FERC : Washington forces electric utilities' hand to connect AI data centers within 60 days

FERC : Washington forces the hand of power grid operators to plug in AI data centers within 60 days

🔎 Silicon is no longer enough, AI is running out of outlets

American artificial intelligence has just hit a wall that no GPU can cross. On June 18, 2026, the FERC (Federal Energy Regulatory Commission) issued six targeted orders that upend the rules of electrical interconnection in the United States.

The message is clear: grid connection queues, which sometimes stretch over five to seven years, have become a national security threat. Washington is no longer letting regional operators manage this bottleneck at their own pace.

The stakes go beyond the energy sector. The global race for compute is now playing out on high-voltage power lines. China, Europe, and the Middle East are accelerating their own AI infrastructures. In the US, mega-data center projects are ready, the GPUs ordered, but they remain idle for lack of being able to plug in.

This FERC decision marks a turning point: for the first time, a federal regulator is treating the interconnection of AI data centers as an absolute priority, with a 60-day deadline imposed on six regional operators to justify or rewrite their rules.

The key points

• The FERC issued 6 orders on June 18, 2026, requiring 6 regional operators (PJM, MISO, SPP, CAISO, NYISO, ISO-NE) to justify their interconnection rules within 60 days or rewrite them.
• AI data centers will have to pay the full cost of grid interconnection — no subsidization by other consumers.
• Connection queues reach 5 to 7 years in some regions, blocking hundreds of GW of compute capacity.
• This decision follows a realization shared by both the Biden and then Trump administrations: the electrical bottleneck has become the main geopolitical barrier to US dominance in AI.

Recommended tools

What the FERC exactly ordered on June 18, 2026

The FERC did not adopt a single rule. It issued six distinct orders, each targeting a specific regional transmission operator (RTO). The approach is surgical: each RTO has different interconnection rules, and the FERC is treating them on a case-by-case basis.

The six targeted operators are PJM Interconnection (Eastern US), Midcontinent Independent System Operator (MISO), Southwest Power Pool (SPP), California Independent System Operator (CAISO), New York Independent System Operator (NYISO), and ISO New England (ISO-NE).

Each order follows the same logic. The FERC is asking the operator to demonstrate, within 60 days, that its interconnection queue rules are in the public interest as defined by the Federal Power Act. If not, the operator must propose a revision.

This "show cause" mechanism is rare. It means that the FERC has already deemed the current rules problematic. The burden of proof is reversed: it is not up to the Commission to prove that the rules are bad, but rather up to the operator to prove that they are good.

According to TechCrunch, this targeted approach avoids the years of proceedings that a general federal rule would have required. It also forces each RTO to confront its own dysfunctions rather than hiding behind collective lobbying.

Why 60 days: the urgency of a paralyzed system

Interconnection queues in the United States have become an industry scandal. According to data compiled by the Lawrence Berkeley National Laboratory (2025), the national queue exceeds 2,600 GW of pending projects, including a growing proportion of AI data centers.

In the PJM region, which covers the Midwest and the Mid-Atlantic, the average connection timeline exceeds five years. Some projects filed in 2021 are still not connected in June 2026. For an AI data center whose useful technological lifespan is 3 to 4 years before GPU obsolescence, waiting five years to connect is an economic non-starter.

The American Action Forum analyzes in its breakdown that these orders fit into a logic of regulatory efficiency. The imposed 60 days do not aim to connect a data center in two months — the physical construction of lines takes much longer. They aim to eliminate the bureaucratic delays that are added to technical timelines.

The real issue is getting projects out of the administrative rut. Currently, an operator can demand sequential network impact studies, multiple cost revisions, and legal back-and-forths that add two to three years of purely regulatory delays. FERC wants this to stop.

This dynamic is reminiscent of what has been observed in New York, where the legislature just sent 7 AI bills to the governor, including a data center moratorium, showing that states are also trying to regain control over AI infrastructure expansion.

The new geopolitical bottleneck: from silicon to high voltage

For years, the dominant discourse on the AI race has focused on chips. US sanctions against China's access to NVIDIA GPUs, the dependence on TSMC for manufacturing, the shortages of H100s and then Blackwells: everything was viewed through the prism of silicon.

