Google has signed one gigawatt of demand response on its data centers, and the load it agreed to cut first is machine learning.
The AI capacity conversation is usually held in units of chips and dollars. In the United States the binding constraint is neither. It is a connection date. Grid operators cannot energize new large loads fast enough, and the interconnection queue is where AI programs now lose years rather than quarters. That changes the question. It is no longer only how much capacity you can buy. It is what you are willing to concede to get it sooner.
Google made the concession explicit. In August 2025 it announced agreements with Indiana Michigan Power and the Tennessee Valley Authority, describing them as the first time it delivers data center demand response "by targeting machine learning (ML) workloads" (Google, August 4, 2025). By March 2026 the same programme reached one gigawatt across five utilities, adding Entergy Arkansas, Minnesota Power and DTE Energy, and Google states these contracts incorporate demand response "as a key resource for new data centers to connect more rapidly to local grids" (Google, March 19, 2026). Flexibility buys a connection date. That is the whole trade.
It is about to stop being one company's arrangement. On June 18, 2026 the Federal Energy Regulatory Commission opened section 206 proceedings against all six US organized markets at once, PJM, MISO, SPP, CAISO, ISO New England and NYISO, directing each to justify or reform how it connects large loads (McGuireWoods, June 2026). One of the five reform categories is service for flexible large loads. The orders point toward non-firm and conditional service options, and toward ramp-up service that lets a site begin operating before its network upgrades are finished (Bracewell, 2026). The responses are due August 17, 2026. That is Monday.
Firm power was the silent assumption under every production workload. It is becoming the premium tier. The fast tier is the one that agrees to stop. Hikari Blue · operator note
The new part is which load is being offered
Demand response is not new. Aluminium smelters and irrigation districts have sold interruptibility for decades. What is new is the class of load on the table. A training run tolerates a pause. A batch scoring job tolerates a pause. A payment authorisation does not. So what is being offered to the grid is not the data center. It is a named set of AI workloads inside it.
That distinction carries the entire arrangement, and almost no enterprise has made it. Ask a CIO which AI workloads can pause for four hours on twelve hours' notice, and the honest answer is usually that nobody has ever been asked. The classification does not exist because until now nothing forced it to exist.
Where this reaches a regulated enterprise
Be precise about the exposure, because it is easy to overstate. A bank or an insurer renting capacity from a hyperscaler does not hold the curtailment risk today. The provider absorbs it. What the enterprise sees is second order: which regions have capacity, and how long a dedicated commitment takes to stand up.
The exposure becomes direct the moment you stop renting. Dedicated capacity, sovereign deployments, colocation for data that cannot leave a jurisdiction: these are precisely the arrangements a regulated firm chooses, and they are the arrangements where the interconnection terms are yours to sign. If your fast path to capacity is a conditional service agreement, then the availability envelope under your AI operating layer is written by a grid operator, and it appears in your third-party risk register or it appears nowhere.
The pressure behind this is not speculative, and it is not settled either. Data centers consumed about 4.4 percent of total US electricity in 2023, and the Department of Energy's Berkeley Lab report puts 2028 in a range of roughly 6.7 to 12 percent (DOE and Lawrence Berkeley National Laboratory, December 2024). Read the second figure as a projection, not a measurement. The width of the range is the point. No one is planning generation against a known number, which is exactly why flexibility is what gets priced.
What an operating layer needs to absorb a pause
Three things, and most stacks have none of them. A declared workload class that states which inference may queue and which may not, signed by the business owner rather than assumed by the platform team. A degradation path that is not an outage: shed the batch, drop to a cheaper model, move interactive traffic to another region. And a record of what was shed, when, and on whose instruction, because a supervised firm gets asked why a control ran late, and "the grid" is not an answer a regulator accepts without evidence.
Build those three and interruptibility becomes something you can sell. Skip them and the only honest reply to a utility offering a faster connection is no, which puts you back in the same queue as everyone else, at the same date.
What to examine before the next capacity decision
- Which AI workloads carry a stated tolerance for delay, expressed in hours and owned by the business, not inferred by engineering.
- Whether any current or pending capacity agreement, direct or through a colocation provider, contains curtailment, non-firm or conditional service terms. Read the interconnection annex, not the service level summary.
- What happens when compute is halved rather than lost, and whether that path has ever been exercised under load.
- Whether a curtailment event would surface in operational resilience reporting or disappear inside a platform availability metric.
Track one number: the share of AI compute hours you could defer by twelve hours without a consequence you would have to explain to a supervisor or a customer. Most firms cannot state it. The ones that can negotiate for capacity from a different position.
The question to bring to the next executive session
Which of our AI workloads would we agree to pause, and what do we want in return?
The six filings land on Monday, and what they propose takes months to settle into tariffs. The classification they will eventually ask for costs nothing to start now, and it is the part no vendor can do for you.
Sources
- Google (August 4, 2025). How we're making data centers more flexible to benefit power grids. Primary source, publisher's own announcement, for the Indiana Michigan Power and Tennessee Valley Authority agreements, for the statement that this is the first time Google delivers data center demand response by targeting machine learning workloads, and for the earlier Omaha Public Power District pilot in which ML workload power was reduced during three grid events. blog.google/inside-google/infrastructure/how-were-making-data-centers-more-flexible-to-benefit-power-grids
- Google (March 19, 2026). Google signed 1 GW of data center demand response. Primary source for the one gigawatt figure, for the five named utilities (Indiana Michigan Power, Tennessee Valley Authority, Entergy Arkansas, Minnesota Power, DTE Energy), and for the statement that these contracts incorporate demand response as a key resource for new data centers to connect more rapidly to local grids. blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/demand-response-data-center-milestone
- McGuireWoods (June 2026). FERC Issues Section 206 Show Cause Orders Directing All Six RTOs/ISOs to Justify or Reform Large Load Integration Rules. Source for the June 18, 2026 issuance, for the six dockets (PJM EL26-67, SPP EL26-68, NYISO EL26-69, MISO EL26-70, CAISO EL26-71, ISO New England EL26-72), for the five reform categories including transmission service for flexible large loads, for the August 17, 2026 response deadline and for the July 20, 2026 generation adequacy report. mcguirewoods.com/client-resources/alerts/2026/6/ferc-issues-section-206-show-cause-orders
- Bracewell (2026). Key Takeaways for Data Center Developers from FERC's June 2026 Show Cause Orders. Second, independent reading of the same orders, used to corroborate the August 17, 2026 deadline and as the source for the direction toward non-firm or conditional service options and ramp-up service as network upgrades are completed. The Commission's own pages at ferc.gov refuse automated retrieval, so the orders are cited here through two law firm analyses that were read in full and that agree on dates, dockets and reform categories. The underlying documents are public in FERC eLibrary under the dockets above. bracewell.com/resources/ferc-show-cause-orders-data-center-interconnection
- US Department of Energy and Lawrence Berkeley National Laboratory (December 20, 2024). 2024 Report on U.S. Data Center Energy Use. Federal source for data centers at about 4.4 percent of total US electricity in 2023 and for the projected range of approximately 6.7 to 12 percent by 2028. The 2028 figure is a projection, not an observation. energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
The Hikari Blue team · Austin, August 2026