Shiftgrid raises $45 million Series A to align AI training with scarce data center power
Sofia Reyes ·
Shiftgrid, an Austin based startup that schedules AI training across sites based on power availability and cost, said Monday it raised 45 million dollars in Series A funding led by Energize Ventures with participation from Coatue and DCVC. The company plans to expand its grid aware orchestration software as data center power scarcity has become a gating factor for model development in the United States and Europe.
The timing underscores a shift in venture allocation from pure GPU supply toward software that cuts idle time and energy costs. Several regions posted record summer loads and long interconnect queues, which is pushing AI builders to treat megawatts as a primary resource alongside chips.
Why investors are moving into power aware AI operations
Shiftgrid says it has 12 paying customers across cloud providers, model labs, and enterprise GPU fleets, with eight million dollars in annual recurring revenue and 138 percent net revenue retention. The company targets gross margins above 75 percent on a subscription and usage based model, and claims early customers have reduced training costs by double digit percentages by moving jobs to lower cost hours and regions.
The product integrates with Kubernetes, Slurm, and Ray to route workloads using real time and day ahead power prices, carbon intensity, and site level constraints. It also links to corporate power purchase agreements so compute teams can prioritize contracted megawatt hours and avoid curtailment penalties, according to the company.
For Energize Ventures, which has focused on climate software, the round extends a thesis that enterprise decarbonization and cost control will be won in software layers. Coatue and DCVC have both been active in AI infrastructure and deep tech, and are now adding exposure where energy markets and compute operations meet.
Competitive field and near term risks
Incumbents are building adjacent features, including cost based schedulers inside major clouds and internal tooling at large model companies. Startups in workload automation, observability, and carbon accounting are encroaching from nearby categories, which puts a premium on utility data access, reliability at scale, and procurement level integrations that are hard to displace.
Buyer scrutiny remains high. Integration into existing MLOps stacks can stretch sales cycles, and savings claims will be tested as power markets normalize outside peak seasons. The durability of revenue depends on expanding from training to inference scheduling and capacity planning, not just arbitrage on summer spikes.
If software that links compute to kilowatts proves sticky, AI infrastructure spend will start to look more like energy procurement. That would pull more climate focused capital into AI operations and push large platforms to expose deeper power signals, reshaping where defensibility sits in the stack.