Every answer you get from an AI chatbot begins with electricity. The words appear on your screen, but the actual work happens in a distant building packed with computer chips. Those chips draw power, move data, and produce enough heat to require heavy-duty cooling from AI data centers.
When you multiply that process across millions of prompts, image requests, and business tasks, you begin to understand why a quick answer to a question that feels weightless becomes a physical demand on power plants and wires.
Electric utilities are being asked to supply that demand in enormous, concentrated blocks. Your average large data-center campus can use as much electricity as a small city, and companies can plan and build one far faster than the utility can accommodate it.
The utility also has to prepare for the hours when customers use the most electricity, even if some of that capacity goes unused during ordinary periods. In short, data centers want power sooner than the grid can provide.
One solution is to build new power plants. But it's a very expensive, time-consuming solution that can take billions of dollars and years to become operational.
However, another solution is to move some of the computer work to another hour.
A chatbot reply usually needs to appear right away, but an internal experiment or an overnight video-processing queue can wait. Software that can tell the difference could slow the work that can wait when electricity is scarce, then let it catch up when more power is available.
A small experiment in Texas shows what that arrangement might look like.
Luxor Energy, a company with roots in Bitcoin mining, teamed up with Bentaus, which makes software that controls how much power computer chips use. Together, they controlled a single Nvidia B200, a high-powered chip built for AI work.
The chip was performing inference, which simply means using a trained AI model to produce an answer, when the software told it to draw less electricity.
The companies say the chip's power draw fell to roughly 25% of normal within half a second, and it processed fewer requests during the restriction.
Ethan Vera, Luxor's chief operating officer, told CryptoSlate that no job failed and no work already in progress was lost. The chip returned to full speed when the restriction ended.
Luxor and Bentaus said their public demonstration caused “no disruption,” but the phrase needs some translation. From the operator's perspective, the job survived and resumed at full speed.
However, customers could still have waited longer for an answer because the chip completed less work during the restriction. Any plan to make AI flexible will depend on how often that delay occurs, who experiences it, and what those customers were promised.
The experiment was a success, but it involved only a single chip. Large data centers contain tens of thousands of chips, along with servers, cooling systems, storage devices, and networking equipment.
The test makes a broader idea easier to see: an AI data center could sort work by urgency and occasionally ask the grid for less.
Texas lacks power to feed the computers waiting
The best example of what happens when new data centers come faster than new power infrastructure is Texas.
The Electric Reliability Council of Texas (ERCOT) operates the grid that serves most of the state. On July 22, electricity use reached a preliminary record of 91,089 megawatts, a number that is unofficial until the data gets finalized.
ERCOT says one megawatt can serve about 250 residential customers during a peak hour. By that rough comparison, the record matched the needs of more than 22 million residential customers at once.
Gov. Greg Abbott said in August that ERCOT was reviewing requests to connect more than 474 gigawatts of new electricity use, with about 90% coming from data centers. One gigawatt equals 1,000 megawatts, so on paper, the queue asks for more than five times the power used during ERCOT's record hour.
Abbott ordered regulators to audit the projects before letting them proceed.
In a July 28 preliminary review, ERCOT found that roughly 205 gigawatts had enough supporting studies to qualify for the first study batch, less than half of the 474-gigawatt total. Abbott's audit interrupted that review.
Regulators gave ERCOT more time on Aug. 20, and the agency said it would send conditional eligibility decisions by Aug. 31. Developers can submit overlapping proposals, hold places for projects that never secure financing, or ask several locations to provide power for one eventual campus.
Texas is conducting the audit partly because the list has become too detached from physical possibility to guide grid planning on its own.
But even with that caveat, 474 gigawatts shows the rush for land with access to large amounts of electricity. Far more machines are proposed than wires are ready to serve them.
A Lawrence Berkeley National Laboratory update published this year estimates that data centers could consume 11.8% of US electricity in 2030. Its low estimate is 9.5%, and its high estimate is 15.3%. The International Energy Agency expects data centers to account for about half of the increase in US electricity use through the end of the decade.
