From Megawatts to Tokens: How NVIDIA Maximizes AI Factory Production

NVIDIA’s blog narrates Emerald AI’s first large deployment: Silicon Valley Power signaled an AI factory to cut draw on a hot August evening; Conductor slowed deferrable work while critical jobs continued. Varun Sivaram’s team watched power fall; the utility has since sent more than 200 demand signals—every one successful in the piece’s telling. Power dropped from four megawatts in the dramatized moment.

Tokens-per-megawatt is becoming the KPI. Factories that can’t flex will wait longer for interconnects.

This is the AEMA story with dirt under the nails.

We’re treating a clean 200-for-200 utility story as promising and still vendor-told.

The operational pattern is right: receive grid signal, shed flexible training/batch, protect latency-critical inference. That’s useful AI infrastructure—software that makes denser compute socially tolerable.

The harm if marketing outruns contracts: communities promised flexibility while always-on agent traffic grows inelastic, or “success” defined as brief dips that don’t help evening peaks. Also, cheering a power drop can hide that the campus still consumes enormous baseline energy.

I’m watching published MW-minutes curtailed, payments/penalties with Silicon Valley Power, and whether other utilities replicate the Conductor pattern on non-NVIDIA stacks.

Emerald AI is an NVIDIA partner building grid-orchestration software. The vignette ties to broader DSX Flex / AEMA messaging in NVIDIA’s September 2026 infra push.