We convert unwanted electricity into a resource that strengthens grids and lowers the cost of power for everyone who depends on them. The people who keep the grid running, and who pay for it, come first.
01 · AI Data-Center Power Demand
The Layer 1 constraint
Every layer of the intelligence economy now rests on a single input that cannot be substituted, stored, or shipped: electricity delivered the instant it is demanded.
Energy sits at the center of human prosperity. Every advance from agriculture to modern computing has come from harnessing more energy and putting it to work. The next advance is constrained not by ingenuity or capital, but by the stability and flexibility of the electric grids that carry it.
NVIDIA describes artificial intelligence as a five-layer industrial stack: applications, models, AI factories, chips and systems, and energy. Energy is Layer 1 because every layer above it depends on reliable, available power.
Hyperscalers are committing roughly $630–700 billion annually to compute infrastructure, yet power availability has become the binding constraint on how far that infrastructure can scale. Without Layer 1, the layers above it do not function.
The scarcity inside Layer 1 is not generated. It is responsive. An AI data center is a firm, non-interruptible baseload: inference workloads carry hard latency constraints, so the load must be met every second. That rigidity is exactly what makes a fully interruptible compute load the resource that completes the pair, a load that absorbs surplus when power is abundant and vanishes in seconds when the system is stressed.
Firm demand and interruptible demand are not competitors for scarce capacity. They are two halves of a balanced pair, and a grid carrying gigawatts of inflexible AI load needs the second half to survive.
This buildout does not happen in the abstract. It lands in real towns, on a grid that ordinary households already pay for, and increasingly it is their bills that climb to fund it. Any honest account of the next decade of energy begins with the people who keep the system running and depend on it, not with the racks of servers it feeds.
Others propose escaping Earth's energy limits by launching compute into orbit. That takes new physics and new infrastructure, and resolves nothing this decade. Type 3 solves the same constraint on the ground, with orchestration of demand-side flexibility we can deploy today. Space-based AI is a 2030-plus experiment. Flexible terrestrial load is a 2026 solution.
Type 3 is a demand response lab. We take fully interruptible compute loads, high-density computing that can shed its entire electrical draw on command and resume without penalty, and turn them into a dispatchable, telemetry-verified grid resource across the ERCOT, SPP, PJM, and MISO markets. The work starts with the data no one else has: how that load behaves under live grid conditions, and the software-defined dispatch and verification we build on top of it. We build the most flexible load on the grid: demand response that lowers the cost of power instead of raising it.
Type 3 is the software-defined demand response infrastructure company that turns fully interruptible compute loads into a dispatchable, telemetry-verified grid resource, demand response that lowers the cost of power instead of raising it.
02 · The Distinction
The category error
A traditional or AI data center cannot shed more than roughly 10% of its load without breaching a service-level agreement. A fully interruptible compute load can shed 100% of its draw inside the dispatch window, repeatedly, at zero operational penalty.
Utility executives and ISO market designers are right to be skeptical of data centers as demand response. When they hear "data center," they picture mission-critical facilities running cloud workloads, financial transactions, and AI inference. Those loads cannot curtail deeply without destroying economic value. The skepticism is warranted. It has just been aimed at the wrong asset.
A fully interruptible compute load is not an IT workload that happens to consume power. It is a silicon energy converter: a variable-load mechanism whose economics are a direct derivative of real-time grid pricing.
There is no delivery deadline, no end-user SLA, and no spoilage. The entire load can drop in seconds and resume when power is abundant, and the only cost of curtailment is the production foregone during the window, a quantity that is precisely calculable in advance and frequently negative after grid credits.
The proven example today is Bitcoin mining: the first compute load to register as a controllable grid resource at utility scale, and the dataset behind every figure on this page.
Because the energy industry filed this load under the same heading as a hyperscale facility, it conflated two distinct asset categories and left the most flexible interruptible megawatts on the grid mispriced and underused.