That time is over. As Ron Schmelzer points out in Forbes, the bottleneck has shifted from the chip to the wall outlet. Models like OpenAI's GPT-5.5 (agentic score of 98.2) or Anthropic's Claude Opus 4.7 (94.3) require clusters of tens of thousands of GPUs, each consuming 700 to 1,200 watts.

A 1 GW mega-data center is the equivalent of the consumption of a city of one million inhabitants. Current projects in the pipeline often exceed 2 to 3 GW. However, the US power grid was not designed to absorb point loads of this magnitude.

The geopolitical dimension is central. China is building its AI data centers with state planning that bypasses regulatory constraints. The Middle East (Saudi Arabia, UAE) offers land, sovereign funding, and fast-tracked procedures. If the United States does not unblock its grid, American companies will go build their compute elsewhere.

It is exactly this type of infrastructure control logic that has driven certain players to massively automate their operations with AI to remain competitive in the face of the growing scarcity of available compute resources.

The 6 targeted operators: who they are and why them

The selection of the six RTOs is not random. Each presents specific problems that FERC wants to address without delay.

PJM: the paralyzed giant

PJM manages the largest electricity market in the world, covering 13 states plus the District of Columbia. It is also the region with the highest concentration of AI data center projects, particularly in Northern Virginia ("Data Center Alley").

PJM's queue exceeds 300 GW. Its "cluster study" rules require projects to be grouped together for grid impact studies, which creates dependencies between projects and slows down the entire process. This is the most critical case.

MISO and SPP: the Midwest under pressure

MISO (Midcontinent) and SPP (Great Plains) are seeing an influx of data center projects attracted by low land prices and access to wind energy. But their grids are designed to transport energy from wind farms to cities, not to absorb massive localized loads.

CAISO: California between virtue and reality

The California system is the most constrained in the country. Carbon neutrality goals coexist poorly with the exponential demand from data centers. CAISO's interconnection rules are among the most complex, with environmental mitigation requirements that considerably lengthen timelines.

NYISO and ISO-NE: the Northeast corridor

New York and New England combine aging grids, high demand density, and rigorous environmental assessment procedures. The New York context is particularly tense as the state recently passed laws strictly regulating AI, creating a paradox between regulation and development.

The "polluter pays" principle applied to data centers

A crucial point of the FERC orders: data centers will pay. The Commission insists that accelerating timelines does not mean socializing costs.

In practice, when an AI data center requests an interconnection that requires transmission grid upgrades (new lines, substations, capacity reinforcement), the applicant bears the full cost. This principle already exists in FERC rules (the "participant-funded" method), but the June 18 orders reaffirm it with unusual clarity.

Why this insistence? Because RTO operators and local utilities have two symmetrical concerns. On the one hand, traditional residential and industrial consumers do not want to pay for lines intended for data centers. On the other hand, data center developers fear that RTOs will pass on shared network costs that do not directly concern them.

FERC rules: each project finances its own interconnection needs. No cross-subsidization. This protects both individuals and developers, but it also means that the total cost of an AI mega-data center skyrockets. A 1 GW project could see its interconnection bill reach $500 million to $1 billion, depending on the distance from the existing grid.

For startups and open source projects trying to democratize access to AI, such as initiatives similar to the MiroFish project which proved that decentralized AI was possible, these interconnection costs represent a new barrier to entry.

What this concretely changes for AI projects

For hyperscalers (Google, Microsoft, Meta, Amazon)

It's a clear victory. Tech giants have the financial backing to fund interconnection costs. What they didn't have was time. Every year of delay on a 100,000 GPU cluster represents hundreds of millions of dollars in lost revenue and a setback in the model race.

With these FERC orders, Google can accelerate the deployment behind Gemini 3 Pro Deep Think (agentic score of 95.4), and Microsoft can push GPT-5.5 without waiting three years for a transmission line. The effect is multiplicative: more performant models require more compute, which requires more electricity, which requires faster grid connection.

For new entrants and AI startups

The equation is more nuanced. Accelerated timelines are beneficial, but the interconnection cost remains a major hurdle. A model developer like Kimi K2.6 (Moonshot AI, agentic score of 88.1) or Z.AI's GLM-5 (82) doesn't have the cash reserves of a hyperscaler to absorb a billion in network costs.

