Every response generated by an AI chatbot begins with electricity. While the text appears instantly on your screen, the heavy lifting occurs in distant facilities filled with computer chips. These components draw electricity, process data, and generate significant heat that requires intensive cooling systems within AI data centers.
When this process scales across millions of daily prompts, image requests, and business workloads, a seemingly weightless digital query transforms into a tangible physical demand on local power plants and electrical grids.
Electric utilities face requests to supply this energy in massive, concentrated blocks. A typical large data-center campus can consume as much electricity as a small city, and tech companies can build and deploy these facilities far more quickly than utilities can upgrade their infrastructure.
Utilities must also maintain reserve capacity for peak demand hours, even if some of that power sits idle during normal operations. Ultimately, data centers are demanding power much faster than the electrical grid can deliver it.
While building new power plants is one option, it is an extremely costly and lengthy process that can require billions of dollars and take years to become fully operational.
However, another option is to shift certain computing tasks to different times of day.
Although a chatbot response generally needs to be delivered immediately, internal testing or an overnight video-rendering queue can afford to wait. Intelligent software can differentiate between these tasks, dialing back flexible workloads when electricity is scarce and ramping them back up when surplus power becomes available.
A recent, small-scale experiment in Texas demonstrates how this approach functions in practice.
Luxor Energy, a firm with origins in Bitcoin mining, partnered with Bentaus, a developer of chip-level power management software. Together, they regulated a single Nvidia B200, an advanced high-performance AI chip.
While the processor was handling inference—the process of using a trained AI model to generate an output—the software instructed it to reduce its electrical consumption.
The project partners reported that the chip’s power draw dropped to roughly 25% of its normal baseline within half a second, resulting in fewer processed requests during the restriction period.
Ethan Vera, chief operating officer at Luxor, noted to CryptoSlate that no computational jobs failed and no active work-in-progress was lost. The processor seamlessly returned to full operating speed once the restriction was lifted.
While Luxor and Bentaus characterized the public demonstration as entirely smooth, that description requires nuance. From the operator’s viewpoint, the task survived the throttling and resumed full performance afterward.
Nevertheless, end users may experience longer wait times because the processor completes less work during a restriction window. Implementing flexible AI infrastructure will ultimately depend on the frequency of such delays, who experiences them, and the service-level agreements promised to customers.
Although the trial succeeded, it involved just one isolated chip. Full-scale data centers house tens of thousands of processors alongside servers, cooling infrastructure, storage arrays, and networking gear.
Despite this scale, the test highlights a broader concept: modern AI data centers could categorize workloads by urgency and strategically reduce their grid draw when necessary.
Texas lacks power to feed the computers waiting
Texas serves as the prime example of what happens when data center development outpaces local power infrastructure.
The Electric Reliability Council of Texas (ERCOT) manages the electrical grid covering most of the state. On July 22, electricity consumption reached a preliminary record of 91,089 megawatts, a figure that remains unofficial pending final data validation.
ERCOT estimates that one megawatt can power approximately 250 residential homes during peak hours. Using this metric, that record consumption matched the simultaneous demands of more than 22 million residential customers.
Gov. Greg Abbott stated in August that ERCOT was reviewing interconnection requests representing over 474 gigawatts of new power demand, with data centers accounting for roughly 90% of that total. Because one gigawatt equals 1,000 megawatts, this queue represents more than five times the electricity used during ERCOT’s peak record hour.
In response, Abbott directed regulators to audit these pending projects before allowing them to proceed.
A preliminary ERCOT review on July 28 indicated that roughly 205 gigawatts had sufficient supporting studies to enter the initial study batch—less than half of the total 474-gigawatt queue. Governor Abbott’s audit temporarily halted that evaluation.
Regulators granted ERCOT additional time on August 20, with the grid operator announcing plans to issue conditional eligibility decisions by August 31. Industry developers frequently submit overlapping proposals, reserve capacity for projects that ultimately lack financing, or request grid access at multiple sites for a single planned campus.
