A $1.05B valuation resting on one 256-GPU test and a breaker nobody has agreed to hand over

What Emerald sells a utility is a forecast: hold this much capacity for me now, and I will give this much back when you ask. That makes the least quotable figure in the Phoenix study the load-bearing one. Across 33 experiments and 212 jobs, the system recorded a 4.52% power-prediction error against average experiment power ^[6], which is roughly a fifth of the reduction it was demonstrating ^[17]. The authors, who are Emerald personnel and partners, wrote that larger deployments with full-site telemetry are needed to measure effects beyond a single cluster ^[7]. That sentence is carrying more weight than the headline result.

The tier structure sets the ceiling on what can be sold. Emerald's tiers permitted throughput reductions of 0% to 50% over three to six hours ^[9], and the Phoenix event sat at the middle of the depth range and the short end of the duration range ^[19]. It also ran about sixteen months before the round ^[18], against batch training, fine-tuning and some inference, with real-time inference, streaming and model-serving untouched ^[8].

Then there is who holds the switch. Under the Santa Clara plan, the utility sends requests during limited periods of grid need, Emerald converts them into approved workload changes, and Emerald reports the response ^[12]. Silicon Valley Power's chief operating officer, Chris Karwick, has said full utility control of the load-side breaker is non-negotiable ^[14]. Those are not the same architecture, and the distance between them is the whole negotiation. Emerald's stated reason an operator would concede any of it is speed: dispatchable AI work should win faster grid connections without leaving operators unable to serve customers ^[1].

The money has a similar gap between the stated and the filed. The August 3, 2026 Form D showed $90,229,639 sold toward a planned $150 million offering, with no valuation or lead investors named ^[4]. That is about 60% of the target, leaving roughly $59.8 million of the offering unsold at that date ^[16].

What is actually novel is the target of the control rather than the idea of flexibility. etalytics and comparable tools change how facility cooling equipment runs without changing compute workloads; Emerald reaches into the job queue ^[15]. Varun Sivaram, a former Biden administration energy official, founded the company in November 2024 ^[2]. The software is the tractable part. The asset being valued at $1.05 billion ^[20] is a signed answer to how often a utility may slow a paying customer's training run, and Karwick's breaker line indicates the two sides are not yet answering it the same way ^[14].