But even with this kind of demand, transmission lines in advanced economies can take four to eight years to complete. The agency says waits for vital equipment, including transformers and cables, have doubled over the past three years.
AI companies tend to talk in chips, but electric systems have to think in cities. An individual B200 can draw as much as 1,000 watts. Nvidia lists maximum power use of about 14.3 kilowatts for a complete eight-GPU DGX B200 server. One megawatt equals 1,000 kilowatts, and Texas's new rules for very large electricity users begin at 75 megawatts.
Under ERCOT's residential-customer comparison, that amount could serve roughly 18,750 customers during a peak hour. It could also power 75,000 one-kilowatt GPUs, at least before adding processors, cooling, networking, batteries, and electrical losses.
So learning how to control and curtail the power use of one of those chips is the first of many, many steps toward understanding how to manage power use across an entire data center.
The sheer complexity of that endeavor, in both software and hardware demands, is why grid planners treat data centers as “firm loads,” meaning electricity must be available whenever they ask for it.
Data center operators want expensive GPUs running continuously because every idle minute delays work that customers are paying for. Thousands of chips working on a single large AI job are tightly interdependent.
At certain points, one group may have to wait for another to finish before it can continue. If you slow down a selected group, the delay can ripple through nearby machines.
But not all the computing work in a data center has to happen immediately or run at full speed. Some jobs are time-sensitive, while others can be delayed or run more slowly with little consequence. Some can even be shifted to another data center where electricity is more readily available.
Each choice comes with trade-offs, but each can reduce the power a data center needs from the local grid at a given moment.
Bitcoin miners taught computers how to yield
The precedent comes from Bitcoin mining on the Texas grid. Bitcoin miners compete to earn rewards by running machines that perform calculations continuously. When a machine shuts down, the miner loses the chance to earn money for that period.
But when power returns, the machine can resume almost immediately. No customer is waiting for a response, and no unfinished computing job has to be preserved.
Texas figured out that the basic idea is called demand response: when electricity gets scarce and expensive, big users get a reason to use less of it.
Bitcoin miners were unusually well suited to the deal. They could shut down when wholesale prices spiked, get paid for cutting power during emergencies, and trim transmission charges by sitting out a handful of critical summer hours.
An ERCOT review in April described crypto miners as the main price-sensitive participants in one of its emergency programs. For a miner, the calculation is simple: when a megawatt becomes more valuable than the Bitcoin the machines might earn with it, turn the machines off.
Luxor supplies Bitcoin miners with software, energy services, and financial products, so it approached AI with an instinct for computation that can be interrupted. The experiment asks whether machines serving customers can inherit some of mining's obedience to electricity prices.
That question is becoming more urgent as miners convert power-rich sites into AI campuses. If the grid trades a Bitcoin mine that can shut down on command for a data center that runs around the clock, it may be giving up a valuable emergency brake.
How flexible a data center can be depends heavily on what its machines are doing.
Training is the long, compute-heavy process of teaching a model, repeatedly adjusting it as it works through enormous amounts of data. Inference is what happens afterward, when someone asks the finished model for an answer, an image, a translation, or a prediction. The two create different opportunities for cutting power.
A long training run can sometimes pause at a saved checkpoint and pick up later, though stopping thousands of machines in sync is not trivial. Inference can consist of millions of smaller requests, some from people expecting an answer immediately and others from automated jobs that can wait in a queue until electricity is easier or cheaper to come by.
Google has been sorting its computing this way for years. In 2023, the company described how it could delay work such as YouTube video processing when a local grid was under strain, or send that work to another region with more power available. Search, Maps, and other services people expect to work immediately stayed online.
Google later brought the same idea to machine-learning workloads. By March 2026, it said it had put one gigawatt of data-center demand response under long-term utility contracts across several US regions.
Some of those deals could also help new data centers connect to the grid sooner.
Researchers are now showing that this can work outside simulations. In a peer-reviewed Nature Energy paper, a team described an experiment at an Oracle cloud facility in Phoenix. Software cut the power used by a 256-GPU cluster by 25% for three hours without pushing priority jobs outside their promised performance levels.
