AI Accelerators & Data Centers
AI data centers grow from megawatts to gigawatts: about 30 GW of AI capacity worldwide, with single sites above 1 GW under construction.
Open in the interactive tree →Specialized chips (GPUs, Google’s TPUs, in-house designs) with high-bandwidth memory are wired into racks drawing over 100 kW and cooled with liquid. According to Epoch AI, training compute for notable models has doubled roughly every six months, and power is the constraint likely to bind first. Electricity, grid connection and cooling have become as important as the chips themselves.
As of October 2026
The Stanford AI Index 2026 reports about 29.6 GW of AI data-center capacity worldwide. Per Epoch AI, OpenAI’s Stargate sites in the US total over 9 GW planned across seven sites (0.3 GW running at Abilene, Texas), most due by Q4 2028. Nvidia reported $96.2 billion revenue for the quarter ending July 2026, of which $89.0 billion from data centers (+117%), with the Vera Rubin platform ramping into full production. The IEA puts data centers at about 415 TWh in 2024 (1.5% of world electricity) and expects around 945 TWh by 2030.
Open steps
- Memory bandwidth per watt Medium AI leverageAI chips wait on memory; taller HBM stacks, new interfaces and in-package memory must raise bandwidth per watt faster than models grow.
- AI-designed kernels and chip layouts High AI leverageHand-tuning GPU kernels and chip floorplans takes experts months; automated search could raise the real utilisation of every accelerator.
- Energy per generated answer High AI leverageServing dominates AI energy use; sparser models, quantisation, speculative decoding and better batching must cut energy per answer without losing quality.
- Faster grid-connection studies Medium AI leverageConnecting a gigawatt site waits years in interconnection queues; automated power-flow studies and load forecasts could shorten the engineering part.
Where AI could help
Low AI leverage. Power, grid links, turbines, memory chips and permits set the pace; AI trims a few percent from kernels, chips and cooling.
- Auto-tune kernels and schedulers for a few percent more work per watt
- Learned control of cooling and workload placement
- Speed up grid-connection studies and siting
- Search chip layouts and memory systems for lower energy
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 July 2016 DeepMind said (own claim) its learned controller cut the energy used for cooling in Google data centers by 40%. source
- In April 2025 PJM, North America's largest grid operator, began a multi-year partnership with Google's Tapestry to bring its interconnection databases and study tools into one AI-supported model (PJM said gains were hard to quantify). source
Prerequisites
- DRAM main memory1966High-bandwidth memory beside AI chips is stacked DRAM
- Open Source & Linux1991AI data centres run on Linux
- Graphics Processor (GPU)1999
- Cloud Computing2006Hyperscale cloud data centres are the base of AI computing
- Deep Learning2012
- Neural Scaling Laws2020Scaling laws justify building ever larger training clusters
- 2 nm Chips & High-NA EUV2025