On July 21, 2025, Anthropic published "Build AI in America," a report estimating that the U.S. AI sector will need at least 50 gigawatts of electric capacity by 2028. The report said a single frontier model at Anthropic would need a 2 GW data center in 2027 and a 5 GW one in 2028, and it proposed faster permitting, federal land for data centers and new transmission. It put a concrete number on what had become the defining constraint of AI infrastructure: not chips alone, but power.
Key Facts#
- Published: July 21, 2025, by Anthropic, with a full report available for download.
- Headline estimate: The U.S. AI sector needs at least 50 GW of electric capacity by 2028.
- Training demand: Anthropic expects total frontier AI training demand in the U.S. to reach 20 to 25 GW by 2028, which it described as roughly twice New York City's peak electricity demand.
- Inference demand: Anthropic said at least as much capacity will likely be needed to serve everyday model use.
- Single-model scale: Anthropic said 2 GW and 5 GW data centers would be needed to develop single advanced AI models for the company in 2027 and 2028, respectively.
- Policy pillars: The report proposes building large-scale AI training infrastructure and broad-based infrastructure for AI innovation nationwide.
- China comparison: The report said China added more than 400 GW of power capacity in the previous year, compared with several dozen gigawatts in the U.S.
What Happened#
The report framed energy as the gating factor for AI development in the U.S. Its central estimate combined two demands. Training frontier models, which Anthropic expected to require 20 to 25 GW in total by 2028, and serving those models to users, which it said would likely need at least as much again. Together they add up to the 50 GW floor that became the report's headline.
The policy proposals were organized into two pillars. The first, focused on large-scale training infrastructure, called for making federal lands available for data centers to avoid prolonged local zoning processes, accelerating environmental reviews under the National Environmental Policy Act, partnering with the private sector on building power lines and ensuring timely grid interconnection for projects important to national security.
The second pillar addressed infrastructure for AI innovation across the country. It called for speeding permits for geothermal, natural gas and nuclear generation, designating National Interest Electric Transmission Corridors, strengthening domestic production of critical grid components and expanding workforce training for the trades that build power and data center infrastructure.
The report closed with a call to act quickly, arguing that decisive action would ensure "the future is built here in America."
Background#
The report followed a visible week of energy-focused activity. On July 15, 2025, Anthropic CEO Dario Amodei took part in the inaugural Pennsylvania Energy and Innovation Summit at Carnegie Mellon University. At that event Anthropic announced a $2 million contribution to the university, split between the Scott Institute for Energy Innovation, for research on AI-driven grid management and efficiency, and the picoCTF cybersecurity education program. In that announcement, Anthropic wrote that "energy is central to winning the AI race."
By mid-2025, AI labs were talking about compute in gigawatts rather than GPU counts, because power, not chip supply alone, determined how fast new capacity could come online. Grid interconnection queues, transmission limits and permitting timelines had become as important to AI roadmaps as accelerator shipments. Anthropic's estimate put a public number on the scale of the problem.
Why It Matters for Developers#
Power constraints reach application developers through three channels: price, availability and rate limits. When capacity is scarce, providers ration it through quotas, regional availability and pricing tiers, and developers feel it as throttled requests or slower access to new models. Several practical habits follow for .NET teams building on hosted models.
- Measure tokens per task, not just requests. Energy and cost both scale with tokens processed. Instrument your application so you know which features consume the most, as our LLM observability and cost guide shows with OpenTelemetry.
- Handle throttling gracefully. Capacity limits surface as HTTP 429 responses and timeouts. Retries with backoff, circuit breakers and fallbacks to other models or regions keep features usable. Our resilience guide covers these patterns with Polly.
- Right-size models. Many tasks, such as classification, extraction and routing, do not need a frontier model. Smaller hosted models or local models with ONNX Runtime or Ollama cut both cost and energy use.
- Plan capacity like any other dependency. For production systems, treat model capacity as a constrained resource with budgets, alerts and fallbacks, the same way you would plan database or network capacity. The scalability and cost interview guide discusses the trade-offs.
The broader lesson is that AI infrastructure has physical limits that software cannot abstract away entirely. Efficient applications will be cheaper to run and more resilient when capacity is tight.
What's Next#
The report's numbers shaped how Anthropic described its own plans over the following months. In November 2025, the company announced a $50 billion investment in U.S. AI data centers with Fluidstack, with sites in Texas and New York, and said the investment supported the goals of the administration's AI Action Plan.
In February 2026, Anthropic went further on energy costs. It pledged to fully fund grid upgrades needed to connect its data centers, to secure new generation that matches its data centers' demand, and to deploy curtailment systems that reduce consumption at peak times. "AI companies shouldn't leave American ratepayers to pick up the tab," the company wrote, while stressing that company-level action is not enough without federal permitting and transmission reform.
Whether the country can add capacity at the pace the report called for remains the open question. It depends on permitting reform, transmission construction and supply chains for grid equipment, all of which move more slowly than AI demand.