Demand Response, Priced and Contracted

Flexibility only becomes a strategy when it is priced and contracted. In the abstract, everyone agrees that a load able to get out of the way is valuable. The discipline is in turning that abstraction into a signed agreement with defined triggers, defined compensation, and defined penalties. This lesson is about the machinery of demand response, which is the umbrella term for programs that pay or reward load for reducing consumption on request.

Begin with the taxonomy, because the words matter in a contract. Emergency demand response is the utility or grid operator calling on you during a genuine reliability event, a moment when the alternative is rolling blackouts. These events are rare, often fewer than ten or twenty hours a year, but participation may be effectively mandatory once you have signed up, and the notice can be short. Economic demand response is different: you curtail because the wholesale price of power has spiked above your threshold, and you are effectively selling your avoided consumption back into the market at that high price. The first is about reliability; the second is about arbitrage.

Then there is the distinction between how fast you must respond. Grid operators procure several products along a spectrum. Frequency regulation requires response in seconds and is the most technically demanding. Spinning and non-spinning reserves respond in ten to thirty minutes. Capacity and peak-shaving programs may give you day-ahead notice. A data center's ability to participate depends entirely on how quickly and how deeply it can modulate its load, and that is an engineering property of the facility, not a financial choice. If your cooling and compute systems can ramp down two hundred megawatts in ten minutes without damage or data loss, you can sell into fast products. If you need an hour of warning, you are limited to slower, lower-value ones.

The economics are the reason this is worth engineering for. Capacity payments in constrained markets can run to tens of thousands of dollars per megawatt-year, and during genuine scarcity events, wholesale energy prices can spike to the market cap, which in Texas is nine thousand dollars per megawatt-hour, more than a hundred times the normal price. A load that can drop during those few hours avoids paying that spike and, in some programs, gets paid for the reduction. Over a year, the value of avoiding the highest-priced hours can materially change the effective cost of power for the facility. The peaks are where the money is, on both sides.

Consider the concrete case of ERCOT, the Texas grid, which has become the proving ground for large flexible load. Texas has an energy-only market with no capacity payments, which means prices must rise high enough during scarcity to incentivize both new supply and demand reduction. This creates sharp, transparent price signals. A data center in ERCOT that curtails during the handful of extreme-price intervals each year can save enormous sums, and ERCOT has been actively developing large-load interconnection processes precisely because so much data center demand is arriving. The Texas experience is showing the rest of the country what large flexible load looks like at scale.

The contractual structure typically involves a curtailment service provider or a direct agreement with the utility. Key terms to understand: the baseline, which is the assumed consumption against which your reduction is measured and which is subject to gaming and dispute; the notification window, which determines which products you can offer; the number and duration of events you must honor; the compensation formula; and the penalties for failing to perform when called. That last item is critical. A demand response commitment you cannot physically honor is worse than none, because non-performance during an emergency event can carry severe penalties and can damage your standing with the operator, which you need for future interconnection.

There is a strategic subtlety specific to AI operators. Your willingness to curtail is only credible if your business can tolerate it, and that depends on the mix of interruptible training versus latency-critical inference in your facility. An operator running mostly batch training has enormous flexibility to sell. An operator hosting real-time inference for a consumer product has far less, because dropping load means dropping service. Sophisticated operators are therefore designing their fleets so that flexibility is a portfolio property: they place interruptible workloads where flexible commitments earn the most, and reserve firm power for the workloads that genuinely need it. This is load shaping as corporate strategy, not just facility engineering.

Faraday's read: demand response is often taught as a sustainability footnote, a way to be a good grid citizen. Treat it instead as a financial instrument. A well-structured curtailment portfolio lowers your effective cost of power, accelerates your interconnection, and hedges you against the price spikes that will define constrained markets for the rest of this decade. The operators who master the contract terms, the baselines, and the penalty structures will run their compute cheaper than those who buy firm power and ignore the market. In a business where power is the first-order constraint, that edge compounds.