The key was deciding where to absorb the slowdown. The software that determines which jobs run and when, called the scheduler, protected urgent work and pulled the power savings from jobs with more forgiving deadlines.
Emerald AI, the company that led that work, announced a $150 million financing round on Aug. 25 that valued it at over $1 billion. It also said its software was operating commercially across entire data centers, drawing multiple megawatts.
Independent performance data for every site aren't available, but even so, the financing shows that flexible AI has moved beyond research papers and into a commercial business.
Other researchers have tried to estimate how much electricity an AI facility could reliably promise to give up during a difficult hour.
A University of Chicago working paper used four years of electricity prices and 49.4 million real inference requests to model the answer. The author estimated that a facility focused on inference could commit to cutting 40% of its demand. A facility running a mix of inference and training could commit 24.6%.
Those percentages fell only slightly when the model expanded to a 10-gigawatt fleet. The main limits came from customer contracts, restrictions on moving work, and the rush of machines returning to full power.
Researchers at the University of Alberta modeled what happens to the grid when AI jobs can be delayed or moved between data centers. In the model’s most stressed scenario, that flexibility cut the amount of power-plant capacity needed by more than 21%. In another scenario, where the local grid was congested, it reduced the total cost of supplying electricity by 3.5%, even though spending on new generation rose 7.1%.
Most of the benefit from delaying jobs appeared within the first three hours, so waiting longer didn’t help much more. Although none of this eliminated the need to build new power plants and transmission lines, it showed the grid could meet more AI demand with less infrastructure and at a lower overall cost.
Four hidden moments can price an entire year
The money behind Luxor’s experiment comes from an unusual feature of the Texas electricity market. Large customers help pay for the high-voltage transmission network, and part of that bill can hinge on how much power they use during just four 15-minute windows all year.
Those windows are the moments of highest systemwide demand in June, July, August, and September, known as the Four Coincident Peaks, or 4CPs.
The catch is that nobody knows exactly when a 4CP is happening until the month is over. So large power users hire forecasters to watch the grid, the weather, and electricity demand and predict when a peak is likely.
If the odds look high enough, they cut their power use for that 15-minute window. Guess right often enough, and the savings on transmission charges can be substantial. That has turned 4CP into a recurring game of prediction and power cuts for factories, Bitcoin mines, batteries, and now, potentially, AI data centers.
That potential payoff makes many false alarms worth tolerating. The latest 2026 PUCT numbers put ERCOT transmission costs at about $6 billion, spread across an average 4CP demand of 80,859.8 megawatts.
That works out to roughly $74.89 per kilowatt per year. At that rate, 100 megawatts of demand during the four peak windows represents about $7.49 million in annual transmission costs.
While the actual bill will vary by utility territory and contract, the financial incentive here is pretty clear. A large data center can have millions of dollars riding on just one hour of electricity use scattered across an entire summer. Cutting power for a few extra hours to capture that hour can be a very good trade.
Luxor decided to throttle the GPU itself, using live grid data to decide when to act. Vera said the company watched for signs that a 4CP window might be forming, then sent its own command to the chip. ERCOT never told the GPU to slow down, and no emergency grid program was involved.
This was essentially a private bet on when electricity demand would peak, aimed at lowering the site’s transmission bill. ERCOT classifies this kind of 4CP self-curtailment separately from the demand-response programs it operates.
That also puts the half-second response time in perspective. A 4CP window lasts 15 minutes, so whether the GPU reaches its lower power level in half a second or several seconds makes almost no difference to the transmission savings.
ERCOT’s emergency program generally gives participating customers 10 or 30 minutes to deliver the power reduction they promised. Some other grid services move faster, requiring customers to start cutting power immediately and reach the full reduction within 10 minutes.
If AI hardware eventually participates in those markets, sub-second control could become more useful. For Luxor, every extra second a GPU spends throttled is a second it could have spent earning money by computing.
Bentaus had already tested the same basic idea at a larger scale. In February, CPower, Bentaus, and Supermicro described a California demonstration using a cluster of servers equipped with B200 GPUs.