Texas initiated the audit partly because the connection queue had become detached from physical grid constraints and real-world feasibility.
Even with those caveats, the 474-gigawatt total underscores an intense land grab for locations with access to massive blocks of power, far outpacing the deployment of physical transmission lines.
An update published this year by the Lawrence Berkeley National Laboratory estimates that data centers could consume between 9.5% (low estimate) and 15.3% (high estimate) of total US electricity by 2030, with a mid-range projection of 11.8%. Meanwhile, the International Energy Agency projects that data centers will drive roughly half of the total growth in US electricity demand through the end of the decade.
Despite this surging demand, constructing high-voltage transmission lines in developed economies typically takes four to eight years. The International Energy Agency reports that lead times for critical equipment like transformers and cables have doubled over the last three years.
While AI firms measure capacity in individual chips, electrical utilities must think in terms of entire municipalities. An individual B200 chip can consume up to 1,000 watts, while Nvidia lists the maximum power consumption of a complete eight-GPU DGX B200 server at approximately 14.3 kilowatts. With one megawatt equating to 1,000 kilowatts, Texas’s regulatory threshold for very large electricity consumers begins at 75 megawatts.
Based on ERCOT’s residential comparison, 75 megawatts could supply roughly 18,750 homes during peak hours, or power 75,000 one-kilowatt GPUs—before accounting for supporting processors, cooling systems, networking hardware, battery backups, and electrical transmission losses.
Consequently, mastering how to control and throttle the energy consumption of a single chip represents an essential foundational step toward managing power efficiency across an entire data center.
The immense hardware and software complexity involved is why grid operators traditionally classify data centers as “firm loads,” requiring a continuous and uninterrupted power supply on demand.
Data center operators prefer running expensive GPUs continuously because every idle minute delays billable customer computations. Furthermore, thousands of chips collaborating on a single large AI workload are deeply interdependent.
If one group of processors is slowed down, bottlenecks can ripple across adjacent hardware.
Nonetheless, not all data center computations require immediate processing at maximum speeds. Certain tasks are time-critical, while others can tolerate delays or reduced processing speeds with minimal impact, or even be migrated to alternative data facilities where electricity is more readily accessible.
Each operational choice involves trade-offs, but collectively they can significantly lower a data center’s immediate draw on the local power grid.
Bitcoin miners taught computers how to yield
This operational precedent largely originates from Bitcoin mining operations on the Texas power grid. Bitcoin miners earn financial rewards by running hardware that performs continuous cryptographic calculations. When a machine is powered down, the operator forfeits mining revenue for that duration.
However, once electricity is restored, the hardware can resume operations almost instantly without waiting for end-user responses or preserving unfinished computing states.
Texas utilities recognized this model as a form of demand response: when wholesale electricity becomes scarce and expensive, large power consumers are financially incentivized to reduce consumption.
Bitcoin miners proved uniquely well-suited for this arrangement. They could power down during wholesale price spikes, earn financial compensation for shedding load during grid emergencies, and lower their transmission fees by sitting out a handful of critical summer peak hours.
An ERCOT review published in April identified cryptocurrency miners as the primary price-responsive participants in one of its emergency demand-response programs. For a miner, the economic calculation is straightforward: if a megawatt of electricity becomes more valuable than the Bitcoin generated by running the hardware, it makes financial sense to shut the machines down.
Because Luxor provides software, energy management, and financial products to Bitcoin miners, the company approached the AI sector with an established understanding of interruptible computing. Their experiment tests whether customer-facing AI hardware can adopt a similar responsiveness to electricity pricing.
This question gains urgency as crypto miners convert power-rich mining sites into AI data centers. If the electrical grid swaps a flexible Bitcoin mining facility that can shut down instantly for a data center that operates continuously around the clock, it risks losing a vital emergency load-shedding resource.
The degree of operational flexibility a data center can achieve depends heavily on its specific computational workloads.
