Eclipse and Energize lead $45 million for GridKite to match AI compute with grid power
Sofia Reyes ·
GridKite, a San Francisco startup that schedules AI training and inference to the lowest cost and lowest carbon power windows, raised 45 million dollars in Series B financing on Monday led by Eclipse and Energize Capital. The company said it will use the funds to expand utility integrations and deepen support for Kubernetes, Slurm, and major cloud spot markets as power constraints tighten across North America and Europe.
The timing matters because investors are recalibrating toward software that arbitrages power price volatility and interconnection delays, not just GPU scarcity. Data center electricity has become a gating variable for AI deployment, and buyers want measurable reductions in energy cost per token and per training step.
Why investors care now
GridKite’s platform ingests locational marginal prices, grid congestion alerts, carbon intensity data, and site level capacity limits, then shifts or shapes jobs across on premises clusters and clouds. Management said live deployments across three regions are averaging 18 to 32 percent energy cost savings with no loss in model quality, supported by automated checkpointing and multi region failover.
Eclipse brings deep industrial and hardware networks, while Energize has portfolio ties to utility and grid software buyers. A smaller participation from a corporate venture arm tied to a major chip vendor, which the company did not name, signals alignment with scheduler and accelerator roadmaps.
The company sells a usage based subscription layered on top of customer infrastructure, with optional revenue share tied to realized power savings. Early customers include two North American hyperscalers and a European research lab under multi year agreements, according to the company. Gross retention above 95 percent and expanding workloads suggest the software embeds into MLOps stacks once deployed.
Competition spans demand response aggregators and carbon accounting tools, which do not control workloads, and cluster schedulers like Run AI or Ray based platforms that do not integrate granular grid signals. Crusoe’s flare gas compute and colocation providers offer alternatives at the infrastructure layer, but they require workload migration rather than policy driven orchestration across existing fleets.
Market context
IEA and utility filings point to data center electricity demand that could double by decade end, with interconnection queues stretching past three years in some regions. Recent curtailments, local moratoria, and rising peak prices have forced AI teams to treat power as a first class constraint, moving budget focus from more GPUs to smarter scheduling and siting.
For the broader tech landscape, the deal reinforces that AI scale depends on energy aware software as much as silicon. Startups that make power an explicit variable in how and where models run will influence data center design, cloud pricing, and the next wave of AI infrastructure standards.