FERC seems aware of this risk. Its orders do not require full upfront payment: progressive financing mechanisms (as construction progresses) remain possible. But the signal is clear: the American electrical grid is not a free public good for AI.

For local utilities

It's an operational headache. FERC orders create enormous pressure on local power companies, which must execute connections faster while managing the technical complexity of multi-GW connections. Many already lack the engineers and equipment (transformers, high-voltage cables) for which order lead times reach 18 to 24 months.

Operators' response: between compliance and resistance

At this stage, the public reactions from the six RTOs are measured. None have openly challenged the orders, which would be politically risky. But behind the scenes, debates are intense.

PJM, in particular, argues that its queues are not solely due to its bureaucracy. A significant share of the delays stems from data center developers' inability to provide the technical information required for network impact studies. According to this reading, accelerating regulatory timelines solves nothing if the applicants themselves are not ready.

FERC anticipates this argument. Its orders include "readiness" requirements: developers will have to prove they have preliminary studies, financing, and construction permits in order before benefiting from the fast-track process. The idea is to prevent the queue from filling up with phantom projects that block others.

Some observers, cited by the American Action Forum, note that the orders could paradoxically lengthen queues in the short term. By lowering the bureaucratic barrier, FERC risks triggering a wave of new applications that will be added to existing projects. The net effect on actual connection timelines remains uncertain.

The parallel with other energy-intensive sectors

AI is not the first sector to hit the wall of electrical interconnection. The crypto-mining industry faced the exact same problem in 2021-2023, with Bitcoin Ethereum projects waiting years to connect.

The difference in treatment is revealing. Mining never received preferential treatment from FERC. AI, on the other hand, is explicitly classified as critical infrastructure. This double standard reflects the political perception: mining is seen as speculation, AI as a strategic investment.

The green hydrogen industry, battery plants, and semiconductor foundries (like CHIPS Act projects) are also facing interconnection delays. But none of these sectors combines such explosive demand growth with such high geopolitical sensitivity as AI.

FERC, with these orders, is creating a precedent that could extend further. If AI gets a fast track, why not battery plants or chip factories? The door is open to a sector-based prioritization of access to the electrical grid, a notion that would have been unthinkable five years ago.

The limits of the FERC decision

That said, these orders are not a magic wand. Several constraints remain fully intact.

The first is physical. Even with perfect interconnection rules, the lines still need to be built. In the United States, building a high-voltage transmission line takes an average of 10 to 12 years, according to the Department of Energy (2024). Federal permitting, environmental studies, eminent domain, local litigation: the FERC controls interconnection rules, not the actual line construction process.

The second limit is generation availability. Connecting a 2 GW data center to the grid is useless if there isn't 2 GW of available generation capacity. However, the closure of coal plants and the insufficient pace of renewables are creating a generation deficit in several regions. Data centers will have to build their own sources (SMR nuclear, solar + storage), which adds years to the timeline.

The third limit is the workforce. The US electrical industry is facing a critical shortage of skilled workers. High-voltage specialized electricians, protection and control engineers, substation technicians: all are in short supply. Accelerating procedures does not create skills out of thin air.

These limits explain why some players are betting on decentralized solutions rather than connecting to the traditional grid. Automating businesses with AI makes it possible to reduce dependence on centralized compute, as illustrated by the approach described in this complete automation experiment in 7 days.

Comparison of interconnection timelines by region (June 2026)

Source: compilation from TechCrunch, American Action Forum, and Forbes data (June 2026). Pending GW are estimates because RTOs do not always publish a breakdown by load type.

The political context: why now

The timing of these orders is not neutral. June 2026 marks a political tipping point on AI in the United States.

The administration has made AI supremacy a national security priority, through executive orders and public declarations. The Department of Energy published a report in May 2026 warning of the risk that power queues could cause the United States to lose its lead in AI compute by 2028-2030.

Congress, for its part, has multiplied hearings on the subject. Senators from both parties have publicly pressured FERC to act. The Commission, traditionally cautious and technical, was therefore pushed toward political action.

