrades that serve large-load projects, such as new transmission lines and substation modifications, can also improve the reliability of power for all users. In the long term, if data centers also act as paying electricity consumers, residential rates could potentially go down. With fixed costs of utilities spread across a larger base of electricity consumption, the portion borne by residential users could be reduced. Behind-the-meter generation, initially a workaround for long connection queues, could evolve into a flexible system that allows for two-way support between data centers and the public grid. But whether this final step will be realized remains uncertain. Thomas concludes, "This is the industry's biggest unknown, and no one can give a definitive answer."