Exascale and the Electricity Bill: What the JUPITER Supercomputer at Jülich Reveals About Powering AI and Climate Science

Key Takeaway

>

AI energy consumption is now a hardware-design constraint rather than an afterthought, and Europe’s first exascale machine — JUPITER at Forschungszentrum Jülich — shows what that constraint looks like in megawatts. The system crossed the exascale threshold because its designers accepted a hard power envelope and engineered around it, which means the lessons travel well beyond one German research campus.

>

• JUPITER was inaugurated at Forschungszentrum Jülich in September 2025 as Europe’s first exascale supercomputer, funded through roughly €500 million split between the EuroHPC Joint Undertaking and Germany.
• On the June 2025 TOP500 list, the JUPITER Booster ranked fourth worldwide at approximately 793 petaflops, drawing around 13 megawatts during that benchmark run.
• That translates to roughly 60 gigaflops per watt on the Green500 list, a figure achieved because the machine leans on NVIDIA GH200 accelerators rather than general-purpose CPUs.
• The International Energy Agency projects global data centre electricity demand will roughly double to about 945 terawatt-hours by 2030, driven largely by AI workloads.
• Direct hot-water cooling and waste-heat reuse reduce the overhead electricity that would otherwise go to chillers, which is where a large share of data centre inefficiency has historically sat.

Europe spent years being a consumer of exascale computing rather than a host of it. That changed in September 2025, when Forschungszentrum Jülich formally inaugurated JUPITER — the continent’s first supercomputer capable of more than a quintillion double-precision calculations per second. The machine matters for climate modelling, materials discovery and European-built language models. It matters just as much as an energy case study, because an exascale system is effectively a small industrial load bolted onto a regional grid.

The uncomfortable arithmetic is this: compute demand has grown faster than the efficiency of the silicon running it. Therefore AI energy consumption has become a procurement question, a siting question and a public-finance question all at once. When Jülich’s engineers specified JUPITER, they were not asking how much performance money could buy. They were asking how much performance a fixed power budget could buy — a different question with very different engineering answers.

This article uses JUPITER’s published figures as a reference point, sets them against the International Energy Agency’s demand projections, and explains why the energy profile of scientific computing has become inseparable from the economics of climate technology investment.

About this analysis

>

Written by the billgates.pk editorial team, which tracks climate technology and global health funding with a focus on stated amounts, timelines and measurable outcomes. Our analytical approach separates institutional positions from individual commentary and follows the money rather than the announcement.

Transparency note

>

This article is independent analysis. It is not affiliated with, endorsed by or produced in cooperation with Forschungszentrum Jülich, the EuroHPC Joint Undertaking, NVIDIA, the Bill & Melinda Gates Foundation or Breakthrough Energy. All figures are attributed to their published sources and linked in the References section. Where a figure has not been published, we state that rather than estimate it.

Why JUPITER Changed the Arithmetic of European Supercomputing

JUPITER crossed the exascale line because European institutions pooled capital that no single member state would have committed alone. The system was procured under the EuroHPC Joint Undertaking, with roughly €500 million covering acquisition and operation, split between EU-level funding and German federal and North Rhine-Westphalian contributions. This co-financing structure matters analytically: because operating costs were budgeted alongside hardware from the start, electricity was treated as a design input rather than a surprise on an invoice.

The performance figures confirm the strategy worked. On the June 2025 TOP500 ranking, the JUPITER Booster placed fourth globally at approximately 793 petaflops on the LINPACK benchmark — the highest-placed European system. Consequently, Europe gained a machine capable of running kilometre-scale climate simulations and training large foundation models without renting capacity from US hyperscalers, which had been the default for European researchers.

JUPITER at a glance	Figure
Location	Forschungszentrum Jülich, Germany
Status	Europe’s first exascale system, inaugurated September 2025
TOP500 result (June 2025)	~793 petaflops, ranked 4th worldwide
Power draw during benchmark run	~13 megawatts
Energy efficiency (Green500, June 2025)	~60 gigaflops per watt
Primary accelerators	NVIDIA GH200 Grace Hopper superchips
Total investment	~€500 million (EuroHPC JU and Germany)
Cooling method	Direct hot-water liquid cooling with waste-heat reuse
Published annual electricity cost	Not stated

How Much Electricity an Exascale Machine Actually Draws

Roughly 13 megawatts during a full-load benchmark run — that is the order of magnitude, according to the Green500 energy-efficiency list published alongside the June 2025 TOP500 results. For context, that is comparable to the continuous draw of a mid-sized industrial plant, and it is drawn by a single room of equipment.

