Request an assessment
Flexible-load interconnection

The grid is already built.
It sits idle 361 days a year.

Utilities size transmission for the single hottest hour of the year. Headroom lets an AI data center take that capacity now, by making the promise to step down on the tight days forecastable, contractual, and provable in the field.

Request a site assessment Watch an event run
ForecastContractDispatchSettleEvidenceExecute
Load duration curve — transmission node 8,760 h · sorted desc
100% 80% 60% 40% 0d 4d 100d 250d 365d UNUSED CAPACITY · 361 DAYS ALREADY BUILT · ALREADY PAID FOR DELIVERED LOAD PEAK — 4 DAYS/YR
Peak window Unused headroom Delivered load
Why finished sites sit dark

Capital is available. GPUs are shipping.
Land is secured. The queue is not moving.

4–7 yrs Interconnection queue Time to connect a large load in constrained Tier‑1 markets such as Northern Virginia.
$2.4M Lost per MW, per year Unbooked revenue while a completed data center waits up to eighteen months for energization.
1.4% North American vacancy Record low across primary power markets. Demand is not the constraint. Delivery is.
$7–25M To build one new MW Capital cost of a megawatt of new firm capacity, on the far side of a three-year transmission queue.

CBRE H2 2025 asking rates · industry reporting on energization delays

The trade

Four days of flexibility, in exchange for power today.

A transmission line is sized for the worst hour of the worst day of the summer. It spends the rest of the year with capacity to spare. The queue is not protecting a shortage of electricity — it is protecting a handful of hours.

A load that can step down during those hours does not require new capacity to be built for it. It requires a contract that says so, and a way to prove it kept the promise.

That trade has been obvious for years. Four things kept it theoretical: the forecast was too loose to underwrite, the grid data sat in four systems that disagreed, there was no product to sign, and nobody could prove after the fact that the load actually came off.

The stack

Six things have to happen.
Headroom does all six.

Turning a firm grid constraint into a flexible tariff is a sequence. Every step depends on the one before it, and the last one is physical.

A node-level model that puts a number on the probability of a constraint, hour by hour, up to 48 hours ahead. It reads grid state, weather, market prices and the site's own telemetry together, so the answer is one number rather than four systems disagreeing. Confidence tightens as the horizon closes, and each downstream gate fires on its own threshold.
The commitment is written as a product, not a letter of intent: notice windows, curtailment depth, response time, duration caps, annual event limits and the penalty structure behind them. Notices fire programmatically to the utility and to the cluster scheduler. A project financier can underwrite this. They cannot underwrite an estimate.
Curtailment depth is composed, not commanded. GPU power capping and deferrable batch queues come off first because they cost nothing. Storage follows. On-site generation is the last resort, not the plan. The site meets its obligation at a fraction of the cost of running the generators, and the compute that matters keeps running.
Performance is priced against the tariff automatically — credits for delivered flexibility, charges where an obligation was missed, reconciled against metered data rather than a monthly spreadsheet exchange. Both sides see the same arithmetic on the same record.
Every forecast, notice, instruction and meter reading is hashed and chained in sequence, so the record of an event cannot be edited after the fact. Regulators do not block flexibility. They block what they cannot independently verify. This is the artifact that ends that argument — and it is the same artifact a utility, an auditor and a lender each need.
The transfer scheme is designed, built, commissioned under real load and witnessed on site — switchgear, transfer controls, protection, generation interface and re-verification on a schedule. Headroom holds the electrical license and self-performs this work rather than subcontracting it. Layers 1–5 are a claim until layer 6 is wired
Machine-readable end to end. The full stack is exposed over REST and Model Context Protocol, so an autonomous GPU scheduler can query headroom, accept a notice and confirm a curtailment without a human in the loop.
One event, end to end

A hot, still evening on a 250 MW site.

Wind drops across the region. The line rating falls 35% on ambient temperature. Node headroom goes to −70 MW. Here is what happens, and when.

