Decoupling and Coordination: The Keys to Making AI Datacenters Constructive Loads in Renewable-dominated Grids

AI and cloud datacenters (DCs) are rapidly growing constant loads that conflict with variable wind and solar generation, challenging both grid reliability and decarbonization. Flexible grid load to match the variable power supply is the key to solving these grid problems, but hardly any commercial datacenters act as flexible grid loads today because it reduces resource/capital efficiency. We explore decoupling datacenter power capacity and grid load using energy resources (e.g. storage, generator) to create datacenter grid load flexibility. Further, we explore the use of decoupling more broadly to ease reliability constraints (datacenter growth) and accelerate decarbonization.

First, we propose "power Middlebox", a new system architecture that realizes decoupling. We define the Middlebox system architecture, frame its objectives, and explore designs (energy resources, extent of decoupling, management) in varied power grid settings. Evaluation shows that Middlebox unlocks 460% or 170% datacenter growth with grid reliability or decarbonization constraints in a wind-dominated grid. Decoupling reconciles the conflict between grid and datacenter needs, enabling constant DC power capacity on 99.9% of days, for a cost equal to 3--5% increase in datacenter TCO. Future technologies are expected to reduce Middlebox cost. Furthermore, workload flexibility studied extensively by others can be exploited to further reduce cost. These results are robust across grid types. Overall, the results show that Middlebox can be deployed in small to large datacenters economically with today's technology.

Second, addressing that the grid is weak in cooperation, we study how to distribute decoupling across datacenters and cooperatively manage that distributed capability to maximize carbon reduction benefits for all. Evaluation shows that optimized distribution must consider site variation. It can deliver >98% of the benefits enabled by maximum local decoupling needs with 30% less total decoupling need. For management, grid control of DC grid load achieves the highest benefits (10--17% grid carbon reduction vs. no decoupling). If the goal is to maintain some datacenter autonomy, PS-GridScale employing 2-way information sharing and control can achieve 84--87% of the grid carbon reduction with decoupling cost reduced by 7--16% vs. grid control. PS-GridScale also outperforms 1-way information sharing (PlanShare) and traditional selfish online control approaches in both benefit and cost. Decoupling may be economically viable, as on average datacenters gain power cost and carbon reduction benefits that are greater than their local costs of decoupling. Skew in cost across datacenter sites suggests public policy will be required to achieve the most efficient decoupling distributions.

To conclude, employing the ideas of decoupling and DC-grid coordination, this dissertation presents a clear pathway to making datacenters constructive loads in renewable-dominated grids, enabling sustainable growth of AI.