The figure becomes more interesting when divided rather than quoted raw. At approximately 60 gigaflops per watt, the JUPITER Booster sits among the more efficient large systems ever benchmarked. This efficiency is driven by architecture, not by restraint: GPU-class accelerators perform far more floating-point work per watt than general-purpose processors on the dense linear algebra that dominates both scientific simulation and neural network training. As a result, the same power budget delivers several times the science it would have delivered on a CPU-only design from the previous hardware generation.

There is a catch that efficiency advocates tend to underplay. Improvements in performance-per-watt have historically been reinvested in larger workloads rather than banked as savings — a rebound dynamic economists have documented across energy-using technologies for decades. Therefore better chips reduce the energy cost of each calculation while total facility demand continues climbing.

Why AI Energy Consumption Outpaced Efficiency Gains

Demand growth has simply been faster than the efficiency curve. The IEA’s Energy and AI report, published in 2025, projects global data centre electricity consumption roughly doubling to around 945 terawatt-hours by 2030 — close to 3% of global electricity demand — with AI-specific workloads the dominant driver of the increase. An earlier IEA analysis, Electricity 2024, had already flagged a range of 620 to 1,050 terawatt-hours by 2026, up from roughly 460 terawatt-hours in 2022.

The national picture reinforces the trend. A 2024 US Department of Energy report prepared by Lawrence Berkeley National Laboratory estimated that American data centres consumed about 176 terawatt-hours in 2023, equivalent to roughly 4.4% of US electricity, and projected a rise to between 6.7% and 12% by 2028. That upper bound implies grid interconnection queues, transmission build-out and generation procurement decisions that take longer than a typical AI product cycle — which is precisely why utilities and regulators became reluctant participants in the AI buildout.

Public scientific computing sits inside this same squeeze, with one important difference. Because facilities such as Jülich are publicly funded and publicly audited, their power figures are published and benchmarked. Commercial AI training runs are rarely disclosed at comparable granularity. Consequently, the best available public data on AI energy consumption at the frontier of compute comes disproportionately from research supercomputers rather than from the companies training the largest commercial models.

Efficiency Engineering: How Jülich Holds AI Energy Consumption Down

The decisive interventions at JUPITER are thermal, not computational.

Hot-Water Cooling and Waste-Heat Reuse

JUPITER uses direct liquid cooling with warm water circulated to the components rather than chilled air pushed through a room. The engineering logic is straightforward: because the coolant leaves the system at a usefully high temperature, the facility avoids much of the compressor-driven refrigeration that has traditionally consumed a substantial fraction of data centre electricity on top of the computing load itself. Forschungszentrum Jülich has also designed the system so that waste heat can be fed into campus heating, which converts a disposal problem into a displaced fuel demand.

The effect of this choice on reported AI energy consumption is significant but easy to misread. Cooling improvements reduce overhead rather than the compute load, which means they lower total facility draw without making any individual training run cheaper in energy terms.

Accelerators and the Performance-per-Watt Trade

JUPITER’s Booster module is built around NVIDIA GH200 Grace Hopper superchips, tightly coupling CPU and GPU memory. This design reduces data movement between processor and accelerator, and because moving data costs more energy per bit than computing on it, cutting that traffic lowers consumption directly. The trade-off is specialisation: code must be rewritten to exploit accelerators, therefore research groups without engineering support capture a smaller share of the available efficiency.

What JUPITER Computes for Climate Science

Climate modelling was one of the explicit justifications for European exascale investment, and the reasoning is physical rather than political. Current global climate models operate at resolutions too coarse to represent convection, clouds and local topography explicitly, so these processes are approximated — the single largest source of uncertainty in regional projections. Exascale capacity enables kilometre-scale simulation, which resolves some of those processes directly instead of parameterising them. The European Commission’s Destination Earth initiative was built on exactly this premise.

The practical consequence is better-grounded adaptation planning. Flood exposure, heat stress and agricultural yield projections become more useful to governments and insurers when modelled at the scale decisions are actually made. There is an obvious irony in burning megawatts to model a warming climate, but the comparison that matters is between the electricity used and the mis-investment avoided. A single poorly sited piece of infrastructure can waste more capital than a supercomputer consumes in energy over its service life.

