Data center energy demand grew 17% in 2025, according to the International Energy Agency, more than five times the 3% growth in global electricity use. Artificial intelligence is the main driver, and much of the new demand is still met by burning fossil fuels. The usual answer is to build more power plants, more transmission lines and more data centers.
Christina Delimitrou, a newly tenured associate professor at MIT, argues for a cheaper first step: make the servers we already have do their jobs properly. An MIT News profile published on 8 October 2026 describes how her group uses machine learning to schedule work, manage shared hardware and debug cloud software, so that operators get more computing out of existing machines. “If we can remove that bloating in a way that doesn’t compromise performance, then we won’t need to build as many new data centers,” she says.
This article explains the scale of the problem, the research behind her claims, what the published results actually show, and where AI-driven efficiency runs out. It ends with practical steps for IT teams that want to cut data center energy waste in their own estates.
Table of contents
- Why Data Center Energy Has Become an Environmental Threat
- The Researcher Rethinking Data Center Energy From the Inside
- The 15 Percent Problem: Idle Servers Waste Data Center Energy
- How Machine Learning Cuts Data Center Energy Waste
- Predicting Problems Before They Burn Data Center Energy
- Cutting Software Bloat to Save Data Center Energy
- The Research Record Behind the Data Center Energy Claims
- Other AI Levers on Data Center Energy
- Where AI Efficiency Stops Saving Data Center Energy
- How IT Teams Can Cut Data Center Energy Use Today
- Frequently Asked Questions About Data Center Energy and AI
- References
Why Data Center Energy Has Become an Environmental Threat
The pressure on the grid is no longer a forecast. It shows up in national statistics, in grid connection queues and in the climate reports of the largest technology companies.
Global demand is on course to double
The IEA’s Energy and AI report put global data center electricity use at about 415 terawatt-hours (TWh) in 2024, roughly 1.5% of all electricity consumed worldwide. It expects that figure to more than double, to around 945 TWh by 2030. The United States accounted for 45% of the 2024 total, China for 25% and Europe for 15%.
The agency’s April 2026 update said data center energy use rose 17% in 2025, and that consumption at AI-focused sites climbed faster still. Applied to the 2024 base, 17% growth adds roughly 70 TWh in a single year. The same update said the capital spending of five large technology companies passed $400 billion in 2025 and is set to rise by a further 75% in 2026.
The United States picture
A December 2024 report from Lawrence Berkeley National Laboratory, published by the US Department of Energy, tracks American data center energy use over time. It found consumption tripled from 58 TWh in 2014 to 176 TWh in 2023, when data centers used about 4.4% of US electricity. By 2028 the laboratory expects between 325 TWh and 580 TWh, or 6.7% to 12% of national demand.
The chart shows how steep that curve is. Each bar is scaled against the 580 TWh upper estimate.
Why emissions follow the electricity
Data center energy is only as clean as the grid behind it. The IEA estimates emissions from data center electricity at about 180 million tonnes of CO2 today, rising to 300 million tonnes by 2035 in its base case and up to 500 million tonnes in its high-growth case.
A June 2026 MIT study by economist Christopher Knittel and colleagues modelled three US grid regions. Compared with a world without new data centers, the growth projected by 2030 raised carbon dioxide emissions by 58% in Texas, 20% in the Mid-Atlantic and 24% in the western states. The local effects, from water use to growing piles of electronic waste, come on top of that.
| Measure | Figure | Source |
|---|---|---|
| Global data center electricity, 2024 | About 415 TWh (1.5% of world use) | IEA, Energy and AI |
| Growth in 2025 | 17%, against 3% for all electricity | IEA, April 2026 |
| Global projection, 2030 | About 945 TWh | IEA, Energy and AI |
| US share of US electricity, 2023 | 4.4% (176 TWh) | LBNL / DOE |
| US share of US electricity, 2028 | 6.7% to 12% | LBNL / DOE |
| Emissions from data center electricity | About 180 Mt CO2 today, 300 to 500 Mt by 2035 | IEA, Energy and AI |
The Researcher Rethinking Data Center Energy From the Inside
Most of the public debate about data center energy is about supply: where to find the gigawatts. Delimitrou works on the demand side, inside the racks. She is the KDD Career Development Associate Professor in Communications and Technology in MIT’s Department of Electrical Engineering and Computer Science (EECS), and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL).