Model training is a prolonged, compute-intensive process that repeatedly adjusts a model as it processes massive datasets. Inference occurs subsequently when a user queries the trained model for answers, images, translations, or predictions. These two phases offer distinctly different opportunities for power reduction.
Long training runs can often be paused at saved checkpoints and resumed later, though synchronizing the shutdown of thousands of machines presents significant engineering hurdles. Conversely, inference consists of millions of smaller, discrete requests—some originating from users expecting instant answers, and others from automated background jobs that can wait in queues when electricity is expensive or constrained.
Google has categorized its workloads using this approach for years. In 2023, the company detailed how it could defer batch workloads like YouTube video processing when local grids experienced strain, or reroute those tasks to alternative regions with surplus power capacity. Core services like Search and Maps remained continuously available online.
Google subsequently extended this strategy to machine-learning workloads. By March 2026, the company announced it had secured one gigawatt of data center demand-response capacity under long-term utility agreements across multiple US regions.
Some of these commercial arrangements could help new data centers connect to the electrical grid much faster.
Researchers are demonstrating that this strategy is viable outside theoretical simulations. In a peer-reviewed paper published in Nature Energy, a research team detailed an experiment at an Oracle cloud facility in Phoenix. Software successfully reduced the power consumption of a 256-GPU cluster by 25% for three hours without compromising priority workloads beyond agreed performance thresholds.
The key was intelligent workload scheduling: the management software protected urgent tasks while absorbing power reductions from jobs with more flexible deadlines.
Emerald AI, the firm that led that research initiative, announced a $150 million financing round on August 25 that valued the company at over $1 billion. The firm also reported that its software is operating commercially across entire data centers, managing multi-megawatt loads.
While independent performance data across every facility is unavailable, this venture funding indicates that flexible AI operations have transitioned successfully from academic research into viable commercial enterprises.
Other academic researchers have attempted to quantify how much electricity an AI facility can reliably curtail during grid-stress events.
A working paper from the University of Chicago analyzed four years of historical electricity pricing and 49.4 million real-world inference requests to model potential savings. The study estimated that an inference-focused facility could commit to reducing its power draw by 40%, while a facility running a hybrid mix of inference and training could curtail 24.6%.
These percentage reductions declined only marginally when the economic model scaled up to a 10-gigawatt fleet. Primary constraints were driven by customer contract terms, restrictions on workload migration, and the electrical surge caused when processors ramp back up to full capacity.
Simulations conducted by University of Alberta researchers evaluated grid impacts when AI workloads are delayed or migrated between data centers. Under the model’s most severe stress scenario, workload flexibility reduced required power-plant generation capacity by more than 21%. In another scenario featuring localized grid congestion, it lowered total electricity supply costs by 3.5%, even though capital spending on new power generation rose by 7.1%.
The majority of benefits from delaying workloads materialized within the first three hours, indicating that extended deferrals yielded diminishing returns. Although these measures do not eliminate the long-term necessity of building new power plants and transmission lines, they demonstrate that the grid can accommodate higher AI demand with less infrastructure and at a lower overall cost.
Four hidden moments can price an entire year
The economic incentive behind Luxor’s experiment stems from a unique mechanism within the Texas electricity market. Large commercial customers contribute to the costs of high-voltage transmission networks, and a portion of those transmission bills is determined by their electricity consumption during just four 15-minute intervals throughout the entire year.
These crucial windows represent the moments of highest system-wide electricity demand in June, July, August, and September, officially designated as the Four Coincident Peaks (4CP).
Because the exact timing of a 4CP interval cannot be known with certainty until the month concludes, large industrial power users hire meteorological and grid forecasters to monitor weather patterns and grid demand to predict peak windows.
If forecasted probabilities are sufficiently high, these consumers curtail their electricity use during those specific 15-minute windows. Accurate predictions yield substantial savings on annual transmission network charges, turning 4CP into a high-stakes game of prediction and load shedding for industrial plants, crypto mines, battery operators, and potentially AI data centers.
