Summary of Key Points
PJM, the largest electricity market operator in the eastern United States, is struggling to cope with the explosive demand for electricity from AI data centers. Outdated market mechanisms and flawed models have led to soaring electricity prices (with household bills increasing by 17%-24%), while new capacity has not kept up with the demand. Additionally, PJM’s governance is trapped in a “voting deadlock,” preventing any meaningful reform. Tech companies such as Microsoft and Google have chosen to build their own power generation facilities to avoid costly and unreliable connections to the public grid, resulting in ordinary households bearing the brunt of the costs. This situation reflects the deep-seated conflict between America’s aging infrastructure and the growing demands of the AI industry.
1. AI Data Centers: A Demanding Presence
The training of large AI models requires numerous GPUs, which consume more power than traditional servers—often exceeding 100 kilowatts per rack. Companies like Microsoft and Google are aggressively acquiring land in the United States to build data centers that are energy-intensive. Over the past 20 years, electricity demand in the eastern region covered by PJM has remained stable, and the grid was optimized for “zero growth.” Suddenly, the massive demand from AI has overwhelmed the infrastructure, which is ill-equipped to handle such a surge.
2. The PJM Capacity Market: How Did Electricity Prices Rise Nearly Tenfold?
The “capacity market” can be understood as a form of “backup power insurance.” Electricity companies not only need to sell electricity but also ensure they have sufficient generating capacity during peak times. PJM auctions this “insurance” annually, and the cost is passed on to consumers in their bills. In 2024, PJM upgraded its model due to the 2022 storms, lowering the reliability rating of natural gas generators by 10%-20%. This led to an apparent reduction in the required backup capacity, but the new model underestimated the actual needs. As a result, auction prices soared from $28.92 per megawatt-day to $270-333 per megawatt-day, resulting in total costs of $63.6 billion—more than an order of magnitude higher than before. Ironically, most of this money went to existing power generators, which paid only 1/20 to 1/40 of the auction price, effectively receiving subsidies.
3. Two Critical Errors in the Model: A Waste of Hundreds of Millions
PJM’s calculation model contained two fatal flaws:
1. The Ignoring of Cold Weather: Cold air increases the efficiency of gas generators, leading to higher power output in winter. The model used summer data, which underestimated the power generation capacity and resulted in the purchase of additional backup capacity worth $13 billion.
2. Outdated Reinforcement Data: After the 2022 storms, power plants were reinforced for winter, but the model still used outdated data on outages. This led to the purchase of an extra 3.8 GW of backup capacity, costing users another $12 billion. These errors caused PJM to overestimate the power shortfall and forced consumers to pay extra.
4. Voting Deadlock: Why Can’t PJM Reform?
PJM’s rules dictate that any reform can be blocked if two of the five stakeholders (power generators, consumers, etc.) oppose it. For example, a winter capacity plan that could save $2.7-8 billion was rejected by power generators, who feared being fined for poor performance in summer. Other reforms have also been thwarted by stakeholders. The FERC (regulatory agency) has issued a final ultimatum, but lacks the authority to initiate necessary changes, leaving the reform process at a standstill. This outdated mechanism from 20 years ago is completely inadequate for the needs of the AI era.
5. Data Centers “Going Off-Grid”: Ordinary Households Bear the Burden
Due to PJM’s slow progress in connecting new data centers (a four-and-a-half-year pause in related research) and the uncertainty of costs, tech companies are opting to leave the public grid:
- Microsoft has signed a 20-year contract to restart a nuclear power plant, while Google has built its own natural gas supply system.
- With the loss of these stable, large users, the public grid’s fixed costs (such as transmission lines and backup capacity) fall on ordinary households.
- PJM is planning to auction 6.8 GW of additional capacity, but the actual shortfall is only 3 GW. Since the cost allocation is uncertain, if large users do not contribute, the burden will once again fall on households.
Consequences: The eastern United States is at a disadvantage in the AI competitiveness race due to unstable power supply, with ordinary households becoming the scapegoats for the system’s failures. This crisis highlights the issue of who will bear the costs of AI’s growth: capable large users have left, leaving ordinary people to bear the consequences. America’s outdated power system is holding back the development of the AI industry.