From Greek geometry to cloud schedulers
Delimitrou grew up in a midsized town on the plains of northern Greece. She credits the country’s mathematical history, and her parents (a chemical engineer and a pharmacist), with her early interest in science. “In Greece, there is a long tradition of geometry,” she told MIT News.
She studied computer engineering at the National Technical University of Athens. Her diploma thesis looked at resource management on one computer running several applications at once. “A lot of the challenges I was looking at then would get much harder if, instead of a single system, you had 100,000 of these systems,” she says. That question became her career.
Stanford, Cornell and MIT
At Stanford, working with Christos Kozyrakis, she studied why large computing systems wasted so much capacity. She continued the work as an assistant professor at Cornell University, then joined MIT EECS in 2022. She now leads a research group there and teaches 6.191, Computation Structures, an undergraduate course with about 350 students each semester.
Her stated research focus has not changed much in fifteen years: improving the resource efficiency of large-scale data centers through quality-of-service-aware scheduling and resource management. What has changed is how central that question is to data center energy policy.
The 15 Percent Problem: Idle Servers Waste Data Center Energy
The core finding behind Delimitrou’s work is simple and uncomfortable. “You would expect, with all the demand for these systems, that they should be running close to 100 percent capacity. But we found that most were running at only about 15 percent capacity,” she says. “This is not a resource-efficient or sustainable way of scaling these systems.”
What the measurements showed
Her 2014 paper on the Quasar cluster manager, written with Kozyrakis, collected the evidence. In a Twitter production cluster managed by Mesos, aggregate CPU use stayed consistently below 20%, even though users had reserved up to 80% of total capacity. A 12,000-server Google cluster run by the more mature Borg system achieved 25% to 35% CPU use. Other analyses cited in the paper put industry-wide utilization at 6% to 12%, and one study estimated Amazon EC2 server use at 3% to 17%.
The chart shows the top of each reported range, so it flatters every case.
Why users reserve the wrong amount
The gap between reserved and used capacity is mostly human. Developers cannot predict how complex code will behave under load, so they ask for more than they need. The Quasar paper found that 70% of workloads overestimated their reservations, by up to ten times, while 20% underestimated them by up to five times. Both errors are costly: the first strands hardware, the second causes slowdowns that people then fix by adding more hardware.
Idle servers are not free
Low utilization would matter less if an idle server drew no power. It does not work that way. In their 2007 paper on energy-proportional computing, Google engineers Luiz André Barroso and Urs Hölzle wrote that “even an energy-efficient server still consumes about half its full power when doing virtually no work”. They found servers spent most of their time between 10% and 50% of maximum utilization, exactly where they are least efficient. Hardware has improved since then, but the lesson holds: spreading work thinly across many machines wastes data center energy.
The arithmetic of packing work tighter
The simplest way to see the stakes is to count servers. Suppose a workload needs the equivalent of 1,000 fully busy servers. At an average utilization of 15%, you need 1,000 divided by 0.15, or about 6,667 machines. At 50% you need 2,000. At 80% you need 1,250.
Fewer machines means less embodied carbon from manufacturing, less cooling, less floor space and fewer new buildings. Because a lightly loaded server still burns a large share of its peak power, consolidation saves data center energy as well as hardware. That is the logic behind Delimitrou’s line about not needing to build as many new data centers.
How Machine Learning Cuts Data Center Energy Waste
If the problem is that people misjudge what their software needs, the fix is to let the system learn it. “Applying machine learning to solve a large-scale system problem was a novel approach at the time. It was a bit risky because people had not yet shown that these techniques would work,” Delimitrou says. “But empirical approaches require a lot of expertise, and the scale of the system is so large that it is difficult for users to manage.”
Paragon: scheduling like a recommendation engine
Paragon, presented at the ASPLOS conference in 2013, tackled two problems at once: interference between applications sharing a server, and the mix of hardware types in a real data center. Instead of profiling every new job in detail, it used collaborative filtering, the technique behind film recommendations, to compare an unknown workload with ones it had already seen.