Treated correctly, it is the rare new load that lowers the cost of the whole system instead of raising it. That is the real shape of data center demand response, and the reason a genuinely grid-interactive data center looks nothing like the hyperscale model. The distinction holds across every attribute a grid operator cares about.
| Engineering Vector | Fully Interruptible Compute Load | Traditional / AI Data Center |
|---|---|---|
| Load interruption | Zero operational penalty; the chips simply stop converting | Catastrophic SLA breach and revenue loss |
| Response profile | Deterministic: sub-ten-minute, automated, total shed | Probabilistic: limited to non-critical ancillary loads |
| Economic model | Energy cost is the primary input; compute is the converter | Compute service is the product; energy is a cost |
| Telemetry requirements | Straightforward: total facility meter, simple baseline | Complex: tenant-level metering, workload-specific baselines |
| Curtailment depth | 100% of load is interruptible | Typically less than 10% of total load is flexible |
| Grid value | Functions as virtual storage plus dispatchable capacity | Functions as firm load with minor flexibility |
In short: a fully interruptible compute load sheds 100% of its draw on command and resumes without penalty; a traditional or AI data center can flex under 10%. They are not the same grid asset.
The physics and economics of grid-integrated flexible load are already proven in the field. What the sector lacks is institutional sophistication: utility-grade telemetry, automated dispatch, clean measurement and verification, and professional market operations. Type 3 supplies that integration layer, and against entrenched skepticism, we lead with proof.
03 · Market by Market
The multi-ISO integration grid
The same fully interruptible compute load earns differently in every market, because each ISO prices flexibility through its own instrument, and in each, that instrument is being rewritten right now around exactly this load profile.
ERCOT's Controllable Load Resource (CLR) framework sat unused from 2004 until 2020, when the first fully interruptible compute load passed its dispatch tests. After Winter Storm Uri, which killed at least 246 people and brought the grid within minutes of uncontrolled collapse, Texas chose to expand CLR participation rather than build an estimated $18 billion in idle peaker plants, a cost that would otherwise have landed on Texas ratepayers. Registered loads now earn two simultaneous streams: a standing availability payment, and a performance credit when dispatched. Riot Platforms, a Bitcoin miner operating a 700 MW interconnection, reported $71.2 million in power curtailment credits across 2023 in its SEC filings, including $31.7 million in August alone.
On December 5, 2025, ERCOT replaced the legacy reserve-margin adder with Real-Time Co-optimization plus Batteries (RTC+B) and Ancillary Service Demand Curves (ASDCs), co-optimizing energy and ancillary services every five minutes. Manual price-watching is now obsolete; automated demand response is now table stakes. Texas Senate Bill 6 (2025) goes further, requiring new large loads of 75 MW or greater to demonstrate curtailment capability during emergencies. That turns flexibility from a bonus into a condition of interconnection.
SPP generates a larger share of its electricity from wind than any other U.S. operator, and wastes more of it. Average hourly wind curtailment reached 1,097 MW in 2023, an eight-fold rise from 136 MW in 2019 (SPP Market Monitoring Unit), and negative price intervals hit 15.2% of the real-time market in 2024. Meanwhile the planning reserve margin is forecast to fall from 24% in 2020 to roughly 11.8% by 2027, breaching the 16% summer minimum, with only 968 MW of dispatchable demand response registered across the footprint at the end of 2024. SPP's CEO has stated on record that generation cannot be built in time, and that demand response must bridge the gap.
The compensation mechanism is Market Registered Demand Response (MRDR), accredited through Effective Load-Carrying Capability (ELCC). An October 2025 study by Energy + Environmental Economics (E3) found that a ten-hour program, the profile a fully interruptible compute load already meets, earns approximately 100% summer ELCC: full capacity credit, equal to a dispatchable generator. A 50 MW interruptible load thus receives the standing capacity payment of a 50 MW gas peaker, without the plant, the fuel, or the emissions. Sited at stranded wind, that same load brings tax base and steady work to communities that host the turbines but have captured little of their value.
