In the rapidly evolving world of artificial intelligence, we’re in the midst of a transition: from a period when developers of large language models concentrated compute resources on training these massive models, to one where computing is shifting to how these models think to a mode of inference and reasoning.
You can see the shift in the steady stream of announcements of “agents” that can take multi-step actions to complete tasks rather than simply answer queries, and of research-oriented tools that most often use “chain of thought” approaches to solve complex tasks, mimicking the work of an analyst.
These approaches use vastly more energy than the already energy-intense chatbots and send a strong signal that load growth forecasts will continue to trend up even in the face of efficiency gains in training.
This mixture of intense competition with China — made more intense by last week’s release of DeepSeek — as well as urgency to deploy data centers at scale, and widespread frustration with the length of interconnection queues continues to fuel a move toward on-site generation.
For instance, GE Vernova is fielding major orders for gas turbines and even announced a partnership to build new generation with renewables developer NextEra; Exxon and Chevron are entering the market with explicit plans for on-site generation; and a number of SMR nuclear companies have said their order books are filling up from data center operators rather than traditional utilities or IPPs.
On-site generation isn’t new to the data center market, but so far it hasn’t been common, for good reason. It’s a large, up-front capital expenditure, either for the data center owner if they are financing it, or for the utility if they are operating as an “energy-as-a-service” provider. And supporting data loads this large is new to everyone. The entire ecosystem around AI data centers and power is at the beginning of an experience curve.
Most on-site generation today is found in heavy industry and large corporate or college campuses. It’s often in the form of combined heat and power (CHP), where a gas generation facility provides power as well as heating and cooling. These tend to be fairly small, well under 10 megawatts, and have long been thought of as a niche market. For existing data centers, on-site generation has typically come in the form of a diesel generator, used for backup power only.
AI-scale data centers are a different beast altogether. Where CHP offers most of its value in efficiency and control; data centers are looking for scale, speed to market, low-carbon energy, and reliability. The value of heat, for now, is an afterthought.
Models for on-site generation
These deployments can come in a variety of models, and the willingness of hyperscalers to pay a premium for speed and reliability is spurring innovative approaches across the market.
For instance, there’s an increase in utility-driven partnership, where the utility provides the on-site generation for the data center operator in advance of building out grid infrastructure to the site. AEP, for example, recently agreed to use up to 1 GW of Bloom Energy fuel cells to power data centers until the necessary transmission grid infrastructure is provided.
Meanwhile, oil and gas majors Exxon and Chevron are leveraging their gas power plant expertise and access to supply to build large-scale power plants on-site, with a plan to lower emissions through integration of CCS over time. They will own and operate these plants and sell power to the data center in what appears to be a behind-the-meter arrangement, devoting all their generation to the data center and avoiding interconnection with the grid — and regulation by FERC — altogether.
Similarly, the small nuclear company Oklo will develop, own, and operate power plants for its data center customers and sell them power via long-term contracts. And microgrid developer Scale Microgrids has modeled developing solar with natural gas backup microgrids for data centers that can provide dedicated power at prices competitive with gas.
Some hyperscalers are building their own generation, but are also drawing from the grid. Project Stargate’s first data center, dubbed Project Ludicrous, will begin with 360 MW of natural gas on-site power from GE Vernova and Solar Turbines, and follow with upwards of 15 GW of additional generation over time as the data center project scales across multiple sites. They will also draw power from the grid where possible.
This combination of activity makes it clear we’re in a period where capex is king. Not only will hyperscalers require massive spending to build the digital infrastructure to power AI, they will also often find themselves building energy infrastructure as well.
We’re in a market phase now where speed is prioritized, so gas generation is often at the head of the line — getting most of the press, and the emissions anxiety. But as the market matures, price and efficiency tend to take over. Gas will be exposed to fuel price volatility and risk, while renewable and storage combinations continue to benefit from cost declines and zero fuel price. Nuclear remains an unknown, with capex and opex still moving targets, but with likely support from the Trump administration.
In general, we expect to see more zero-carbon experimentation and solutions, including in on-site power, as the hyperscalers try to limit their emissions growth. With Meta, Google, Amazon, and Microsoft indicating plans to spend between $60 billion and $90 billion each on infrastructure capex in 2025, there is ample room to innovate. The on-site option is clearly gaining traction.
A version of this story was published in the AI-Energy Nexus newsletter on January 22. Subscribe to get pieces like this — plus expert analysis, original reporting, and curated resources — in your inbox every Wednesday.