It is in this context of federal pressure that state legislative action also falls. In New York, the sending of 7 AI laws to the governor illustrates the tension between regulation and development. One of these laws imposes a moratorium on new data centers, in direct contradiction with the federal objective of acceleration.

By acting at the federal level, FERC partially bypasses these state-level blockages. The Federal Power Act gives the Commission authority over interstate electricity commerce, which preempts local regulations that would slow down interconnection.

AI Models and Consumption: The Frightening Figures

To understand why FERC is taking action, one must look at the consumption figures associated with current models. The growth is exponential.

OpenAI's GPT-5.5, the leader in the agentic leaderboard with 98.2, requires an order of magnitude more compute than GPT-4. Each training cycle of this scale consumes tens of GWh. In inference, a model of this size serving millions of simultaneous users consumes the equivalent of a small city.

The table below illustrates the correlation between model performance and estimated energy requirements:

Note: consumption per request are estimates based on published architectures and energy efficiency reports from NVIDIA H100/B200 clusters (2025-2026).

The trend is undeniable: the higher the models climb in performance, the more their consumption per request increases. Efficiency optimizations (quantization, sparse attention, distillation) do not compensate for the increase in model size. The power grid is therefore facing a demand curve it has never seen before.

❌ Common mistakes

Mistake 1: Confusing regulatory deadlines and construction timelines

Many commentators have interpreted the 60 days as the timeframe to physically connect a data center. This is wrong. The 60 days only apply to the RTOs' response regarding their interconnection rules. Building the lines will still take years. FERC is eliminating bureaucratic delays, not physical ones.

Mistake 2: Thinking that data centers don't pay for the grid

The idea that FERC is giving data centers a "free pass" is widespread but incorrect. The June 18 orders explicitly reaffirm the "participant-funded" principle: the developer pays the entirety of the interconnection costs that apply to them. The acceleration is procedural, not financial.

Mistake 3: Believing that FERC solves the generation problem

Connecting a data center to the transmission grid does not create electricity. If the region does not have sufficient generation capacity, the data center will remain dependent on clean energy sources that it will have to build itself (solar PPA, battery, nuclear SMR). FERC is addressing the transmission problem, not the generation one.

Mistake 4: Ignoring the risk of windfall gains

By accelerating procedures, FERC could encourage speculative projects: developers filing interconnection applications without any real intention of building, simply to reserve a spot in the queue or to speculate on the value of the authorization. The orders include safeguards, but their effectiveness remains to be proven.

❓ Frequently Asked Questions

Can FERC force a state to accept a data center?

No. FERC regulates the interstate transmission grid, not local land use. A state or municipality can still refuse a building permit. But if the data center has its local permits, FERC ensures that the electrical interconnection is not the bottleneck.

Will residential consumers pay for AI data centers?

No, this is explicitly excluded by the orders. The "participant-funded" principle means that the data center finances its own interconnection costs. Residential rates are not directly affected, although increased overall demand could influence energy market prices in the long term.

What is the connection between these FERC orders and New York's AI laws?

There is a direct tension. New York has adopted a moratorium on data centers, while FERC is accelerating their interconnection. In practice, federal law (Federal Power Act) prevails over state law regarding the interstate commerce of electricity, but the legal conflict could materialize.

Do the FERC orders apply to modest-sized projects?

The orders target "large load interconnections," generally above 20 MW. A small 5 MW data center is not directly concerned. But the systemic effect (freeing up capacity in the queues) could indirectly benefit smaller projects.

What happens if an RTO does not comply within 60 days?

FERC can impose rules by direct order, initiate sanction procedures, or as a last resort, revoke the RTO's operator license. In practice, no RTO will take this risk. Compliance will be negotiated up to the last day, but it will happen.

✅ Conclusion

The FERC has just officially acknowledged what the industry has been whispering for two years: the AI war is won on high-voltage power lines, not just in data centers. The six orders of June 18, 2026, do not solve everything — the physical construction of lines, the labor shortage, and the generation deficit remain massive walls. But they send a clear political signal: the electrical bottleneck is now being treated as a national security issue, and operators who drag their feet will be bypassed. The 60 days are ticking.