The International Energy Agency published the report Key Questions on Energy and AI and almost all the coverage stayed with the noisy half: data centres. The other half, the one that talks about energy savings in industry, has barely been told.
What you’ll find
- 950 TWh: what data centres will consume in 2030
- 13 exajoules of energy savings in industrial processes
- Between 3 and 10 points less energy cost
- 40% of agentic AI projects will be cancelled before 2028
- 563 industrial decarbonisation projects in Europe
- Why energy savings don’t happen on their own
- What this means in a real plant
- Physics-informed models
- The 2026 regulatory calendar
- Frequently asked questions
The report was published on 16 April 2026, and its first half is the one that has circulated: data centres will go from 485 TWh in 2025 to around 950 TWh in 2030. The second half matters much more to a factory.
That second half says that applying artificial intelligence to optimise production processes can cut energy costs by between 3 and 10 points and generate energy savings of more than 13 exajoules by 2035, equivalent to 3% of the planet’s final energy consumption.
AI is set to consume close to 3% of the world’s electricity and, applied well, can save 3% of the world’s final energy.The two halves of the same IEA report
Here are the five figures worth keeping to hand in 2026 and, above all, what it takes for those energy savings to reach an actual plant rather than stay on a slide.
01. 950 TWh: what data centres will consume in 2030
Data centre electricity consumption grew 17% in 2025, and AI-focused centres’ consumption jumped 50% in that same year. The IEA projects total consumption will double to 950 TWh by 2030, around 3% of global electricity demand.
The money follows that curve. Capital expenditure from just five tech companies exceeded $400 billion in 2025, and the IEA expects it to rise a further 75% in 2026. That’s an enormous figure, and it explains why the public debate about AI and energy has become almost a debate about cooling, grids and connection permits.
For an energy-intensive industry, however, that headline changes nothing day to day. What does change day to day is in the next figure.
02. 13 exajoules of energy savings in industrial processes
The IEA calculates that widespread deployment of AI applications that already exist to optimise industrial processes could deliver energy savings of more than 13 exajoules by 2035. That’s 13 billion gigajoules: 3% of the world’s final energy consumption.
In the previous report, Energy and AI (April 2025), the agency put the same idea another way: the possible energy savings are equivalent to more than Mexico’s entire current energy consumption. Two different reports, the same conclusion.
It’s worth underlining the nuance that usually gets lost: the IEA isn’t talking about technology yet to be invented. It’s talking about existing applications that are used today in a handful of plants and haven’t been rolled out more widely. The bottleneck for energy savings isn’t the algorithm. It’s deployment.
Key takeaway
The potential isn’t in some future technology. It’s in deploying, in your own plant, what already works in others.
03. Between 3 and 10 points less energy cost
The range the IEA gives for AI-driven process optimisation is a reduction of between 3 and 10 percentage points in energy cost. In a plant with €20 million in revenue that spends 15% on energy, each point of energy savings is €30,000 a year.
That range matches what’s seen in the field. The industrial lighthouses analysed by McKinsey report average energy savings of 10 to 25%, and specific cases such as Arçelik document 20%. The range is wide because it depends on the process, not the software brand.
The practical conclusion is simple: energy savings are measured process by process. A reheating furnace, a compressor, a drying line and an industrial chiller plant each have different curves and different margins. Promising a single energy-savings percentage for the whole factory is the first sign that someone hasn’t looked at the data.
04. 40% of agentic AI projects will be cancelled before 2028
Gartner forecast that more than 40% of agentic AI projects will be cancelled before the end of 2027, due to rising costs, unclear business value and insufficient risk controls. This isn’t a figure about the technology: it’s a figure about how these projects are being bought.
The same pattern repeats in industry. According to Deloitte data reported by Manufacturing Dive, 87% of manufacturers have launched at least one generative AI pilot and three out of four companies plan to deploy agentic AI within two years, but only one in five has its data model ready to sustain it.
In other words: there are far more pilots than systems in production. And the reason is almost never the model’s capability, but siloed data, signal quality and the gap between the operations department and the systems department.
