AGI's hidden bottleneck: power grid can't handle 100GW demand

The drive toward artificial general intelligence rests on an implicit assumption: that we can scale compute indefinitely. Scaling laws have held for a decade, with compute doubling roughly every four months. But the physical world imposes a hard ceiling. A single NVIDIA GB200 rack draws 120 kilowatts. Scaling to one million such accelerators would require approximately two gigawatts of continuous power. At a mature AGI layer—some estimates suggest 100GW—that equals Japan’s entire electricity consumption. This is not a future problem; it is an present-day constraint. In 2019, training a single large NLP model emitted carbon equivalent to five cars over their lifetimes. By 2023, data center power constraints had already become a critical operational issue in Virginia, the world’s largest data center market, where new builds face multi-year grid connection delays.

The major players are all grappling with this physics barrier. AMD’s MI300X draws around 700 watts per unit, offering better efficiency than NVIDIA’s previous generation but still clustering power demands that strain regional grids. Cerebras’s Wafer-Scale Engine, the WSE-3, consumes roughly 15 kilowatts per system and boasts superior compute density per watt, yet scaling to millions of accelerators would require multi-gigawatt infrastructure. Google’s TPU v5p sets a benchmark for efficiency within its own datacenter pods, but Google’s own power draws are already testing grid limits. Hyperscalers like Microsoft have allocated over $50 billion for AI data centers by 2025, with power purchase agreements now a core part of planning. Even so, the grid connection queue in many regions stretches years.

The obvious counterargument is that efficiency gains will save us. History shows power usage effectiveness and chip efficiency improve by roughly 20% per generation. But compute growth has tracked approximately 4x per year—a race that efficiency is losing. The alarmist 100GW figure may itself be an overestimate; AGI could be achieved with far fewer parameters via sparsity, pruning, or distillation. Novel compute paradigms such as optical or analog computing could slash power by orders of magnitude. Superconducting logic operating near zero resistance is under exploration. And the grid itself can be expanded with dedicated small modular nuclear reactors or solar/wind farms, which several tech companies are already planning. Yet none of these are proven at scale. The hidden risk is not that we hit 100GW, but that the infrastructure bottleneck caps growth well before that. For investors, NVIDIA’s soaring stock faces a ceiling if hyperscalers hit limits and pivot to energy-efficient alternatives from startups like Groq or Graphcore. The global AI chip market, projected to reach $200 billion by 2030, may see a shift in winners based on power efficiency.

Ultimately, AGI is not just an AI problem—it is an energy infrastructure problem. The companies that will lead are not only those with the best algorithms but those that secure dedicated power sources or radically break the energy-compute curve. The grid engineer’s analysis, while provocative, serves as a necessary corrective to the techno-optimism that assumes scaling laws can ignore physics. The bottom line: AGI’s timeline will be determined as much by power plant construction as by model architectures.