Emerald AI is worth $1.05bn on the theory that grid headroom is a scheduling problem

The $150mn is more than double every dollar Emerald had raised before it ^[25], and it values the software at roughly $900mn before the cheque cleared ^[24]. What that price buys is a claim about sequencing: that the quickest way to more compute is not more copper but a contractual right to consume less at particular moments. Emerald Conductor sorts computing jobs by how much delay each customer will tolerate, and when a utility signals strain it slows, pauses, caps or moves the ones that can wait ^[6]. Most efficiency software goes at the cooling equipment and leaves the computing alone ^[7].

The demonstration record is narrower than the pitch. Across 33 experiments in the Phoenix field test the system managed 212 jobs without breaking a predefined service tier, and its power prediction was off by 4.52% against average experiment power ^[16]. Marcus Schuler, writing for Implicator, set out what that does and does not cover: Emerald personnel and partners wrote the study, it covered one pre-profiled cluster, and what got slowed or paused was batch training and fine-tuning, with real-time inference, streaming and model serving left alone ^[17]. The authors say measuring anything beyond a single cluster needs larger deployments with full-site telemetry ^[18]. The company's headline figure is more than 100 gigawatts of untapped capacity on the existing US grid ^[9].

The reason utilities are listening anyway is arithmetic they cannot argue with. Silicon Valley Power serves roughly 55 data centres across 20 square miles, including Nvidia and Intel ^[12]. Its director Nicolas Procos told The New York Times that every scrap of spare capacity is now spoken for, leaving a choice between building out the system and finding creative solutions ^[13]. Building grid infrastructure can take a decade ^[10], and the International Energy Agency expects data centres to drive nearly half the growth in US electricity demand through 2030 ^[11]. Santa Clara chose the software, and its Flexible Load Interconnection Program trades expanded grid access for verified, dispatchable flexibility ^[14].

Which puts the value of this company in a place the funding announcement does not describe. Silicon Valley Power's chief operating officer Chris Karwick has said full utility control of the load-side breaker "is non-negotiable" in exchange for faster interconnection ^[21]. Operators resist that, because an abrupt shutdown can damage expensive hardware ^[22]. As of late June, Schuler reported, no standardised binding agreement between a utility and a data centre existed ^[23]. The scheduler is the easy part; the enforceable interruption right is the asset, and it has not been written yet.

Europe is where the queue bites hardest, and 63% of new European capacity is now landing outside the big five markets because of connection queues and land ^[20]. RWE and Siemens both put money in ^[4], and Emerald's fifth commercial demonstration ran in London with National Grid among the partners ^[19].