This substantial financial payoff makes tolerating occasional false alarms economically viable. According to 2026 Public Utility Commission of Texas (PUCT) figures, ERCOT transmission costs total roughly $6 billion, distributed across an average 4CP demand baseline of 80,859.8 megawatts.
This equates to approximately $74.89 per kilowatt annually. At that rate, maintaining 100 megawatts of demand across the four peak windows translates to about $7.49 million in yearly transmission expenses.
While actual billing calculations vary across utility territories and individual contracts, the financial incentive is clear. A large data center can have millions of dollars riding on just one hour of aggregate electricity consumption spread across a summer. Curtailing power usage for a few extra hours to successfully capture those peak windows is a highly advantageous trade.
In their test, Luxor throttled the GPU directly by utilizing live grid data to time their intervention. Vera explained that the company monitored indicators suggesting a 4CP window might be forming and independently transmitted curtailment commands directly to the processor, without instructions from ERCOT or participation in formal emergency grid programs.
This execution was essentially a private commercial wager on peak grid demand designed to lower facility transmission fees. ERCOT classifies this type of 4CP self-curtailment separately from its standardized demand-response programs.
This dynamic also contextualizes the half-second response time. Because a 4CP window spans 15 full minutes, whether a GPU reaches a lower power draw in half a second or several seconds makes virtually no difference to the resulting transmission savings.
By contrast, ERCOT’s formal emergency response programs typically grant participating customers 10 to 30 minutes to achieve pledged power reductions. Specialized grid stabilization services operate more quickly, requiring participants to begin shedding load immediately and reach full reduction within 10 minutes.
Should AI hardware eventually participate in those ancillary markets, sub-second hardware control could offer greater operational value. For Luxor, however, every extra second a GPU spends throttled represents time it could have spent generating revenue through computations.
Bentaus has previously demonstrated similar capabilities at a larger scale. In February, CPower, Bentaus, and Supermicro showcased a California demonstration utilizing a server cluster equipped with B200 GPUs.
Those project partners reported that the server cluster responded to a wholesale electricity market signal in less than 20 milliseconds, slashing its power consumption by up to 75% while maintaining contracted performance benchmarks.
While the Texas test was narrower—involving a single GPU responding to a transmission-billing incentive—it adds another real-world validation to a concept shifting rapidly from individual processors to multi-node clusters and utility-scale programs.
Important details regarding the Texas demonstration remain undisclosed. The participating companies did not reveal which specific AI model was executing, the exact duration the GPU remained throttled, its initial baseline power draw, the magnitude of throughput reduction, or the impact on user response latency.
Although Luxor’s representative in the Texas electricity market verified the recorded power drop, no independent third-party analysis of the test has been published.
The trial proved that a single B200 running an inference workload can absorb a steep power restriction without dropping active computations. However, it remains unverified how this impacted user wait times, whether identical methods would apply to varied training or inference workloads, or how much energy the technique could save across an entire facility.
A GPU represents only one component of a data center’s total energy bill. Cooling infrastructure, networking switches, storage hardware, fluid pumps, and power conversion units also consume substantial electricity. Consequently, cutting an individual chip’s power draw by 75% does not mean the overall data center facility draws 75% less power from the utility grid.
Total reduction measured at the main building meter could prove considerably smaller.
Luxor is already preparing a subsequent test utilizing a cluster of Nvidia H100 GPUs in Texas. Vera noted that scaling up requires developing orchestration software capable of identifying which workloads can safely slow down while coordinating interdependent hardware units, all while respecting customer performance agreements.
Each step up in scale—from a single GPU to server blades, rack units, and entire data centers—introduces compounding complexity. Expanded equipment draws more power, interconnected machines must synchronize movements, and diverse customer workloads may exhibit varying tolerance for operational slowdowns.