05. 563 industrial decarbonisation projects in Europe
The EU Innovation Centre for Industrial Transformation (INCITE), managed by the Joint Research Centre (JRC), identified in June 2026 a total of 563 demonstrator projects with innovative techniques ready or close to market for decarbonising energy-intensive industry.
Three sectors account for 65% of those projects: iron and steel, cement and lime, and chemicals including fertilisers. The first two generate more than 40% of greenhouse gas emissions from all of the EU’s energy-intensive industries.
The takeaway for an industrial SME is twofold. There’s money and political focus on decarbonising processes. And there’s a clear bias towards big technological leap projects (hydrogen-based steel, lower-clinker cement) that leave out most existing plants: the ones where energy savings have to come from the furnaces already installed.
Why energy savings don’t happen on their own
Plant data rules
A model that optimises a thermal process needs reliable signals: temperatures, flow rates, composition, cycle times, output quality. In many plants that information exists, but it’s scattered across the SCADA, the historian, the ERP and a spreadsheet maintained by one specific person.
Before talking about models, it’s worth reviewing what’s measured, how often, and with what error. We’ve covered this in this guide on industrial data: without signal traceability there’s no sustainable energy saving, only a nice-looking pilot.
The model degrades
An analysis published in Nature (2022) covering 128 machine learning models found performance degradation over time in 91% of cases. In a plant this has a name: the raw material changes, a burner is replaced, a setpoint is adjusted, and the model stops getting it right.
That’s why energy savings aren’t a project that gets delivered and closed. They’re a system that’s monitored, retrained and audited. Anyone selling a model without a maintenance plan is selling half the problem.
What this means in a real plant
At Deduce Data Solutions we work exactly in that space: existing thermal processes, plants that are already built, and data that was never designed with a model in mind. The GASAI system, developed with Reinosa Forgings & Castings, is aimed at reducing natural gas consumption in industrial furnaces.
Industrial validation was carried out on the five furnaces of a main forging line, under SODERCAN’s INNOVA programme. The technical detail, regulatory context and sources are in the Reinosa case study.
The context helps explain why gas is the lever with the most room to move. According to Eurostat, the iron and steel sector consumed 907 PJ in 2023, and natural gas accounts for 31.3% of industrial energy consumption in the EU. In Spain, according to IDAE, industry accounts for 52% of natural gas use.
On top of that is the price of CO₂. With emission allowances moving in the range of €70 to €100 a tonne and the CBAM now in force, every point of energy savings today pays back twice over: the bill, and the allowance you don’t have to buy.
How to bring an AI energy-savings project down to earth
If the IEA’s figures describe the potential and Gartner’s describe the failure, the difference between the two lies in the method. This is the order that works in heavy industry:
- Choose a process, not a factory. A furnace, a line, a compressor. With measurable consumption and a named owner.
- Audit the signal before the algorithm. What sensors exist, how often they record, what gaps the historical data has, and what error each measurement carries.
- Set the baseline using your own data. Without an honest baseline, any energy-savings percentage is impossible to prove in an audit or for a public grant.
- Incorporate the physics of the process. Energy and mass balances constrain what the model can propose and rule out setpoints that are impossible on the plant floor.
- Validate in parallel before touching setpoints. The model recommends, the operator decides, and it’s compared against normal operation for weeks, not days.
- Plan retraining from day one. With alert thresholds and a named owner for follow-up.
None of these six steps requires replacing the furnace. That’s exactly the point: the energy savings the IEA describes are achieved in existing installations, not only in the new plants that make the headlines.
Key takeaway
An energy-savings project without a documented baseline isn’t a project: it’s a demonstration.
Physics-informed models: why they matter in a furnace
A purely statistical model learns only what it has seen. If a situation isn’t in the historical data, the model extrapolates it without judgement and can propose a setpoint that breaks an energy balance or a metallurgical limit. In a reheating furnace that’s not a precision error: it’s a defective part.
Physics-informed neural networks (PINNs) build the process equations into the training itself. The model doesn’t just minimise error against the data; it also penalises solutions that break known physics: heat transfer, mass conservation, combustion kinetics.
