This is a week where you have to wonder if there’s room for all the foundation large language models out there.
New sectors of the tech industry tend to begin with many players and land on one or two. There were once many search providers; now there is (primarily) Google. There were many social networks; then Meta (including Instagram) and TikTok. Sometimes infrastructure-based markets support more players — particularly if governments insist on it to avoid monopolies — but there still tends to arise a clear leader.
For the first year of the generative AI revolution, it felt like OpenAI was that company. But Google quickly followed, then Anthropic, Meta’s Llama, Mistral in France, and now even China’s DeepSeek. The race is high-stakes because if history is a guide, being second is far from being first, and third can be distant enough to be considered niche.
In the past week, xAI released its latest update, dubbed Grok 3, and early tests show it to be on par with or ahead of its rivals along many metrics. It was trained on xAI’s Colossus supercomputer, which has 200,000 NVIDIA H100 GPUs, powered by 15 gas turbines — resulting in a great deal of emissions.
The crowded market raises the question of whether a shakeout is coming. Will users coalesce around a single leader, while others lose share and the ability to finance data centers?
This question has real implications for utilities and IPPs as they react to gigawatt-scale requests for power and infrastructure without knowing if these players will be solvent a decade from now. Moody’s Ratings warned last summer of the multiple types of exposure that vertically integrated utilities have as they navigate this sector: the demand-side risk of overbuilding infrastructure for data center demand that either doesn’t materialize or flattens too soon, along with the supply-side risk of promising capacity to data center operators and being unable to meet those obligations in the timeframe that suits the customer. Both present credit risks for the utilities, and as a result Wall Street is watching the space closely.
A recent wave of earnings announcements from U.S. electric utilities shows we’re still very much in the early stages of this infrastructure bubble market.
Indeed, rather than the skepticism about the AI market’s longevity that dominated industry conversations when DeepSeek’s model was released just weeks ago, it was clear from utility leader presentations that they see the demand as real. They’re getting firm orders with financial commitments, not just requests to get into utility queues.
Reading the utility earnings tea leaves
Utilities reported an increase in capital spending to deploy infrastructure and power to data center operators, and they all continue to see long-term prospects for growth. A few highlights:
Dominion Energy’s territory is home to the largest concentration of data centers in the world. And the utility’s leaders described a market still very much accelerating, with 40 GW of data center capacity in various stages of contracting, up a remarkable 19 GW from July 2024.
Of that total, 26.2 GW are at the earliest stage of contracting, known as the substation engineering letter of authorization (SELOA), up from 7.2 GW in July 2024. Those in the final stage, where an electric service agreement (ESA) has been signed, have increased to 9 GW; these include an element of take-or-pay to the contract, so revenue is guaranteed. Dominion also raised its five-year capital plan through 2029 by 16% to roughly $50.1 billion, a reflection of its anticipated need to serve this load growth.
Meanwhile, Duke Energy’s five-year $83 billion capital plan — which is a 12% increase from the former plan — will add upwards of 320,000 miles of power lines and 5 GW of new natural gas generation in service by year-end 2029.
Between 2025 and 2027, Duke will bring 900 MW of solar into service in Florida. Starting in 2027, the utility plans to bring 1.5 GW of solar into service annually in the Carolinas, as well as additional battery storage; and in the mid-2030s, it plans to offer the option of deploying SMR nuclear generation in the region.
Northern California’s PG&E, focused largely on resilience and its avoidance of utility-caused wildfires in its territory for two years. But the utility also framed data centers and load growth in the context of affordability. Their pipeline for 15 data center developers is now over 5.5 GW, an increase of 2 GW since their last forecast in June 2024. The value to ratepayers, they calculate, is a 1% to 2% reduction in electric bills for each GW they provide to data center operators.
AEP, one of the country’s largest power providers, also highlighted the value of serving the large loads associated with data centers. They are seeing demand from data centers in Indiana, Ohio, and Texas, and expect commercial gigawatt-hour sales to grow by 23.9% in 2025, compared to 10.6% in 2024. This represents “customer commitments for approximately 20 GW of load through 2029 driven by data center demand and economic development with 2/3 from the commercial class and 1/3 from the industrial class.”
AEP also expects to see retail load growth of 8% to 9% annually from 2025 to 2027. Their five-year capital plan is $54 billion, up 25% from the previous plan, and includes an incremental $10 billion to meet demand growth.
DTE had a similar story. They increased their capital plan by $5 billion from their earlier plan, which they attributed to load growth, and pointed to their growth data center customers as a means to keep rates lower than average. They pointed to the SUPERNAP data center campus deal with Switch that could scale to 1.4 GW of load through 2032, and another unnamed data center deal that could scale to 2.1 GW over time.
And as for the DeepSeek discourse? Well, utility CEOs and analysts largely set aside the question of whether the more efficient model would reset expectations for energy; it featured only as an aside in most earnings calls.
A version of this story was published in the AI-Energy Nexus newsletter on February 19. Subscribe to get pieces like this — plus expert analysis, original reporting, and curated resources — in your inbox every Wednesday.
