Grid-only power is not enough: energy-system integration as the option
Power for AI data centers is becoming difficult to secure by simply "connecting to the grid." EPRI's "Powering Intelligence 2026" shows that U.S. data centers could account for 9-17% of national electricity consumption by 2030, while warning that grid-only dependence could mean up to 10 years before energization in some regions (EPRI Executive Summary). Demand rises in a few years, but generation, transmission, and substations take many years to build. That timing gap is pushing operators toward energy-system integration, combining their own power, storage, and operating flexibility.
This article separates the main measures, onsite generation, stationary battery storage (BESS), and demand response, into "direction" and "current adoption level."
Onsite generation: not waiting for the grid queue
The most direct measure is onsite generation at the facility. Procurement is already moving: Baker Hughes announced that it will supply 1.21GW of natural-gas turbine generators for Crusoe through Boom Supersonic, with firm orders totaling 1.3GW for AI data-center power (Baker Hughes). On the supply mix, EPRI states that under current policy natural gas is likely to dominate near- and medium-term incremental supply, with 2025-2030 additions of 6.6-13.7GW per year, above the recent five-year average of 5.7GW per year (EPRI). Fuel-cell vendor Bloom Energy also says, based on its own decision-maker survey, that onsite power will expand, but this should be discounted as vendor survey and intent data rather than a neutral market forecast (Bloom Energy).
Onsite generation is not a cure-all. Fuel procurement, emissions rules, environmental permitting, and maintenance capacity introduce separate costs and constraints. It can bypass grid waiting time, but site-specific regulation and fuel conditions decide whether it works.
Stationary BESS: absorbing peaks and grid constraints
Stationary battery storage (BESS) absorbs the steep load swings of AI workloads and smooths peaks, easing interconnection conditions. Battery deployment is accelerating across the U.S.; EIA expects 18.2GW of new utility-scale battery storage in 2025, up from 10.3GW in 2024, and says solar and batteries will account for 81% of 63GW of new generating capacity (EIA).
Data-center-driven demand is also starting to be seen as a growth driver. Industry-data-based reporting has argued that data centers will lift behind-the-meter battery demand (secondary reference: Axios), but we treat those analyst forecasts as directional evidence rather than figures to take at face value. In practice, primary data showing colocated capacity or contracted volumes for individual data centers remains scarce. The appropriate view today is that the direction is clear, while scale still awaits primary-data coverage.
Onsite generation
Use gas turbines or fuel cells to avoid waiting for the grid. Baker Hughes has firm AI data-center power orders totaling 1.3GW. Fuel, emissions, and permitting constraints remain.
Stationary BESS
Absorb sharp load swings and reduce peaks to ease interconnection. U.S. new battery storage is expected to reach 18.2GW in 2025 (EIA).
Demand response and flexible operation
Limit or shift compute loads to lower peaks. Data centers begin to act as grid assets themselves.
Grid-side technology such as SST
800VDC peripherals are in implementation, but solid-state transformers (SST) for data centers remain at the research and design-proposal stage.
Demand response and flexible operation: turning data centers into grid assets
Measures are also advancing on the demand side, not only on the supply side. Google announced that agreements with multiple U.S. utilities include a combined 1GW of data-center demand-response capacity. The mechanism curtails or shifts some machine-learning workloads to reduce power demand during specified periods (Google). Google is also using external flexibility, including an agreement with Voltus to aggregate up to 100MW of distributed resources in the PJM region (Google).
On the research side, EPRI's "DCFlex" initiative is testing whether real-time flexible data-center operation can support grid reliability and faster interconnection, with more than 60 companies participating (EPRI DCFlex). The direction is clear: data centers are moving from "loads that only consume power" toward "adjustable assets that support the grid." Market design, including which wholesale power-market products should apply, remains a future issue.
SST and 800VDC peripherals: what is practical today?
Technology refresh is also discussed on the grid-equipment side, but the "direction" and "adoption level" need to be separated carefully. ABB says high-density AI workloads are pushing conventional AC distribution close to physical limits, describes 800VDC distribution as suitable, and presents MVAC-to-LVDC converters and the IEC-certified SACE Infinitus solid-state breaker as implementation elements (ABB).
By contrast, solid-state transformers (SST) themselves, as substitutes for long-lead-time iron-core transformers, remain at the research and design-proposal stage for data centers. Academic work proposes and evaluates SST configurations converting 10kV AC to an 800V DC bus, but we have not confirmed commercial deployment or proof that SST has avoided transformer lead times. It is a promising direction, not a planning assumption yet.
What each role should check next
Energy-system integration turns the strategy for securing electricity itself into a design target, beyond optimizing a power topology.
Site development
Which configuration should be assumed: grid only, onsite generation, storage, or a mix? Fuel, emissions, and permitting for onsite generation affect site decisions.
Procurement
Check lead times and multiple sources for generation equipment, BESS, and 800V-compatible devices. Long-lead components must not become the project bottleneck.
Operations and grid coordination
Decide how far to use demand response and flexible operation. Contracts with wholesale markets and grid operators become a new competitive variable.
Technology planning
Separate research-stage and implementation-stage grid technologies such as SST. Pulling them forward too aggressively creates planning risk.
Data-center power is shifting from a contest over "how many facilities to build" to one over "how reliably power can be secured." How operators combine the grid, onsite generation, storage, and demand response will divide data-center siting and investment decisions in the AI era.