This has two very practical consequences for energy savings. The first is that less historical data is needed to get a usable model, which is decisive in plants with limited instrumentation. The second is that the recommendations are explainable to a production manager, because they rest on quantities they already know and monitor.
Academic literature on physics-informed machine learning in manufacturing has grown strongly in recent years, with specific reviews covering forming processes, heat treatment and additive manufacturing. It is, for now, the most solid route for bringing energy optimisation to processes where errors carry a physical cost.
Agentic AI and a process model are not the same thing
Much of 2026’s noise comes from mixing up two different things. A conversational agent that looks up manuals, drafts reports and coordinates tasks solves an information problem. A process model that anticipates a furnace’s temperature solves a physical problem. They’re bought, validated and maintained differently.
That distinction explains Gartner’s figure. When an organisation applies the buying method for an office tool (a quick pilot, a flashy demo, no baseline) to a physical problem, the project ends up cancelled even when the technology was appropriate.
For an energy-savings goal, the starting point is always the process and its instrumentation. The conversational layer can come later, and it helps a lot with day-to-day operation, but it doesn’t replace the model that understands the furnace.
The regulatory calendar pushing energy savings
The 2026 regulatory context strengthens the economic case. The Carbon Border Adjustment Mechanism (CBAM) entered its definitive phase on 1 January 2026, with declaration and certificate-purchase obligations for imports in sectors such as iron, steel, cement, aluminium and fertilisers.
In parallel, the European AI Act has applied generally since 2 August 2026. The digital omnibus package pushed back the obligations for Annex III high-risk systems to 2 December 2027, and those for AI embedded in regulated products, such as machinery, to 2 August 2028.
That postponement has been read as a breathing space, but it’s worth understanding properly: what’s delayed is the formal requirement, not the need for traceability. A system that acts on plant setpoints needs a record of decisions, human oversight and model documentation, both for the future audit and to justify energy savings for a grant.
In practice, plants that already document their baseline and retrainings will reach 2027 without surprises. Those that don’t will have to reconstruct, after the fact, a history nobody kept.
What to ask before signing a project
These five questions separate a serious energy-savings project from a well-staged demonstration:
- Against what baseline will the savings be measured, and who validates it?
- What specific signals does the model need, and which are missing from the plant today?
- How does the system behave in a situation that isn’t in the historical data?
- What happens when the raw material changes or a piece of equipment is replaced?
- Who keeps the model, the data and the documentation once the contract ends?
If the answers to these five questions are clear and in writing, the project has a good chance of ending up in production. If they’re vague, it will most likely swell Gartner’s 40%.
The gap left by the data-centre debate
There’s an interesting asymmetry in the public conversation of 2026. AI’s consumption gets discussed on every front page; the energy savings AI can bring to industry barely gets a mention, despite being in the same report and of a similar order of magnitude.
That asymmetry has practical consequences. A management committee that has only read the headlines associates AI with energy cost and reputational risk. A committee that has read the full report also associates it with a lever for competitiveness against the CBAM and the electricity bill.
The difference between the two readings is five figures. And, above all, the willingness to look at your own plant’s data before commenting on everyone else’s.
How much energy could you be saving in your process?
We’ll review a specific process with you, its data and its baseline, before talking about models.
Frequently asked questions about AI-driven energy savings
How much energy saving is realistic in a thermal process?
The IEA puts the energy-cost reduction at between 3 and 10 points; McKinsey documents 10 to 25% in reference plants. In combustion processes with high operational variability, the energy-savings margin tends to be larger. The only way to know is to measure the baseline of the specific process.
Do I need to change the machinery?
In most cases, no. Energy savings come from operating existing equipment better: setpoints, sequences, timing and anticipation. The relevant investment is usually in targeted instrumentation and data integration, not in replacing the furnace.
What if my historical data is patchy?
That’s normal in installations that are decades old. A historical record with gaps doesn’t stop you starting, but it shapes the scope of the first phase and requires a data-capture period before committing to energy-savings targets.
Keep reading: energy efficiency with AI · predictive models for industry · operational planning with AI
