Resources
The AI Landscape
AI from 2022 to 2032, on one timeline. Layers shows which parts of the AI stack, from materials to applications, held it back each year. Power flow follows electricity from the grid to matrix multiplication. Cost flow follows the budget to useful math and explains the gap with a roofline chart. World energy shows the global power build-out and how much of it goes to data centers and AI.
2026
ProjectionA stacked diagram of seven layers, Materials at the bottom and Applications at the top. Each layer lists its components as buttons; each component is marked as a bottleneck, tight, or not binding for the selected year and scenario. Select a component to read why, or a layer title for an overview.
- Bottleneck
- Tight
- Not binding
- Projected (hatched)
From the grid to the chip
Where does the power go?
Follow the highlighted power into the next bar. Each bar expands its parent to full width; every watt stays relative to 100 W from the grid.
The arithmetic is only part of the story
View as table
Loading power estimates…
From the budget to the math
Where does the money go?
Energy is a small slice of what an accelerator costs. What matters is how much of each dollar buys work the chip can actually do.
What the budget actually buys
View cost data as a table
Why the gap exists
The roofline: where each workload sits
H100-class roof: 3.35 TB/s memory bandwidth, 989 TFLOP/s dense BF16 compute.
Dashed roof — Illustrative inference-optimized design: same compute, 4× memory bandwidth.
Training lives on the flat roof, where a GPU is near-optimal. Decode lives on the slope, where performance is set by bytes per second, not FLOPs. Inference ASICs exist to move the slope.
View roofline data as a table
Assumptions: H100-class accelerator at ~$25–40k installed, 5-year hardware amortization, ~1.2 kW all-in per accelerator at $0.08/kWh, ~$10M per MW facility build, inference fleet utilization ~55%. Estimates, not vendor data.
Loading cost estimates…
Installed generating capacity
Gigawatts by source. Click a column to select its year.
Datacenter electricity use
Terawatt-hours a year; the selected year shows its share of world demand.
Show the data as a table
* Projected
Historical and current values reflect public reporting; projections are illustrative scenarios, not forecasts.