The Grid Ends at the Last Recoverable Device

Big Tech’s announced AI data-center plans may require 44GW of additional capacity by 2028. Some developers can face waits of up to seven years for a grid connection. The electricity can exist, the project can be financed, and the site can still lack usable power.

Developers, grid operators, battery companies, chip suppliers and governments are responding to the connection bottleneck with local generation, storage and flexible load.

Grid queues separate national capacity from the campus

Big Tech may need 44GW nationally. A developer needs deliverable power at one site, and transmission constraints, interconnection procedures and local infrastructure can separate those two measures.

Developers confronting long waits are turning to aeroderivative turbines and diesel generators. AI labs are also deploying on-site gas generation. FERC is trying to accelerate data-center connections, while AI companies are engaging with the regulator.

Developers are paying for controllable power without waiting for the whole grid to move. Local generation, storage and demand flexibility can justify higher energy costs when the alternative is years of delay.

AI developers can use long-duration agreements to turn promised megawatts into financeable capacity, as the emerging market for contracted AI infrastructure demonstrates. A contract secures an entitlement. During congestion, equipment failure or a transition to local generation, the site’s controls determine whether that entitlement becomes usable service.

On-site power turns the campus into a distributed system

A conventional grid connection gives operators a relatively simple power input. A site combining grid supply, local generators, batteries and adjustable computing loads is a different machine. Its equipment can be available, unavailable, charging, discharging, curtailed, isolated or degraded. Operators must coordinate those states.

Lunar Energy shows what coordination looks like at a smaller scale. It began deploying its own household batteries in 2025 and raised $232 million after developing software that synchronizes batteries across homes into virtual power plants. The scale differs from an AI campus, but both require operators to make many modest assets behave like one dependable resource.

Controllers must know what is connected, which devices are healthy, what each asset may do and how the system should respond when telemetry becomes unreliable. Operators need update, exception and restoration paths. Without trustworthy state data, abundant nameplate capacity amounts to optimistic spreadsheets with cables attached.

An isolated battery or generator can cut an energy bill or bridge an outage. Once operators integrate generation, storage and computing demand, a controller failure can affect both campus uptime and grid behavior.

A self-protecting site can destabilize the grid

National Grid, Nvidia and other industry participants found in a trial that AI data centers could adjust consumption when asked rather than drawing peak power continuously. The compute load itself can help relieve grid stress.

Data centers switching to local generation can also make unannounced disconnections from the grid, potentially contributing to cascading outages. A site may protect itself while creating an unexpected imbalance elsewhere.

Facility operators need controls that verify a grid event before islanding, notify the grid operator, preserve a safe local mode and reconnect under an agreed protocol. They must also distinguish a failed sensor from a real grid disturbance.

FERC can speed a connection, and National Grid can request flexible demand. Unannounced transfers to local power fall outside both mechanisms. Grid operators and site owners need explicit notice, curtailment and reconnection duties at that boundary.

Investors fund capacity while Ireland makes operators bring power

By December 16, 2025, shares in Chinese power-equipment maker Sungrow had risen 130% that year, the Financial Times reported, as exports boosted margins despite U.S. tariffs. Lunar’s February 2026 financing backed the control layer alongside the hardware. Investors are funding both added capacity and its coordination.

Ireland is assigning more responsibility to data-center operators. The Wall Street Journal reported on June 8, 2026, that its Bring Your Own Power policy requires new facilities to provide on-site power plants or contract for energy produced nearby. In December 2024, the Financial Times reported that NERC had warned rising data-center demand could strain U.S. and Canadian grids and increase blackout risk.

A turbine, battery or inverter’s nameplate rating shows investors its megawatts. The rating omits rollback, replacement continuity and restoration time. When buyers base awards on price and capacity alone, hardware suppliers capture the scarcity premium while recovery remains unpriced.

Embedded vendors can control recovery

The United States reportedly urged NATO allies to use defense spending to replace Huawei components in networks and critical infrastructure. The request makes replacement capacity a defense requirement.

A reported five-year, roughly $295 billion Chinese AI data-center plan would source more than 80% of its technology from domestic suppliers, including Huawei. If implemented, the plan would use infrastructure procurement to build an integrated domestic stack.

Seven years after U.S. restrictions cut Huawei off from advanced American chips, the company has reportedly been rebuilding its chip business and pursuing logic-stacking technology. A teardown found that 57% of components in Huawei’s Mate 70 Pro and Pura 80 Pro were made in China, compared with 32% in similarly priced Huawei phones in 2023. U.S. restrictions strengthened the incentive for domestic substitution.

The reports establish neither a Huawei inverter deployment in the United States nor an FCC inverter import ban. They support a narrower conclusion: governments are scrutinizing the components, software and recovery dependencies embedded in critical infrastructure.

For networked power equipment, buyers must ask who can authenticate privileged access, issue updates and audit changes. They also need to know whether operators can isolate compromised nodes and restore service without a single vendor’s intervention. When only the vendor can recover a device, a cheap purchase becomes an expensive dependency.

Procurement teams must test the exception path

Data-center buyers have to add recovery tests to capacity, efficiency, warranty terms and purchase price. Networked equipment introduces propagation risk because shared control logic can spread one defective update across many devices.

Procurement teams can make deployment accountability concrete before acceptance. The site team should deploy a signed update to a subset of devices, roll it back and verify the audit trail. It should also cut telemetry, rotate privileged credentials, isolate a failed unit and restore service with a documented spare.

Site teams can use those drills to test corrupted software, lost credentials, failed communications and supply interruption separately from ordinary operation. Buyers learn little about those exception paths from normal uptime. If the site team cannot complete the drills without vendor improvisation, warranty and availability metrics leave control of recovery unresolved.

Forty-four gigawatts on paper cannot shorten a seven-year connection queue. Developers can spread generation, storage and flexible demand across the campus, but every added asset creates another state to observe, isolate and restore. For an AI campus, the grid now ends at the last device the operator can recover without losing control.