On the morning of November 5, 2026 — the day after U.S. midterm results came in — OpenAI executives gathered for an unscheduled meeting. According to sources cited by foreign media, the discussion centered on whether to revise the company's 2027 data center capacity targets. No decision was reached, but the anxiety was now on the table.

Over the past several months, major data center projects across the United States have been hitting the brakes. Texas's outgoing Governor Abbott froze new data center grid interconnection and air permits starting in August, ordering an audit to run through year-end. Newly elected Governor Gina Hinojosa has pledged to halt all new data center approvals until the state legislature passes resident-protection legislation — a process that won't conclude until the legislative session ends in May 2027. Ohio's new Governor Amy Acton proposed a conditional pause requiring new projects to cover their full power costs and build only on former industrial sites. Georgia's incoming Governor Keisha Lance Bottoms plans to sign an executive order halting all new construction in the state on her first day in office.

According to Tokenpost, at least 21 large data center projects were cancelled across the United States between July and November 4 of this year — nearly matching the 23 cancellations recorded for all of 2025. In its IPO prospectus filed October 26, Anthropic listed the moratoriums as a risk factor, noting that some jurisdictions are considering or have already enacted pauses that could delay or block the computing power infrastructure it depends on, and that opposition to data centers is becoming more organized across jurisdictions.

So what happens if you take this trend to its logical extreme — if every planned data center in the United States simply cannot be built? What does that mean for AI, today and in the years ahead?

What Gets Cut Is Everything Beyond the Initial Doubling

The boundaries of this thought experiment need to be set clearly. The extreme scenario rests on three assumptions: all planned projects nationwide that have not yet broken ground are frozen, extending the current roughly 41% pause coverage to the entire country; all projects already under construction are completed on schedule, consistent with how existing moratoriums in Texas and elsewhere apply only to new approvals; and chip supply remains unaffected, with electricity and land as the only binding constraints. This analysis covers the U.S. only and does not address the possibility of moving computing power offshore.

According to Newmark estimates, U.S. data centers currently consume around 60 GW of electricity — a figure that encompasses all data centers, not just AI. Estimates of capacity under construction vary by source: Cleanview puts it at roughly 57 GW, Wood Mackenzie at around 80 GW, and DCMap's August figures at approximately 90 GW. Under this scenario, all of that capacity would be completed on schedule, representing the entire near-term addition to U.S. computing power.

The truly massive figure sits in the "planned but not yet started" category. DCMap's August count puts this at 228 GW; Wood Mackenzie has disclosed a cumulative pipeline exceeding 430 GW. A range of roughly 230 to 340 GW is a reasonable working estimate.

Based on these figures, a rough calculation suggests that under the extreme scenario, total U.S. data center capacity would plateau around 2028 at approximately 120 to 150 GW — roughly two to 2.5 times today's level — and stop growing after that. For context, Synergy Research had previously forecast U.S. capacity to double by 2029; if the full pipeline were built out, that figure would instead be more than five times today's. What the extreme scenario cuts is everything beyond that initial doubling.

Translating this into chip counts makes the scale more tangible. One gigawatt of power can support hundreds of thousands of Blackwell-class GPUs: on a theoretical per-card all-in power draw of 1.8 to 2.3 kW, that's roughly 430,000 to 550,000 units; at a real-world facility like Abilene, the figure is closer to 270,000 per GW. Freeze two to three hundred gigawatts of planned capacity, and you've effectively left tens of millions of cutting-edge AI chips without a home.

This scenario is extreme, but not implausible. Cleanview estimates that the states where newly elected governors have pledged moratoriums have a combined planned capacity of 158.6 GW — about 41% of the national total — with Texas accounting for 88.4 GW, Ohio 21.9 GW, Georgia 20.2 GW, New Mexico 11 GW, and New York 9.9 GW. In other words, nearly half of the freezes in this extreme scenario already have corresponding policy commitments in the real world. Of the 69 large projects cancelled since 2024, 33 were driven primarily by local opposition. Senator Sanders has also introduced a national moratorium bill, though its prospects for passage remain slim.

One caveat deserves honest acknowledgment: the planning figures are heavily inflated. The Texas grid operator ERCOT has a large-load interconnection queue of roughly 474 GW across about 1,800 projects — approximately five times the grid's peak demand — most of which are speculative placeholder applications. So the effective computing power actually cut by the extreme scenario would be smaller than the headline numbers suggest. Even stripped of that padding, it would still represent a sudden hard brake on supply.

Training Runs Through 2027 Are Fine — the Wall Hits in 2028

If the extreme scenario materialized tomorrow, users would likely notice nothing for a year or two. Frontier models through 2027 would continue to train and iterate on schedule.

That's because the training infrastructure supporting this generation of models is already built or nearly so. According to Epoch AI's database, the largest known AI campus is Anthropic and Amazon's facility in New Carlisle, Indiana, at 1,037 MW; second is xAI's Colossus 2 at 946 MW. OpenAI's campus in Abilene, Texas, is planned for 1.2 GW — six of its eight buildings had been delivered as of September, with full completion expected by year-end.