T − 48 hours
54%
Advisory

Non-critical batch training pre-drains. Storage begins staging state of charge. Nothing is committed yet.

T − 24 hours
70%
Notice

The notice gate fires to the utility and to the cluster scheduler. Power-capping profiles lock for the window.

T − 4 hours
91%
Firm

On-site fuel logistics warm up. Storage discharge availability is validated against the committed depth.

T − 1 hour
95%
Dispatch

70 MW comes off in 41 seconds. Signed, hashed and time-stamped into the evidence chain as it happens.

Each gate fires on its own threshold as the forecast tightens. No human in the loop.

Dispatch economics

Curtailment is not a diesel bill.

The reason flexible load has a reputation for being expensive is that the industry treats curtailment as a single lever: shed everything, run the generators, absorb the cost. Composing the depth across four sources instead — free ones first — changes the arithmetic entirely.

GPU workload power capping 22 MW < 60 seconds $0 / MWh
Deferrable batch training queue 9 MW < 5 minutes $0 / MWh
On-site storage discharge 14 MW < 2 minutes $18 / MWh
On-site generation & transfer 25 MW 41 seconds $186 / MWh
Blended flex tier 70 MW Sub-minute $70 / MWh
Free firstGeneration last
$70 / MWh
Composed dispatch
$186 / MWh
Generation only
Where this category stops

Most of it ends at the notification.

Software companies cannot go down the stack. Electrical contractors cannot come up it. The regulator needs both halves signed by the same party.

Software-only platforms

  • Forecast an event and send an email or an API call.
  • Depend entirely on customer self-reported performance data.
  • Cannot sign a binding attestation to a utility commission.
  • Subcontract the high-voltage work to somebody else.

Headroom

  • Forecasts, contracts, dispatches and settles the event.
  • Engineers and builds the high-voltage transfer scheme itself.
  • Signs the attestation the commission accepts.
  • Re-verifies performance on site, on a schedule, under load.
Why now

Flexible load is becoming a tariff class.

FERC has directed all six grid operators to create a class of transmission service for loads that can flex. The tariffs are being drafted now.

Oct 2025DOE directs FERC to act on large-load interconnection.
Jan 2026SPP framework approved — interconnection in 90 days, against a five-year average.
Jun 2026Show cause orders issued to PJM, MISO, SPP, CAISO, ISO‑NE and NYISO.
Aug 2026RTO responses and tariff filings due.
2027Flexible tariffs in force. Every developer needs one.

The first accepted evidence package becomes the template everyone else is measured against.

Who this is for

Four parties, one record.

Data center developers

Energize in months instead of years, on capacity that already exists. The interconnection becomes signable and the site becomes financeable.

Utilities & ISOs

Serve new load without new build. Higher asset utilization, incremental revenue, and a counterparty whose performance can be checked rather than trusted.

Regulators

An independently verifiable record of every notice, instruction and meter reading — the thing that has been missing from every flexibility filing so far.

Ratepayers

New load carried on infrastructure already built and already paid for, rather than a rate case for capacity used four days a year.

Engagement

Assess. Build. Meter.

Each stage produces something the next one needs — and something you can take to a utility, a commission or a lender on its own.

Assess Weeks · per site

A measured flexibility envelope for the specific node: how much depth is real, how fast it can be delivered, what it costs, and a straight answer on whether this site can be energized at all.

Deliverable — capability envelope & go / no-go
Build One-time · engineered & witnessed

The transfer scheme designed, built and commissioned under real load, witnessed on site. This is the step every software-only vendor has to hand to somebody else.

Deliverable — signable interconnection, financeable site
Meter Annual · per MW under management

Forecast, notices, dispatch, settlement and the evidence chain running continuously, with re-verification on a schedule so the capacity never comes back into question.

Deliverable — a record that holds up on review
Start here

Tell us the node and the load.

An assessment starts with three things: where the site sits on the network, how much power it needs, and when it needs to be live. We will tell you what is actually available there.

We reply within two business days. No mailing list.