Data Center Islanding Interconnection Is the Harder Case
Data center islanding interconnection looks like less work for the utility. It is more. The protection scope nobody puts in the interconnection budget.
Data center islanding interconnection is being specified from two directions at once, and the two directions want the same piece of equipment for opposite reasons.
On one side, data center operators want to run without the grid. Capgemini Research Institute, surveying energy and data center executives across 21 countries, found that 88 percent of North American respondents view the ability to island from the grid during an outage as an important competitive advantage, with 86 percent saying the same globally. Twenty-nine percent have already put behind-the-meter generation on site. Another 39 percent expect to consider it in the next one to two years.
On the other side, regulators want a breaker they can open. On July 24, 2026 the Public Utility Commission of Texas approved a co-located AI data center in Hutto, in Oncor territory, on the condition that the load can be curtailed in full within 30 minutes during a grid emergency, with, in the order’s words, “physical breaker disconnection if necessary.” ERCOT gives 60 minutes notice when practicable, or the operator may voluntarily commit to a 10-minute response instead. Utility Dive reported the conditions on approval.
A site engineered to run independent of the grid and a site required to accept a commanded full disconnect need the same physical asset: a full-load-rated point of separation with relaying behind it and a settled answer to who is allowed to trip it. Almost nobody is budgeting the utility side of that asset, because the mental model says a customer with its own generation is less work for the utility. It is more.
Anti-Islanding Protection on a Large Load Is Utility-Side Scope
Curtailable and islanding sound like the same customer described twice. They are two different protection problems, and data center islanding interconnection is the expensive one.
A curtailable load stays grid-following. The utility needs to be able to reduce or remove it, which is the hardware we covered when load flexibility became an equipment spec: utility-owned load-side breakers, relays, reclosers and a telemetry path to prove the reduction happened.
An islanding load can energize its own bus. That single capability adds a list of things the grid-following case never needs. The point of separation has to be named in the interconnection agreement, and named specifically enough that both parties know which device it is, who owns it, who sets it and who tests it. Anti-islanding protection on a large load has to keep a customer-owned source from backfeeding a de-energized utility feeder, which is a lineman safety requirement and not a commercial term, and it has to work alongside whatever transfer scheme data center designers have already committed to inside the fence. Reclose blocking has to hold the utility recloser out while the island is alive. Sync-check relaying has to govern the return to parallel, along with an agreed resynchronization procedure. And protection coordination has to work twice, because the fault current the site contributes while grid-parallel is not the fault duty it sees running islanded, and both cases have to coordinate.
None of that appears on a conventional transformer-and-switchgear list. It is the part of the scope that gets discovered during the protection study, which on a long-lead equipment calendar is late.
Texas Wrote a Behind the Meter Interconnection Spec Into an Order
The most useful phrase in the Texas approval is “physical breaker disconnection if necessary.” A commission wrote a hardware capability into a load interconnection approval. That is a different instrument from a contractual curtailment obligation backed by a penalty. Somebody has to buy the device, install it, own it, set it, test it and maintain it.
Unpack what a 30-minute full-load drop of a campus that size actually implies. It implies an interrupting device rated for the full site load rather than a distribution recloser, which at that scale puts it in transmission or subtransmission class switchgear. It implies a control path with settled authority, because the question of who issues the trip, the grid operator, the wires utility or the customer’s own energy management system acting on an instruction, is a design input and not a footnote. A 30-minute window is generous enough for a human in the loop; the voluntary 10-minute alternative is not, and effectively requires an automated path. It implies metering and communications back to the grid operator, since an obligation nobody can verify is not a condition. And it implies protection coordination for an intentional full-load separation, including what happens to the co-located generation at the instant the load goes away, because the generation does not disappear when the load does.
The scale detail worth carrying: the two data centers at the Hutto site total roughly 525 MW of load against approximately 265.5 MW of co-located wind. The developer argued the second phase should face different treatment. The commission held that allowing selective curtailment “would undermine the goal of ensuring generation capacity is available to the grid during emergencies.”
Read that arithmetic before you read the press release. Co-location adjacency does not convert a grid-dependent load into a behind-the-meter one. A campus whose load exceeds its co-located generation is a grid customer wearing a co-location structure, and the commission priced it that way. Buyers and developers who assume proximity to a generator changes the interconnection category are mispricing the interconnection.
The pattern is established rather than emerging. This was the third curtailment-conditioned large-load approval logged in the same month, after Georgia Power’s arrangement with OpenAI and Duke’s Customer Protection Plus filing. More telling than the conditions themselves is that the commission adopted ERCOT’s generic reliability conditions rather than negotiating a project-specific carve-out. Generic conditions applied case after case are how a requirement turns into a standard term, and a standard term is how it turns into a product specification. The same logic drove the PJM large load curtailment rule toward a registry and a telemetry obligation.
The Fight Nobody Has Had Yet Is Over Trip Authority
Here is where the two motives collide.
The operator wants the island because uninterrupted operation is the product. The regulator wants the disconnect because dropping load is the last step before rolling outages. Both are asking for a breaker at the same electrical boundary. They are not asking for the same control philosophy.
