US Data Center Freeze Threatens to Halve AI Compute Growth by 2028
TREE NEWS reports: On November 4, 2026, the day after the US midterm elections, OpenAI executives met to discuss whether to reset their 2027 data center capacity expectations. No decision was reached, but the anxiety was palpable. Across the country, large data center projects are being halted: Texas Governor Greg Abbott froze grid connections and air permits starting in August, pending an audit through year-end; newly elected Governor Gina Hinojosa has promised to pause all new data center approvals until the legislature passes protective measures—and the legislative session doesn’t end until May 2027. Ohio’s new governor, Amy Acton, has proposed a conditional moratorium requiring new projects to bear full power costs and locate only on former industrial sites. Georgia’s incoming governor, Keisha Lance Bottoms, plans an executive order to pause new construction statewide. Between July and November 4, at least 21 major US data center projects were canceled, nearly matching the 23 cancellations in all of 2025. Anthropic’s October 26 IPO prospectus lists moratoriums as a risk factor, noting that opposition is becoming more organized across jurisdictions.
What This Means for AI Compute
If all planned but unstarted US data center projects were frozen—a scenario that is extreme but not absurd—the US would lose roughly 230 to 340 GW of planned capacity. Current US data center power demand is about 60 GW. Projects under construction (estimated at 57–90 GW) would be completed, bringing total capacity to around 120–150 GW by 2028—about 2 to 2.5 times today’s level, but far below the 5x growth expected if all planned projects proceed. That’s a 50% cut to potential compute growth. In terms of chips, 1 GW powers roughly 270,000 to 550,000 Blackwell-class GPUs, meaning tens of millions of advanced AI chips would be left without a home. The affected states—Texas, Ohio, Georgia, New Mexico, and New York—account for 158.6 GW of planned capacity, or 41% of the national total. So nearly half of the freeze scenario is already backed by policy commitments.
Market Implications
The immediate impact on AI progress would be muted: models slated for 2027 are already training in existing or near-complete facilities. The real damage hits next-generation models planned for 2028–2030, which require multiple gigawatts. Frontier training power demand is growing 2.2–2.9x annually, and by 2030, the largest single training run could need 4–16 GW. With new power frozen, effective compute growth—combining physical compute (4–5x per year) and algorithmic efficiency (about 3x per year)—would drop from 12–15x annually to around 6x. AI would slow, not stop. But the squeeze would force a reallocation of existing power from training to inference, leading to token rationing and a pivot toward efficiency: better performance-per-watt chips (Nvidia’s Vera Rubin claims 10x Blackwell), distributed training across smaller facilities, more investment in post-training and reinforcement learning, and edge computing. For investors, this scenario favors companies with existing data center capacity and energy-efficient hardware. It also narrows the gap between frontier and open-source models, benefiting efficient model developers like DeepSeek, Qwen, and Kimi, whose token share on platforms like OpenRouter has surged from 1.2% in late 2024 to nearly 30% in 2025 and 50-60% in 2026. Companies with large, permitted pipelines—like Meta’s Hyperion in Louisiana—gain a competitive moat. Utilities and grid operators in moratorium states face reduced demand growth, while those in permissive states could see accelerated investment.
Key Takeaways
- US data center moratoriums could cut planned AI compute growth by half by 2028, delaying next-generation models but not stopping progress.
- Effective compute growth would slow from 12–15x to ~6x annually, compressing the lead of frontier labs and giving open-source challengers a better shot.
- Investors should watch for winners in energy-efficient chips, distributed training, and regions with supportive policies; losers include utilities in moratorium states and companies reliant on large, centralized training clusters.
- The token economy will shift toward rationing and efficiency, making performance-per-watt the key metric for AI infrastructure.
