For conventional grid-tied projects, developers know their interconnect agreements and offtake terms, and storage specifications follow from those fixed parameters
The System Pressure
The power behavior of AI compute differs from every prior data center load type. GPU clusters ramp from light draw to near-full capacity and then collapse back down in a pattern that repeats continuously across training and inference jobs. That cycle operates faster than gas engines and grid regulation mechanisms can respond. The consequence is not inefficiency alone — it is physical stress on generation equipment and, for grid-tied facilities without buffering, a disturbance that propagates outward into shared infrastructure.
According to Wärtsilä, which is actively positioning itself as a power systems integrator for large AI campuses, the only workable solution is interposing battery energy storage between the AI load and whatever generation source supplies it. Storage absorbs the ramp event and presents a stable draw to the generation asset, making the fluctuation invisible to the grid and protecting equipment from repeated mechanical stress. The physics of rapid load cycling support that framing regardless of its commercial source.
The Drivers, Dependencies, and Constraints
Three converging forces are making BESS integration structurally necessary rather than merely preferable.
First, interconnection timelines are narrowing viable options. With grid queues running three to seven years in major markets, many AI data center developers are moving to islanded or off-grid configurations with dedicated on-site generation. Those configurations remove the grid as a natural buffer, meaning storage must perform that function entirely. Wärtsilä’s position is unambiguous: operating an AI data center on islanded power without BESS is effectively impossible, because gas engines alone cannot track the load profile fast enough without risking mechanical damage.
Second, BESS sizing for AI facilities is not a solved calculation. For conventional grid-tied projects, developers know their interconnect agreements and offtake terms, and storage specifications follow from those fixed parameters. AI data center developers arrive without a confirmed load target. Wärtsilä’s starting framework — match battery power capacity one-to-one with the expected load swing, begin with a two-hour duration — gives a first-principles anchor. A 600MW swing implies a 600MW storage system as the opening assumption. Even so, the firm is explicit that full system modeling is required before any final specification can be locked; generation mix, grid connection terms, and the actual load profile must all be resolved first.
Third, the load profile is not reliably known in advance. Chip providers including NVIDIA do not publish the power demand curves that their accelerators produce under real workloads. Those profiles can shift across chip generations, and changes in training methodology can alter consumption patterns at a facility that has already been built and specified. A BESS sized against one generation of GPU cluster may fall outside specification when the compute stack turns over, leaving the power system operating beyond its design envelope with no straightforward correction path.
The Texas project Wärtsilä announced in April — a 790MW off-grid plant using 42 natural gas engines — illustrates the sequencing pressure. Generation equipment and large power transformers carry the longest procurement lead times. Batteries, at roughly 18 months, can follow. But facilities that defer BESS procurement past the initial build phase are accumulating operational risk that may not surface until first compute load arrives.
Open Dependencies
Several material gaps prevent a clean system read at this stage.
Gigawatt-scale islanded grid operation has no prior operational precedent. Wärtsilä acknowledges that islanded systems in its existing experience top out around 100MW — an order of magnitude below what some AI campus designs now require. Control architecture, maintenance scheduling, and state-of-charge management at that scale introduce complexity that has not been tested in a live AI compute environment. Coordinating generation dispatch, storage cycling, and planned outages across a system with no grid fallback is a qualitatively different engineering challenge from anything currently in service.
The vendor source constraint also deserves explicit acknowledgment. The framing of BESS as non-negotiable comes from a company that manufactures and sells energy storage systems. The operational conclusions are directionally sound given the physics involved, but independent validation from grid operators, ISOs, or engineering research institutions would strengthen both the sizing methodology and the risk characterization before it becomes a procurement standard.
Finally, the regulatory and commercial implications of large islanded AI facilities remain underexplored. A facility operating at 790MW or beyond with no grid interconnection avoids queue delays but also loses frequency support, grid backup, and potential demand response revenue. Those tradeoffs are not yet part of a settled industry framework.
The Operating Exposure for Global Heads of Data Center Energy
The most immediate budget and planning exposure is the BESS line item arriving later in the sequence than it should. If generation equipment procurement drives the schedule and storage is treated as a follow-on decision, BESS specifications will be locked before the actual load profile is validated against real compute behavior. That sequencing error is difficult to correct after construction — inverter ratings, transformer capacity, and physical site layout are not easily revised around a storage system that needs to scale.
The load profile opacity adds a second layer of procurement risk. Teams negotiating long-term power supply agreements or specifying storage capacity are working against a demand variable that the chip manufacturer controls and does not disclose. When the compute stack refreshes, the power infrastructure assumption changes with it. Contractual BESS performance warranties are unlikely to account for that scenario unless explicitly negotiated.
For facilities pursuing islanded configurations, the control system complexity at gigawatt scale is an unpriced engineering risk. It does not transfer directly from smaller islanded systems or conventional grid-tied operations. Budgets that treat this as a standard integration effort are likely underprovisioned.
Signals the System Is Shifting
Three developments would confirm this dynamic is hardening from best practice into industry standard. First, BESS appearing as a contractual requirement in AI campus power supply agreements — rather than as a discretionary add-on — would signal that grid operators, insurers, or offtakers are formalizing what is currently described as operational necessity. Second, chip manufacturers publishing load profile data for power infrastructure planning purposes would materially change the accuracy of storage sizing and reduce the retrofit risk described above. Third, a gigawatt-scale islanded AI facility completing its first operational year would produce the first real performance dataset for control systems at that scale — and either validate or force revision of current engineering assumptions.
Until those markers appear, BESS for AI data centers is a structurally sound requirement being executed against incomplete information. That gap is the live risk.
Sources
- Energy-storage — Wärtsilä on why AI data centres need BESS to smooth demand and avoid ‘destroying the grid’ (Link)