The companies said the cluster responded to a signal tied to the state’s wholesale electricity market in less than 20 milliseconds and cut its power use by as much as 75%, while still meeting its promised performance levels.
The Texas experiment is smaller and much narrower: one GPU responding to a specific transmission-billing incentive. But it adds another real-world test to an idea that has already moved from individual chips to server clusters and utility programs.
Important gaps remain in what we know about the Texas test. The companies haven’t disclosed which AI model was running or how long the GPU stayed at reduced power. They haven’t said how much electricity it was using beforehand, how much its computing throughput dropped, or how much longer requests took to complete.
Luxor’s representative in the Texas electricity market verified the power reduction, but no independent analysis of the test has been published.
The test showed that one B200 running an inference workload could take a steep power cut without losing the work already in progress. It is uncertain what that did to user wait times, whether other inference or training workloads would respond the same way, or how much electricity the technique could save across an entire data center.
A GPU is only one part of a building’s power bill. Cooling systems, networking equipment, storage, pumps, and power conversion also consume electricity. So cutting a chip’s power by 75% doesn't mean the data center draws 75% less power from the grid.
The reduction measured at the building’s meter could be considerably smaller.
Luxor is already preparing its next test, this time with a group of Nvidia H100 GPUs in Texas. Vera said scaling up means building software that can figure out which jobs can safely slow down, then coordinate the machines working on them. It also has to respect whatever performance customers were promised.
Every jump in scale, from one GPU to a server, a rack, and eventually an entire data center, adds another layer of complexity. More equipment draws power, more machines have to move together, and more customer workloads may or may not tolerate a slowdown.
Texas is starting to require some of that flexibility. Senate Bill 6, passed in 2025, requires certain large power users connecting from 2026 onward to cut consumption during severe grid emergencies. It also calls for a program that would pay sites using at least 75 megawatts to reduce demand when trouble is expected.
At the same time, the state is also rethinking 4CP. Its four summer peaks can miss the evening and winter hours when the grid is under more stress. Regulators have proposed replacing it with 12CP, which would base transmission charges on one 30-minute peak each month.
ERCOT reached a similar conclusion in an April review: Texas has plenty of demand response, but it doesn’t always show up when the grid needs it most. 4CP drives most of those power cuts, but its summer peaks can miss the hours when demand is high and wind and solar output is low.
ERCOT said that mismatch is a problem. New power plants and transmission lines take years, but flexible demand can be added in months. The challenge now is making sure that flexibility shows up at the right time.
The hardest part is proving that a data center can cut power reliably. If the grid is counting on 50 megawatts to disappear, it needs to know how much the site would have used otherwise, then verify the reduction with meter data.
It also needs to know how long the cut can last and what happens when the GPUs ramp back up. Bring thousands of them back at once, and the data center could create a fresh power spike.
That makes customer contracts an essential but overlooked part of the equation. A data center could keep interactive and safety-sensitive work running normally while putting jobs like internal experiments, indexing, or overnight processing into a flexible tier.
Customers might pay less for that flexibility, while the grid pays the data center to deliver a predictable, measurable power cut when needed.
That would make one fact about AI impossible to ignore: not every computation is equally urgent. The industry already sorts work by price, speed, and compute cost, so electricity could become another variable in that calculation.
When the grid gets tight, one image might take longer to render or a training run might slip to tomorrow, while other services keep moving. Instead of treating every GPU cycle as equally important, data centers could start distinguishing between what needs to happen now and what can wait.
Luxor’s half-second power cut was the easy part. Doing this across thousands of GPUs, without breaking promises to customers and while delivering megawatts the grid can actually count on, will be much harder.
But that’s also where the idea gets interesting, because AI has a power problem and the grid has a flexibility problem, and data centers happen to be right in the middle. They're packed with machines doing work that can sometimes move by seconds, minutes, or hours without anyone noticing.
If operators can turn that flexibility into dependable power savings, AI’s enormous appetite for electricity could become something the grid can actually work with. That could make the next phase of the AI buildout as much about using power at the right time as finding enough of it in the first place.
The post AI data centers are learning the power trick Bitcoin miners mastered first appeared first on CryptoSlate.