The authors tested it on systems of up to 1,000 servers on Amazon EC2. In a 2,500-workload scenario, Paragon met performance guarantees for 91% of applications while significantly improving utilization. Schedulers that ignored hardware differences, ignored interference, or simply picked the least-loaded server managed similar guarantees for only 14%, 11% and 3% of workloads.
Quasar: ask for performance, not for cores
Quasar, published in 2014, changed the contract between users and the cluster. Users no longer reserve raw resources. They state a performance target, and the system decides how many resources, of which type, are needed to meet it, adjusting as the workload changes.
On a 200-server EC2 cluster, Quasar improved resource utilization by 47% at steady state while meeting performance targets for every type of workload. In data center energy terms, that is close to half again as much useful work from the same machines.
Sinan: managing microservices
Cloud applications then changed shape. Developers began splitting them into dozens of small services spread across many servers. “But the servers were not built for this new style of application design,” Delimitrou says, so she rebuilt her earlier approach for that world.
Sinan, from 2021, uses machine learning models to predict how one microservice’s resources affect the whole application’s response time. It then allocates resources tier by tier without overprovisioning. On a hotel booking application, it used 25.9% fewer resources on average than other methods that met the same quality target, and up to 46%. On a more complex social network, it saved 59% on average and up to 68.1%, which the authors describe as handling about twice the requests per second without more hardware.
Predicting Problems Before They Burn Data Center Energy
Wasted capacity is one drain. Software failures are another. When a cloud application slows down, operators often respond by throwing more servers at it, and slow recovery means hardware runs for hours doing little useful work.
Seer: early warning for cloud applications
Seer, presented at ASPLOS in 2019, uses deep learning on the tracing data cloud systems already collect. It learns the patterns in time and across services that come before a quality-of-service violation, warns the cluster manager, and lets it act before users notice.
The paper reports that Seer anticipated violations 91% of the time and avoided them altogether in 84% of cases. In the authors’ local cluster it pinpointed the microservice that started the problem 89% of the time. MIT News says this averts the widespread slowdowns that can follow when a developer tries to fix a problem by hand.
Sage: no labelled data required
Seer had a practical weakness: it learned from traces labelled with the true cause of each failure, which production clouds rarely have. Sage, developed with Google and published in 2021, uses unsupervised learning instead. It builds a causal graph of how services depend on each other and tests possible fixes in a generative model.
In experiments on local clusters and on Google Compute Engine, Sage identified the root cause of performance violations with more than 93% accuracy. Because it needs no labelled failures, it is much closer to something a production operator could run.
From faults to security flaws
Delimitrou has since extended the same debugging approach beyond coding errors to security issues that can expose user data to attackers. That matters for data center energy too, because a compromised or unstable service wastes cycles. It also links her efficiency work to mainstream cybersecurity, where anomaly detection on traces is already a familiar tool.
Cutting Software Bloat to Save Data Center Energy
The second half of Delimitrou’s argument is about software. “There is a lot of bloating, especially on the software side of these systems,” she says. Code written for convenience, layered frameworks and network-heavy designs all consume cycles that do no useful work for the user.
Fitting software to the hardware
Her group now uses AI to redesign software systems so they better fit the hardware they run on. Earlier work showed how much microservices change the bottlenecks. A large share of each request’s time now goes on processing network messages rather than useful computation. The group’s Dagger project moved that networking work onto a reconfigurable chip, and its research page reports 1.3 to 3.8 times higher remote procedure call throughput per core than earlier hardware and software designs.
Ditto: cloning what academics cannot see
A growing obstacle is access. Early cloud research used commodity servers anyone could buy. Today’s hyperscalers run proprietary hardware and software that university teams cannot inspect, so a fix that works in the lab may fail in production.
Ditto, presented at ASPLOS in 2023 with co-authors from Meta and Intel, attacks that problem. It builds a synthetic clone of a cloud application that reproduces its CPU, memory, network and kernel behavior without revealing the original code. Companies can then share clones of production services with hardware vendors and researchers.
