The Computing Energy Wall
Demand for computing grows faster than chips get efficient: scaling runs into energy, heat and grid limits.
Open in the interactive tree →Since Dennard scaling ended in the mid-2000s, power density has risen, and the doubling time of transistor density has stretched from about two to about three years. Moving data now costs more energy than computing itself. AI training grows much faster than chip efficiency improves.
As of October 2026
Data centers used about 415 TWh in 2024 (1.5% of world electricity), and the IEA expects around 945 TWh by 2030. According to Epoch AI, power is the constraint likely to bind first: a frontier training run of 2e29 FLOP by 2030 would need about 6 GW, while OpenAI’s Stargate alone plans over 9 GW across seven US sites. The Landauer limit (about 3e-21 J per erased bit at room temperature) lies many orders of magnitude below what computers spend per operation, so physics is not yet the limit, but grid connections and cooling are.
What is missing
- Low-energy memory and data movement (computing in or near memory)
- New power sources and faster grid connections (often years of waiting today)
- More efficient algorithms and models (sparsity, smaller models, quantization)
- Cooling and waste-heat reuse at gigawatt scale
Becomes possible once solved
- AI scaling without a grid bottleneck
- Powerful AI on phones and robots without the cloud
- Data centers that do not need their own power plants
Open steps
- Smaller, sparser, cheaper models High AI leverageCut energy per useful answer through sparsity, quantization, distillation and architectures that keep quality while using less compute.
- Kernel and scheduler tuning High AI leverageGet more useful work per watt from existing accelerators and clusters through tuned kernels, compilers and job placement.
- Computing in or near memory Medium AI leverageMove less data by computing inside memory arrays; prototypes exist but need accuracy, endurance and software support.
- Energy-aware chip design Medium AI leverageLower energy per operation through layout, circuit and architecture search across the chip stack.
- Cooling and heat reuse at GW scale Low AI leverageRemove and reuse heat from gigawatt campuses with less water and electricity; liquid cooling at that scale is new.
Where AI could help
Medium AI leverage. AI helps with chip layouts, kernels and model efficiency, but gigawatt grid connections, cooling and new power plants are physical build-outs.
- Auto-tune GPU kernels and schedulers (AlphaEvolve-style) for a few percent more work per watt
- Search chip layouts and circuit designs for lower energy per operation
- Find more efficient model architectures, quantization and distillation recipes
- Optimize data-center cooling and workload placement
Shown so far
- In May 2025 Google said (own claim) that a scheduling heuristic found by AlphaEvolve recovers about 0.7% of its worldwide compute, and a 23% faster kernel cut Gemini training time by about 1%. source
- In September 2024 Google DeepMind said AlphaChip reinforcement-learning layouts were used in its last three TPU generations (according to the company; the original method is disputed). source
Prerequisites
Unlocks
- Post-Silicon Computers2040s?