Compute Governance
Observatory
An open observatory for the compute behind frontier AI: what training runs cost in chips, energy, and dollars, and which of them today's regulatory thresholds would catch.
Dashed line: the EU AI Act's 1025 FLOP presumption of systemic risk. Models to its right would be presumed in scope today.
Frontier AI is built from compute, and compute can be counted.
Chips are physical, scarce, and traceable. That makes training compute one of the few levers on AI development that can actually be measured and verified. This project makes the arithmetic behind it explicit, sourced, and checkable.
See how the estimates are validated →From parameters to policy
Five views on one question: how much compute does a model take, and does it matter?
- 01
Compute, energy & cost
Enter parameters, tokens, and hardware to get training compute, wall-clock time, energy, and cost. Every figure is labelled reported, calculated, or estimated.
- 02
Hardware
11 accelerators from V100 through B200, MI300X, and TPU v4/v5e/v5p, with sourced peak throughput, power draw, and representative cloud pricing.
- 03
Validation
The core 6ND formula is checked against 5 public training runs and lands within a few percent of their reported compute.
- 04
Compute landscape
Notable models from OpenAI, Google DeepMind, Meta, Alibaba, DeepSeek, and TII by compute consumed, alongside known clusters by installed capacity.
- 05
Regulatory thresholds
The EU AI Act's 10²⁵ FLOP systemic-risk presumption, live-compared against any estimate. Editable, and extensible with your own thresholds.
Explore the observatory
The estimator
Model a training run from first principles and see what it would take.
Explore →Training clusters
4 known clusters drawn to scale, from accelerators to data halls.
Explore →Thresholds
Which models and clusters a compute threshold would catch, and by how much.
Explore →Checked against real runs
The estimator's compute formula, set beside what labs actually reported.