Open benchmarks
The group also maintains DeathStarBench, an open-source suite of realistic microservice applications such as a social network, a hotel reservation site and a movie review service. Its research page says the suite has more than 20,000 unique clones on GitHub. Shared benchmarks make claims about data center energy savings easier to test and compare.
The Research Record Behind the Data Center Energy Claims
The table gathers the main systems, with the results each paper reports. These are research results on test clusters, not audited savings from production fleets.
| System | Published | What the learning does | Reported result |
|---|---|---|---|
| Paragon | ASPLOS 2013 | Classifies new jobs by similarity to past ones | Guarantees met for 91% of 2,500 workloads |
| Quasar | ASPLOS 2014 | Sizes resources from a performance target | Utilization up 47% on 200 EC2 servers |
| Seer | ASPLOS 2019 | Deep learning on traces to predict slowdowns | 91% anticipated, 84% avoided |
| Sinan | ASPLOS 2021 | Allocates resources per microservice | 25.9% to 59% fewer resources on average |
| Sage | ASPLOS 2021 | Unsupervised root-cause analysis | Over 93% root-cause accuracy |
| Ditto | ASPLOS 2023 | Clones an application’s behavior, not its code | Matches CPU, memory and performance metrics |
Other AI Levers on Data Center Energy
Server scheduling is one layer. AI is being applied to the building, the storage, the software and the grid connection as well, and MIT has published several examples this year.
Cooling
The best-known case is from 2016. Google applied DeepMind’s machine learning to its data center cooling and reported a consistent 40% cut in the energy used for cooling. After electrical losses and other overheads, that equalled a 15% reduction in power usage effectiveness (PUE) overhead. Cooling also drives water use, which we covered in our look at what water usage effectiveness really shows.
Storage
In April 2026, a team led by MIT graduate student Gohar Chaudhry and associate professor Adam Belay described Sandook, a system that balances work across pools of solid-state drives. It raised application throughput by 12% to 94% and improved storage capacity use by 23%, reaching 95% of the drives’ theoretical maximum. “We want to be able to maximize the longevity of these very expensive and carbon-intensive resources,” Chaudhry said.
AI agents and workflows
The same team, working with Microsoft Azure, built Murakkab to configure multi-step AI agent workflows automatically. In tests it used about 35% of the computation of other methods and about 27% of the energy, as we reported in our story on the AI workflow tool that could cut cloud energy use.
Measuring before managing
Another MIT and IBM team built EnergAIzer, which estimates how much power an AI workload will draw on a given chip in seconds rather than hours, with about 8% error. Fast estimates let operators compare options on data center energy before they commit hardware. Chip choice matters too, which is why performance per watt has become a headline metric.
Timing and the grid
Knittel’s June study found that letting data centers shift more than 20% of their load to off-peak hours could cut power system costs by up to 5% in Texas. Other research on data center flexibility points the same way.
| Layer | Example | Reported effect | Main limit |
|---|---|---|---|
| Servers and scheduling | Quasar, Sinan | 47% higher utilization; up to 68.1% fewer resources | Needs control of the whole cluster |
| Cooling | DeepMind at Google | 40% less cooling energy | Cooling is only part of the load |
| Storage | Sandook | 23% better drive capacity use | Tested on a 10-drive pool |
| AI workflows | Murakkab | About 27% of the energy of other methods | Agent workloads only |
| Grid timing | Flexible load (Knittel) | Up to 5% lower system costs in Texas | Can raise emissions on some grids |
Where AI Efficiency Stops Saving Data Center Energy
None of this makes the environmental threat disappear. The same sources that report the gains also show where they end.
Efficiency can be eaten by demand
The IEA says the power needed per AI task is falling at a rate unprecedented in energy history, yet total data center energy use still rose 17% in 2025, because more people use AI and energy-hungry uses such as agents are spreading. Economists call this the rebound effect. Cheaper computing invites more computing, so efficiency on its own may slow emissions growth without reversing it.
Black boxes are hard to trust
“One of the challenges when it comes to applying AI to these systems is that the AI is not interpretable,” Delimitrou says. An operator who cannot see why a scheduler packed two services together will hesitate to hand it a production fleet. Her group now builds explainability into its tools, so engineers can check the answers and learn how to design better systems.
