Emerald AI Series A raises $150 million for power-flexible data centers

This funding signals a shift in how AI infrastructure interacts with the energy grid, moving from passive consumption to active load management. For streaming and AI providers, this technology offers a path to bypass the decade-long wait for new grid infrastructure by utilizing up to 100 gigawatts of untapped capacity on existing lines. As data centers are projected to drive half of U.S. electricity demand growth through 2030, software-defined power flexibility becomes a critical operational requirement for scaling high-density compute. Watch for the commercial launch of the Vera Rubin AI Research Factory later this year to validate if these power-flexible 'AI factories' can maintain uptime during extreme grid stress.

Emerald AI enters a rapidly expanding market for software-defined power management in AI data centers. The company's Emerald Conductor platform, which dynamically adjusts compute workloads and onsite energy resources during grid stress events, competes with a growing field of startups and incumbents targeting the same bottleneck. In August 2026, Intel presented three silicon platforms at Hot Chips 2026 specifically designed for agentic AI workloads across data center, inference, and edge tiers, including the Crescent Island GPU with up to 480 GB of LPDDR5X memory in a 350-watt air-cooled PCIe form factor. That thermal envelope design philosophy mirrors Emerald AI's core thesis: maximizing compute density within existing power and cooling constraints rather than waiting for new grid capacity.

The business case for power-flexible infrastructure is being validated by enterprise adoption data. According to the Salesforce 2026 Connectivity Benchmark, the average enterprise now runs 12 AI agents, with multi-agent adoption projected to surge 67% by 2027, creating compounding demand for data center capacity that existing grid connections cannot easily serve. Emerald AI's backers, including Energize Capital and DCVC, are betting that software orchestration of power consumption can unlock stranded capacity on existing transmission lines faster than building new generation or transmission assets. The company's $1.05 billion valuation reflects investor confidence that utilities and grid operators like National Grid will increasingly contract with data centers for demand-response services, turning power flexibility into a revenue stream rather than merely a cost constraint.

Technical research is beginning to quantify the workload characteristics that make power-flexible orchestration both necessary and complex. A new benchmark suite called AgentSysBench, published in August 2026, found that in five of ten representative agentic applications, non-LLM components such as tools and environments dominate latency, with sandbox working-set memory peaking at 28 GB per session. The study also demonstrated that task-aware serving reduces latency by 29 to 40 percent and state offloading cuts memory usage by 4.6 times, suggesting that intelligent workload scheduling of the kind Emerald Conductor performs can yield substantial efficiency gains. Separately, researchers have applied causal reasoning agents to O-RAN network forensic triage, achieving 83 percent accuracy on delay anomaly classification through graph soft-prompting techniques, a methodology that could extend to power-grid anomaly detection in data center orchestration systems. These findings underscore that as AI workloads grow more heterogeneous and stateful, the software layer managing their interaction with physical infrastructure becomes a critical differentiator for operators like Emerald AI.