What gets stuck is the next generation of models — those slated to arrive between 2028 and 2030, requiring multiple gigawatts of power just to train.

Epoch AI data shows that the electricity required for frontier training runs is growing at 2.2 to 2.9 times per year, with current frontier runs already exceeding 100 MW. Epoch expects multiple campuses to break the 1 GW threshold within the next year. Roughly two years out, the largest training runs may require the equivalent of around 2.5 million H100-class chips; following that trajectory, by 2030 a single top-tier training run could require 4 to 16 GW.

For comparison, the largest campus today has just crossed 1 GW. Epoch has also noted that data center capacity is effectively the ceiling on training scale: a model can only be trained as large as the campus that houses it.

And many of the mega-projects aimed at that 2028-and-beyond window have already hit walls.

OpenAI's Stargate program targets more than 9 GW by 2029, with most of its seven U.S. sites originally slated to come online at scale in 2027 and 2028 — concentrated in Texas, New Mexico, and Ohio, all moratorium states. The planned expansion adjacent to Abilene has already been abandoned. In New Mexico, Project Jupiter — a 2.45 GW development in Doña Ana County — was rejected in March by the state land office over its gas pipeline application. At the Ohio site in Lordstown, satellite imagery from April still showed no major construction activity.

Meta's roughly 1 GW Prometheus campus is also in Ohio. Only its Louisiana project, the 5 GW Hyperion facility, has avoided this round of moratoriums.

Effective Compute Growth Falls from the Teens to Around Sixfold

With the computing power base for next-generation training stranded, how much does AI progress actually slow? Answering that requires breaking down the sources of compute growth.

According to Epoch AI, since 2020 the training compute for frontier language models has grown roughly four to five times per year. Over the same period, training electricity consumption has grown 2.2 to 2.9 times per year. The roughly twofold gap between these two figures is largely attributable to improvements in chip efficiency — more computing power extracted from each watt of electricity.

Beyond hardware, algorithmic improvements have also contributed. A commonly cited figure is roughly three times improvement in algorithmic efficiency per year; research by Ho et al. in 2024 put it at around 2.7 times. When Epoch later incorporated post-training into its calculations, the estimate rose to roughly ten times per year — though with a range from 2x to 50x. This analysis uses the more conservative 3x figure.

A rough calculation follows from these figures. Under normal conditions, the "effective compute" available to AI each year grows by roughly 12 to 15 times — the product of physical compute growth (four to five times) multiplied by algorithmic improvement (around three times).

In the extreme scenario, new electricity comes to a halt after 2028. The most powerful driver of physical compute growth disappears, leaving only the efficiency gains from successive chip generations — roughly a twofold improvement per year. Multiply that by the 3x algorithmic gain, and effective compute annual growth falls from 12 to 15 times down to roughly 6 times.

Using Epoch's more optimistic algorithmic estimate of around 10 times per year, baseline effective compute would normally grow 40 to 50 times annually; in the extreme scenario that falls to roughly 20 times. Under both assumptions, the electricity freeze cuts effective compute growth to roughly 40 to 50 percent of what it would otherwise be. Algorithmic progress, however fast it runs, cannot fill the gap left by frozen electricity capacity.

There is an additional risk embedded in this. The efficiency gains from new chips assume those chips can actually be installed in functioning data centers. Once total facility power is capped, the difficulty of upgrading power and cooling systems may erode even that twofold efficiency dividend.

Even so, roughly sixfold annual growth remains a substantial number. The conclusion from this exercise is straightforward: in the extreme scenario, AI slows meaningfully — but it does not stop.

This accounting only covers training compute, however. Zoom out to the full operation of AI systems, and a second competition would begin the moment electricity is capped.

Every Megawatt Becomes a Choice Between Training and Inference

In the past, when compute was scarce, you built more data centers. In the extreme scenario, the total power envelope is fixed — and every watt directed toward training has to come from inference, agents, and post-training.

The industry's compute mix has already been shifting toward inference. Deloitte projects that inference's share of total AI compute was roughly one-third in 2023, reached about half in 2025, and is expected to hit roughly two-thirds in 2026.

The internal allocations at AI labs are shifting too. SemiAnalysis estimates based on the combined compute of OpenAI and Anthropic show that pre-training accounted for 67% of compute in early 2024, while post-training, reinforcement learning, and inference represented about 5% and 29% respectively. Since then, pre-training's share has declined each quarter — 63% in Q2, 57% in Q3, and down to just over 50% in 2025 — with post-training rising steadily. The latest estimates suggest the two are now roughly equal. Researchers have noted that this is a share shift: total lab compute has multiplied many times over, so the absolute amount going to pre-training has not necessarily fallen.

The more acute pressure comes from product design. Reasoning models generate lengthy internal chains of thought before responding — invisible to the user, but consuming electricity at every step. Gartner data shows that an AI agent completing a task consumes five to 30 times as many tokens as a standard chat interaction. Goldman Sachs forecasts that global token consumption will grow 24 times between 2026 and 2030.