Closed hardware limits outside research
Hyperscalers now design their own chips, servers and software stacks. University researchers cannot test on them, which is why tools like Ditto exist. It also means most of the largest potential savings sit inside a few companies, whose disclosures vary. Our coverage of how the AI race weakened climate pledges at Google and Amazon shows how far promises and practice can drift.
Flexibility is not automatically green
Knittel’s study found that in the Mid-Atlantic, flexible data centers could raise system-wide CO2 emissions by about 3%, because shifted load can keep a coal plant running. “That’s why we have policy,” he said.
AI still needs auditing
Delimitrou is careful about her own tools. “You still have to use AI carefully. While it can greatly accelerate the application development side, we still need to audit it,” she says.
How IT Teams Can Cut Data Center Energy Use Today
You do not need a hyperscale fleet to apply these lessons. Most organizations run some mix of on-premises servers, colocation and public cloud, and all three waste capacity in the same ways.
Measure utilization honestly
Start with average and peak CPU and memory use per host and per virtual machine over at least a month. Many estates look busy because reservations are high, while actual use sits in the range the Quasar paper described.
Set targets in performance, not in hardware
Define service level objectives for latency and throughput, then let autoscaling and right-sizing tools find the smallest footprint that meets them. That is the Quasar idea in commercial form.
Use tracing to find waste and faults early
Distributed tracing and anomaly detection catch slowdowns before they turn into emergency scale-outs. Treat repeated over-provisioning as a defect to fix in the code.
Consolidate and retire
Move lightly used workloads onto fewer hosts and switch off what is left. Our data center operations and cost optimization teams can help plan that work.
| Step | What to measure | Why it cuts data center energy |
|---|---|---|
| Baseline utilization | CPU and memory per host, 30 days or more | Shows stranded capacity |
| Right-size reservations | Requested vs used resources per workload | Fewer idle machines drawing power |
| Performance-based scaling | Latency and throughput targets | Capacity follows demand |
| Tracing and anomaly alerts | Request paths and error rates | Stops panic scale-outs |
| Consolidation | Hosts under 20% average use | Fewer servers, less cooling |
| Supplier questions | PUE, utilization and carbon data | Moves work to efficient sites |
Frequently Asked Questions About Data Center Energy and AI
Who is Christina Delimitrou?
She is a newly tenured associate professor in MIT’s Department of Electrical Engineering and Computer Science and a member of CSAIL. Her group applies machine learning to make large cloud systems more efficient, secure and reliable.
How much data center energy is wasted on idle servers?
There is no single figure. Studies cited in her Quasar paper put average server utilization between 3% and 35%, depending on the operator, and she says most systems she studied ran at about 15%. A 2007 Google study found that even an efficient server drew about half its peak power while doing almost no work.
Can AI really reduce data center energy use?
Yes, at the level of individual systems. Published results include 47% higher utilization (Quasar), up to 68.1% fewer resources (Sinan) and 40% less cooling energy (DeepMind at Google). Whether total use falls depends on how fast demand grows.
Does AI make the problem worse as well as better?
It does both. AI is the main driver of new data center energy demand, and the same machine learning techniques are among the best tools for using that capacity efficiently.
What should a business do first?
Measure real utilization across servers and cloud accounts, then right-size reservations against performance targets. It is usually the cheapest and fastest saving available.
References
Using AI to mitigate the growing environmental threat of data centers (MIT News)
Data centre electricity use surged in 2025 (IEA)
Energy and AI: Executive Summary (IEA)
DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers
Paragon: QoS-Aware Scheduling for Heterogeneous Datacenters (ASPLOS 2013)
Quasar: Resource-Efficient and QoS-Aware Cluster Management (ASPLOS 2014)
Sinan: Data-Driven, QoS-Aware Cluster Management for Microservices
Sage: Practical and Scalable ML-Driven Performance Debugging in Microservices
End-to-End Application Cloning for Distributed Cloud Microservices with Ditto
Christina Delimitrou: Research
The Case for Energy-Proportional Computing (Google Research)
DeepMind AI Reduces Google Data Centre Cooling Bill by 40%
How data centers can better manage energy use (MIT News)
Helping data centers deliver higher performance with less hardware (MIT News)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