What This Means for Climate Tech Funders and Grant Applicants

Compute has become a line item in climate technology diligence. Investors who once assessed hardware, feedstock and offtake agreements now assess training budgets and inference costs, because a model-dependent business with unhedged electricity exposure carries a risk profile closer to heavy industry than to software. This shift explains why grid, storage and firm clean power have attracted capital alongside the AI applications that depend on them — a pattern visible across climate tech and AI health investing portfolios that pair compute-intensive tools with the generation needed to run them.

For researchers writing proposals, the practical implication is narrower and more useful: state the compute requirement explicitly, name the facility or allocation being sought, and quantify the energy footprint rather than leaving reviewers to guess. Publicly funded exascale time is allocated competitively, and proposals that treat power as a real constraint tend to read as more credible than those that do not.

Frequently Asked Questions

How much electricity does the JUPITER supercomputer use?

Approximately 13 megawatts during full-load benchmarking, according to Green500 figures published in June 2025. That draw delivered around 793 petaflops, giving roughly 60 gigaflops per watt. Actual operational consumption varies with workload mix, and Forschungszentrum Jülich has not published a single annual electricity figure for the system.

Is AI energy consumption actually a problem for electricity grids?

It is a timing problem more than a volume problem. The IEA projects data centre demand reaching about 945 terawatt-hours by 2030, near 3% of global electricity. The difficulty is that AI facilities are built in months while transmission and generation take years, so local grid constraints bite before national supply does.

Why does an exascale system need liquid cooling?

Because air cannot remove heat fast enough at exascale power densities. Direct hot-water cooling carries heat away more efficiently and leaves the coolant warm enough for reuse in campus heating. This reduces the refrigeration electricity that sits on top of the computing load, lowering total facility demand without changing the compute itself.

Who paid for JUPITER?

The EuroHPC Joint Undertaking and Germany, roughly €500 million in total. EU-level funding was matched by German federal and North Rhine-Westphalian contributions covering acquisition and operation. This co-financing model exists because exascale capital and running costs exceed what most individual member states commit to a single research facility.

Limitations and What This Analysis Does Not Establish

Research limitations. Benchmark power figures measure a specific, highly optimised workload under controlled conditions. LINPACK is not representative of climate simulation or transformer training, therefore the ~13 megawatt figure should be read as a reference point rather than an operating average. Published annual energy totals for JUPITER were not available at the time of writing, and IEA projections to 2030 are scenario-based rather than forecasts — they shift with assumptions about chip efficiency, model size and deployment rates.

Alternative approaches. Efficiency is not the only lever. Workload scheduling tied to grid carbon intensity, siting near existing clean generation, smaller task-specific models in place of general-purpose ones, and renting commercial capacity instead of building sovereign facilities all reduce emissions or cost through different mechanisms. Each involves trade-offs in data sovereignty, utilisation and capital commitment that this article does not resolve.

Professional consultation. Anyone making funding, siting or procurement decisions on the basis of these trends should consult grid engineers, energy economists and the relevant funding authority directly. Allocation rules, eligibility criteria and tariff structures differ by jurisdiction and change without wide publicity.

Conclusion

JUPITER’s significance is not that it is fast. It is that its energy profile is published, benchmarked and therefore arguable — which is more than can be said for most of the compute being built today. The machine demonstrates that exascale performance within a roughly 13-megawatt envelope is achievable when accelerators, hot-water cooling and waste-heat reuse are designed together rather than bolted on sequentially.

The broader implication is that AI energy consumption will be governed by infrastructure timelines, not by software ambition. Transmission lines, interconnection queues and firm clean generation move on decade-long schedules, while model sizes move on quarterly ones. Consequently the organisations that treat power as a first-order design constraint — as Jülich did — will build the systems that actually run, and the ones that treat it as an operating detail will meet the constraint anyway, just later and more expensively.

References

• International Energy Agency, Energy and AI (2025 report) — https://www.iea.org/reports/energy-and-ai
• International Energy Agency, Electricity 2024 (analysis and forecast to 2026) — https://www.iea.org/reports/electricity-2024
• Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report (prepared for the US Department of Energy) — https://eta.lbl.gov/publications/2024-united-states-data-center-energy
• TOP500, June 2025 list — https://www.top500.org/lists/top500/2025/06/
• Green500, June 2025 energy-efficiency list — https://www.top500.org/lists/green500/2025/06/
• Forschungszentrum Jülich, Jülich Supercomputing Centre — https://www.fz-juelich.de/en/ias/jsc
• EuroHPC Joint Undertaking — https://eurohpc-ju.europa.eu
• Destination Earth, European Commission — https://destination-earth.eu