Regulatory frameworks in Texas are beginning to mandate this type of operational flexibility. Senate Bill 6, enacted in 2025, requires specific large electricity consumers connecting to the grid from 2026 onward to curtail consumption during severe grid emergencies. It also establishes a compensation program paying facilities consuming at least 75 megawatts to reduce demand during anticipated grid strains.
Concurrently, state regulators are reevaluating the 4CP mechanism itself. Because the four summer peak hours can overlook evening and winter periods when the grid experiences greater stress, regulators have proposed transitioning to a 12CP model, which would base transmission fees on a single 30-minute peak demand interval each month.
ERCOT reached a similar conclusion in its April review: while Texas possesses ample demand-response capacity, it does not always materialize when the grid requires it most. Although 4CP drives significant load shedding, its summer-centric peaks miss periods when overall electricity demand is high and intermittent wind and solar generation are low.
ERCOT identified this structural mismatch as a major challenge. Because constructing new power plants and transmission lines requires years, flexible demand resources can be integrated within months. The core challenge lies in ensuring that flexibility is deployed at precisely the right moments.
The most demanding aspect is verifying that a data center can curtail power reliably. If grid operators rely on 50 megawatts of load shedding, they must establish baseline consumption metrics and independently verify reductions using utility meter data.
Operators must also predict how long curtailments can be sustained and manage the electrical surge when processors ramp back up. Restoring thousands of high-power GPUs simultaneously risks creating an abrupt secondary power spike.
This dynamic makes customer contracts a critical component of the equation. Data centers could maintain continuous operation for interactive, safety-sensitive workloads while shifting non-urgent tasks like internal testing, data indexing, or overnight batch processing into flexible tier queues.
Customers could receive discounted rates for accepting flexible execution timelines, while grid operators compensate data centers for delivering predictable, verifiable power reductions during grid emergencies.
This operational reality reinforces a broader truth about artificial intelligence: not all computational tasks carry equal urgency. Because the tech industry already prices workloads based on speed, cost, and compute resources, electricity consumption may become another core variable in that economic calculation.
When electrical grids tighten, an image rendering task or training batch could be deferred to off-peak hours while essential services remain uninterrupted. Rather than treating every GPU cycle as equally vital, data centers can begin distinguishing between immediate computational needs and deferrable tasks.
Luxor’s half-second power throttling demonstration was the straightforward part. Scaling this capability across thousands of GPUs without violating customer agreements—while delivering dependable megawatts that grid operators can actually count on—will prove significantly more challenging.
Nevertheless, this intersection creates a compelling opportunity: AI faces a severe power challenge, the electrical grid faces a flexibility crisis, and modern data centers sit directly at the convergence of both. These facilities house vast arrays of computational hardware performing tasks that can frequently be shifted by seconds, minutes, or hours without disrupting end users.
If data center operators successfully convert that inherent operational flexibility into dependable energy savings, AI’s immense appetite for electricity can become manageable for utility grids. This transformation could make the next phase of AI infrastructure development as much about consuming energy at optimal times as it is about securing raw generation capacity.
Frequently Asked Questions
What is 4CP and why does it matter to data centers?
4CP stands for Four Coincident Peaks, which are the four 15-minute intervals of highest system-wide electricity demand in Texas during June, July, August, and September. Large power users pay transmission network costs based on their consumption during these exact windows. Cutting power during these predicted peaks can save companies millions of dollars annually.
How did Luxor and Bentaus control the GPU’s power?
The companies used specialized software paired with live grid data to instruct a single Nvidia B200 GPU to draw less electricity while performing AI inference. The chip’s power draw dropped to roughly 25% of normal within half a second without losing active work or failing jobs.
Why can’t all AI computing be delayed during grid shortages?
While background tasks, internal experiments, and overnight batch processing can wait, interactive applications like real-time chatbot responses require immediate processing. Furthermore, thousands of GPUs working together on complex tasks are tightly interdependent, meaning slowing down one group can cause processing bottlenecks across the entire cluster.
