Episode Summary
Executive Summary: The episode examines how software can make AI data centers flexible enough to act like grid assets instead of grid liabilities. Host David Roberts and Emerald AI founder Varun Sivaram discuss Conductor’s orchestration of compute, batteries, and on-site resources to reduce peak demand, speed interconnection, and improve grid reliability. The conversation also addresses enforceability, technical failure modes, and the climate implications of flexible AI load.
Main Topics: Data center flexibility as grid value: The core idea is that data centers can reduce or shift demand during peak periods so they do not force new grid buildout, potentially allowing much more load to connect to the existing system. Emerald AI’s Conductor platform: Sivaram explains that Conductor coordinates temporal, spatial, and resource flexibility—slowing jobs, moving workloads, and using batteries/generators—to create a controllable AI load. Reliability, liability, and enforceability: The episode digs into who bears responsibility if flexibility is misdeclared or fails, and argues that flexibility must be contractually and technically enforceable for utilities to trust it. Technical failure modes and scaling: Roberts presses on telemetry delays, unreachable nodes, congestion, and scaling from small demos to multi-megawatt facilities, while Sivaram says these are solvable engineering problems discovered through demos. Economic value and incentives: The main value for data centers is faster and larger power connections, while existing data centers may also gain revenue from demand response and flexibility markets. Climate, affordability, and grid planning: Sivaram argues flexible AI load can help absorb renewable energy, avoid some gas buildout, and reduce rate pressure by using existing infrastructure more efficiently. Future of AI as a ‘meta-grid’: Sivaram projects a future where AI workload shifting across locations and substations becomes a major part of electricity system operation, complementing the physical grid.
Key Arguments: Flexible data centers can avoid raising grid peaks, reducing the need for new generation and transmission upgrades. Load flexibility can be achieved through three complementary levers: temporal flexibility, spatial flexibility, and resource flexibility. Conductor acts as the orchestration layer that coordinates GPU scheduling, job movement, and onsite batteries/microgrids. The technology is only useful if it is production-grade, measurable, and contractually enforceable; otherwise utilities will not rely on it. Data centers can provide much larger value through faster interconnection than through demand response payments alone. Existing technical problems—telemetry delays, unreachable nodes, congestion—are real but manageable with abstraction, buffers, and fault tolerance. Flexible AI loads could lower costs for other customers by making better use of existing grid assets and avoiding unnecessary peaks. As AI becomes a larger share of electricity demand, flexibility may become a standard feature of new AI facilities rather than an edge-case response. Flexible data centers should help renewables by matching load to variable generation, potentially reducing the need for gas peakers and curtailment.
Data Points: Potential new load on existing grid: 100 gigawatts - Cited from a 2025 Duke Nicholas Institute study if new loads curtail demand for just 0.5% of uptime. Curtailment window: 0.5% of uptime - Condition under which the Duke study found major new load could fit without new generation. Expected connection delay avoided: 12 years - Sivaram says flexible data centers may get earlier grid connections instead of waiting over a decade. Example connection increase: 50 MW to 75 MW - Used to illustrate how an existing site might expand with flexibility. Demo scale at Hillsboro: 250 kW - Small demonstration cited as a test case for Emerald’s software. Demo scale at NVIDIA/PG&E site: Less than 1 MW - Sivaram says the Hillsboro demo used Grace Blackwell 300s at under a megawatt scale. Commercial deployment scale: Multi-megawatt - Sivaram says Emerald has moved beyond demos into real deployments. Future test center scale: 96 MW - Virarubin test center with NVIDIA and Digital Realty expected to come online later in the year. Target facility scale: 100 MW - Sivaram says the Virarubin project will be the world’s first 100-MW scale power-flexible facility. Grid event duration tested: 24 hours - Emerald says it tested whether a local substation outage could be handled for a full day. Telemetry interval problem: 15 minutes - One demo had data arriving only every 15 minutes, complicating control decisions. Performance tracking events: 200+ grid events - Sivaram says published work showed precise performance across more than 200 grid events. Dispatch price threshold: $2,000 per MWh - ERCOT example of willingness-to-curtail being represented as an economic number. Market monitor critique: “regulatory fiction” - PJM market monitor’s phrase describing data center flexibility claims as unenforceable in current form.
Pivotal Quotes: "I believe that makes them energy technologies." — Varun Sivaram: Sivaram argues that controllable data centers should be treated as grid assets, not just loads. "If we're going to make this work, if data centers are going to get faster and larger grid connections in return for promising flexibility, that promise needs to be legally enforceable and technically enforceable." — Varun Sivaram: On the question of whether flexibility can be trusted by utilities and regulators. "In the long run, AI should really be flexing a lot more often." — Varun Sivaram: Describes a future where flexibility is routine, not just an emergency measure.
Implications: If this works at scale, AI data centers could connect faster, support renewables, reduce grid upgrades, and lower rate pressure. But the industry will need stronger measurement, contracts, and operational proof before utilities trust flexibility as standard practice.