Feed those numbers into a fixed power budget, and every lab faces the same zero-sum question: each megawatt either goes toward training the next, larger foundation model or toward serving current users and agents. Routing hundreds of thousands of chips into a bigger pre-training run means rationing the inference capacity that existing users depend on. Prioritizing inference means falling behind on the next generation.

With the option to simply add more power off the table, the question becomes: what paths remain for AI to continue advancing under a fixed electricity and compute ceiling?

Four Routes to Doing More With Less Power

Under the pressure of an extreme scenario, efficiency technologies that have long been dismissed because "you can always add more cards" would become mainstream solutions.

The first is a change in the performance metric — shifting hardware evaluation from raw specifications to tokens-per-watt. In a facility with a fixed power budget, swapping older chips for more efficient newer ones is the only remaining form of capacity expansion.

NVIDIA's preferred benchmark is now "tokens per second per megawatt of total facility power" — a denominator that includes everything: cooling, networking, the whole campus. The fact that this metric has become standard tells you something: chipmakers have already internalized that the customer's bottleneck is electricity, not card count. According to NVIDIA, Vera Rubin delivers 10 times the per-watt performance of Blackwell, and up to 30 times that of the GB300 in agentic workloads. Independent tests have produced similar figures: in CoreWeave's benchmarks running the same DeepSeek R1 workload at equivalent throughput speeds, the Vera Rubin NVL72 produces 10 times the tokens per megawatt of the GB200 NVL72. Rubin doesn't enter volume production until the second half of the year, however, and most customers won't have widespread access until 2027.

The second route is distributed training across multiple data centers. Moratoriums have consistently targeted large facilities requiring hundreds of megawatts or more. A natural workaround is splitting training jobs across smaller, geographically dispersed facilities in different jurisdictions. Policy pressure would translate directly into engineering demand for distributed training.

The major players are already experimenting. Google has described training Gemini across synchronized clusters in multiple locations. Microsoft has linked data centers in Wisconsin and Georgia. CoreWeave and Google Cloud have announced cross-cloud training. Google's DiLoCo research demonstrates that training can run across remote facilities connected by standard internet links, while tolerating hardware failures.

The technical challenges are real, though. DiLoCo's own authors acknowledge the technique can only supplement, not replace, very large single-site clusters. A 2026 simulation study found that when facilities are tens to hundreds of kilometers apart, network latency significantly degrades the overlap of computation and communication, leaving substantial compute idle while waiting for data.

The third route is shifting R&D emphasis from pre-training toward post-training and reinforcement learning. Pre-training is most dependent on a single massive cluster and is therefore the first casualty in a constrained environment. If multi-gigawatt campuses can't be built, engineers will redirect effort toward post-training, reinforcement learning, and architectural improvements — mixture-of-experts models, reduced-precision computation — that are less demanding of cluster scale. Cursor has stated publicly that its Composer 1.5 model consumed more compute in post-training than in pre-training.

The fourth route is edge deployment and small models offloading demand from the cloud. By 2026, distillation has become the dominant method for building smaller models. Quantized small models can run structured tasks directly on smartphone NPUs; models below one billion parameters can generate roughly 12 to 15 tokens per second. A growing number of applications are adopting a hybrid architecture: a local small model handles most requests, with only the most complex queries passed to a cloud-based large model.

Who Benefits Inside These Constraints

The extreme scenario caps electricity on U.S. soil. For two other categories of players, that constraint opens a window: model companies that have spent years on the efficiency path, and AI startups preparing to enter the market.

Consider the gap between frontier labs and those trying to close it. According to Epoch AI's estimates from earlier this year, the strongest open-source models lag closed frontier models by an average of about four months — roughly the gap between GPT-5 and GPT-5.5. In the past, frontier labs used annual effective compute growth in the low double digits to continuously widen that lead. In the extreme scenario, that pace drops to around sixfold, making each new capability gap harder to establish and the distance challengers need to close shorter.

Then there are the players already ahead on the efficiency curve. Chinese model companies — DeepSeek, Qwen, Kimi, GLM, MiniMax, and others — have built their reputations on open-source releases, low cost, and mixture-of-experts architectures. That approach is now showing up in adoption data.

According to joint research from OpenRouter and a16z, Chinese open-source models' share of token consumption on the OpenRouter platform rose from roughly 1.2% at the end of 2024 to nearly 30% in 2025. Into 2026, multiple tracking sources put that share between 50 and 60 percent. a16z partner Martin Casado has also estimated that roughly 80% of startups using open-source model stacks are running Chinese models.

It's worth noting that OpenRouter is a single platform that excludes ChatGPT's own traffic and enterprise private deployments. But the direction is clear: when the entire industry has to answer the question of how much intelligence it can extract from each kilowatt-hour, the teams that have been working on that problem the longest will find their experience worth considerably more.