Layer 1 · Where the power comes from
Demand response and curtailable load
Operators who can genuinely pause or shift work, which for training is more plausible than it sounds.
The shape of it
- Lead time
- Months
- Capital cost
- Low — it is mostly contractual and operational
- Carbon
- Neutral to positive
- Maturity
- proven
Verified 2026-09-06.
The catch
Requires workloads that tolerate interruption. Inference serving real users does not. Training checkpoints do.
What it actually decides
Increasingly the price of getting connected at all — several utilities now favour flexible loads in the queue. The cheapest megawatt is the one you agree not to draw at peak.
What this is powering
A GB200 NVL72 rack draws 120–132 kW and a GB300 pushes 135–200 kW, against a 2026 average rack of about 27 kW. The silicon decision and the power decision are the same decision, made eighteen months apart.
Layer 2 — Silicon · GB200 NVL72 · The 150 W inference option
Weighed against
- Demand response and curtailable load vs Behind-the-meter generation — Two ways to get connected sooner: agree to use less, or avoid the wires entirely.
Others in where the power comes from
- Grid interconnection — 4-5 years typical, approaching 5 on average
- Nuclear PPA (existing plants) — 1-3 years to contract; the plant already exists
- Small modular reactors (SMR) — 2030s for meaningful capacity
- On-site natural gas turbines — 12-24 months, the fastest firm power available
- Behind-the-meter generation — Depends on the generation, but it skips the queue
- Solar + battery storage — 18-36 months