FERC Order 2222 is the federal directive opening organized wholesale markets to aggregated distributed resources. Each RTO translates it into a local participation model, which lets distributed flexible compute reach premium market revenue it could never access at a single small site.
In PJM, the largest competitive electricity market in the world, three consecutive capacity auctions have cleared at or against the FERC price cap, $329 to $333 per MW-day. Data centers account for 94% of the 30 GW of projected load growth through 2030, and congestion costs rose 64% in a single year. PJM then lifted demand response accreditation from 69% to 92% ELCC for the 2027/28 delivery year by rewarding around-the-clock availability, a 33% increase in capacity revenue from the same registered megawatt, favoring exactly the 24/7 interruptible profile. When a hyperscaler commits $1.6 billion to restart the 835 MW Three Mile Island reactor for round-the-clock power, the market has priced reliable capacity at the capital-allocation level.
In MISO, the largest AC transmission system on earth, the 2024 Reliability-Based Demand Curve (RBDC) drove the summer capacity price to a record $666.50 per MW-day, a 22-fold jump. Alongside it, $1.8 billion in annual congestion and 508 MW of average hourly wind curtailment reward loads sited at the source of generation, turning a congestion cost that consumers ultimately bear into local economic activity. MISO consolidated its programs into Load Modifying Resources (LMRs) and Demand Response Resources (DRRs) after more than 40% of one DR class failed performance tests, a direct demand for the deterministic, metered, verifiable response that a fully interruptible compute load is built to deliver.
04 · The Social Contract
The constraint capital can't buy
A gigawatt the public refuses to host does not get built, at any price. The deepest constraint on the AI buildout is not silicon, generation, or capital. It is the consent of the people who keep the grid running and pay its bills.
The fight over data centers is not really about data centers. It is about who carries the cost and risk of an enormous buildout, and who captures the gain. When households watch their bills rise to power compute they will never use, and workers watch an industry arrive that they fear will hollow out their livelihoods, resistance is not ignorance. It is a rational response to being treated as the cost center for someone else's returns.
U.S. utilities requested more than $29 billion in rate increases in the first half of 2025 alone, double the prior year (per utility rate-case filings), with data-center infrastructure among the drivers. Every firm, must-serve load deepens that pressure, because the substations, transformers, and transmission built to serve it are paid for, in large part, by everyone else on the system. The hyperscale model pushes its grid costs onto the public. That is the real source of the backlash, and no advertising campaign or paid endorsement will dissolve it.
Type 3's model runs the other way. A fully interruptible compute load consumes power that is already being wasted, defers the buildouts that raise bills, and is structured to lower the cost of electricity for every household and business on the grid. It can be placed where it brings work and tax base to towns that sit beside stranded energy generation and capture little of their value. It is the version of the compute buildout that pays the public back instead of billing them for it.
This is not philanthropy; it is the engineering reality of the system. The megawatts the intelligence economy needs run through thousands of interconnection points, each one approved by a regulator, a utility, and a community that can say no. An industry that treats those communities as obstacles to be bought will fall short of the capacity its own business depends on. Resentment does more than dent a reputation. It sets a hard ceiling on how much you can build.
So Type 3 builds with the people who live next to the substation, not over them. We meet towns, ratepayers, line workers, and the creative communities who fear what this technology will do to their work, face to face, and we judge our own results partly by what they get back: lower bills, local jobs, and a grid that is more reliable because it has enough demand response. The grid belongs to the public. Infrastructure that forgets that does not, in the end, get built. And it should not.
05 · Reference
Automated DR · Bitcoin mining · ELCC · Battery storage
Ref · Sources
Ref · Glossary
67 clinical definitions · Updated June 2026
Every term this manifest uses: Controllable Load Resource, FERC Order 2222, RTC+B, Market Registered Demand Response, and 63 more, has a standalone, citable definition in the Type 3 glossary.
Open the Full Glossary →