An operator specifying for autonomy wants separation that is fast, local, automatic and returns to parallel on its own terms. A regulator specifying for reliability wants separation that a third party can command on notice and that stays open until released. Those are different relay settings, different communications assumptions and, in the end, a different answer to a governance question that is currently being deferred: whose signal opens the breaker, and whose permission closes it.
That question is cheapest to answer while the interconnection agreement is a draft. It is most expensive to answer after the switchgear is ordered, because the settings, the control wiring and the communications path all follow from it.
The Onsite Generation Data Center Buyers Are Choosing Has Changed
There is a second change worth tracking, because it moves the protection review before anyone has finished the first argument.
Global Energy Monitor’s August 2026 release puts 189 GW of proposed US gas capacity tied to data centers, out of 378 GW of US gas capacity in development. Inside those figures is a line most coverage skipped: reciprocating engine capacity announced for data centers reached roughly 45 GW, roughly triple where it stood six months earlier.
Engines are not small turbines. A multi-unit reciprocating plant is a different paralleling problem, a different fault-current contribution, a different harmonic and governor response, and a different set of protection settings from a single large machine. If the onsite generation being proposed to a municipal utility or a cooperative is increasingly banks of engines, then the interconnection review that utility is being asked to perform is changing shape at the same time its volume is rising. That is a staffing question as much as an engineering one, and it lands on exactly the buyers least equipped to absorb it. The same shift is visible in the broader move toward hyperscaler-owned generation.
One boundary on those numbers, because it matters for how you use them. The 189 GW and 378 GW are capacity in development, not capacity being built. Global Energy Monitor’s own figures show 52 GW of that US total under construction, and only 16.9 GW of the data-center-tied portion. Announced generation is an option book, not a build plan, and a load forecast resting on it deserves the same discount as a pipeline figure read against contracted load.
What to Ask Before the Interconnection Agreement Is Signed
Four questions, and none of them requires a protection study to ask.
Ask where the island separates, by device, not by concept. An interconnection agreement that describes a capability without naming the point of separation has deferred the entire specification. Name the device, then settle ownership, setting authority and test responsibility for it in the same paragraph.
Ask who holds trip authority and who holds close authority, and write both down separately. They are not the same permission and they will not necessarily sit with the same party.
Ask what the site does at the instant the load separates. If there is co-located generation, the answer is a real engineering result and not a reassurance. A commanded full-load drop leaves generation running into a changed system, and the response of that generation is part of the utility’s problem whether or not the utility owns it.
Ask what the onsite generation actually is before you accept a protection scope. Engines and turbines produce different studies. A scope written against one and delivered against the other gets revised at the worst point on the calendar.
The survey numbers say this conversation is coming to more utilities than have staffed for it. Bloom Energy, which sells onsite generation and therefore has an interest in the finding, reported from interviews with 156 data center operators that 61 percent would choose onsite power over relocating a project if the grid is unavailable. Treat the vendor-sponsored projections about 2030 as projections. The directional point stands on its own: grid independence is being treated as a product feature rather than a backup arrangement, and that is a change in what a utility is going to be asked for.
Where the Protection Scope Shows Up First
The full protection-scheme matrix for data center islanding interconnection, device class by device class with the coordination cases written out, belongs in a scoped report rather than a blog post. What travels here is the framing, and the framing is the part that has to arrive before the purchase order.
We read the dockets and the approval orders so the terms show up on your checklist while the interconnection agreement is still a draft.
The Feeder is our free monthly read on what changed and what it means for the next order. No cost, one email a month.
Related Reading
- Data Center Load Flexibility Is Now an Equipment Spec
- PJM Large Load Curtailment Rule: The Gear It Requires
- Hyperscalers Own Generation: The Data Center Buyer Shift
Frequently Asked Questions
What is data center islanding?
Running the facility disconnected from the utility grid, carried entirely by onsite generation, rather than using that generation as backup behind a normally closed service. The site separates at a defined point and operates as its own electrical island until it resynchronizes and returns to parallel.
Does an islanding data center reduce the utility's interconnection work?
No. A customer that can energize its own bus has to be prevented from energizing a de-energized utility feeder, which is a lineman safety requirement rather than a commercial one. That adds utility-side anti-islanding and directional relaying, a defined point of separation, reclose blocking and sync-check on return, none of which a conventional grid-following load needs.
What did the Texas commission require of the Crusoe and Ensign site?
The Public Utility Commission of Texas approved the co-location on July 24, 2026 on the condition that the load can curtail in full within 30 minutes during a grid emergency, with physical breaker disconnection if necessary, on 60 minutes notice when practicable, as reported by Utility Dive.
Does co-locating next to a generator make a data center behind the meter?
Not on its own. At the Hutto site the two data centers total roughly 525 MW of load against about 265.5 MW of co-located wind, so the campus cannot carry itself at full load and remains a grid dependent load. The commission read the arithmetic that way and required full site curtailment rather than curtailing only the grid supplied portion.
Know what changed before your next quote
Free Member tier. Pick your topics. Get a weekly digest filtered to what you actually buy.