The Headroom Was Always There
A turbine maker, a grid-equipment giant and a utility funded a company whose product is using less electricity. That cap table is the story of the week.
A turbine maker, a grid-equipment giant and a utility funded a company whose product is using less electricity. That cap table is the story of the week. What happened Three events in seven days, from three different layers of the same stack. 25 August — Emerald AI raised $150 million at a $1.05 billion post-money valuation, bringing total funding above $220 million. The round was co-led by Energize Capital and DCVC. The product makes data centres throttle power draw dynamically in response to grid conditions . 24 August — NVIDIA's Groq 3 LPX entered full production , an inference accelerator extending the Vera Rubin platform. 256 language processing units per liquid-cooled rack. NVIDIA claims up to 35× more inference throughput per megawatt — its own figure, against a baseline it chose. Nebius is the first cloud to adopt it, and it follows NVIDIA's $20 billion Groq deal. 18 August — Pennsylvania signed Executive Order 2026-05 , officially titled "Protecting Pennsylvania Consumers from Data Center Impacts." It removes AI data centres from Fast Track permitting, requires legally binding GRID commitments plus local approval before the environment department will even review an application, and prohibits nondisclosure agreements on data-centre projects. A startup, the dominant chip vendor, and a US state — all responding in the same week to the same fact: generation cannot be built fast enough. Why it matters The binding constraint in AI infrastructure has moved twice. First it was chips . Then it was electricity . It is now becoming electricity at a particular moment , which is a different problem with a different solution. Grids are dimensioned for the hour of highest demand and spend the remaining eight thousand hours of the year operating well below that line. A system with substantial headroom at 3am and none at 6pm can host far more load than its peak capacity suggests — provided the load will move . Emerald's proposition is that AI workloads are unusually well suited to this. A meaningful share of them — training runs, batch inference, embeddings backfills, evaluation sweeps — genuinely do not care what hour they execute in. That is where Emerald's claim of 100+ GW of untapped US grid capacity comes from — a claim about opportunity size, not delivered capacity. Not new generation. Existing headroom that inflexible load cannot use. The cap table is the argument Energize Capital and DCVC co-led the round. But the names worth noticing are further down the list: GE Vernova (turbines), Siemens (grid equipment) and RWE (a utility). These are the companies that profit most from building new generation. They have taken positions in a company whose product reduces the need for it. Read charitably, they see flexibility as complementary — it defers capital expenditure and buys time for construction already underway. Read sceptically, they have run the numbers on turbine delivery schedules and concluded that demand must bend because supply will not arrive