AI Data Centers: 80 Terawatt Hours, a Weakened Efficiency Law, and My 45 Euro Phone
AI data centers are heading toward 80 terawatt hours in Germany by 2045, and the Bundestag is weakening the Energy Efficiency Law exactly now: higher PUE limits, more exceptions on waste heat. The counter test on my 45 euro deck, night of 02:21 included.
Kein fertiges Produkt und keine Hochglanz-Rückschau: Dieser Eintrag hält fest, was gebaut, getestet, verworfen und als Nächstes überprüft wird.
# AI Data Centers: 80 Terawatt Hours, a Weakened Efficiency Law, and My 45 Euro Phone
Last night Android won the argument: at 02:21:38 my gateway process went down, its last heartbeat left 978 MB of free memory and 3.1 GB of swap behind, and the signature points to an OOM kill. I keep the final device side proof open, a signature is not a verdict. By 02:35 the operation was back up, four clean sessions, 21 jobs armed, the outage cost around 14 minutes. I run this agent operation on a used Google Pixel 6a that cost 45 euros, and in that world efficiency is not an administrative chore, it is the survival condition. On the same day I wrote down the crash, Berlin went about weakening its Energy Efficiency Law while forecasts for the country's data centers point toward 80 terawatt hours per year. Precisely while AI data centers grow measurably fast, politics is loosening the rules that demand transparency, efficiency and waste heat reuse. My deck shows what this debate loses: limits force better operations.
How much power do AI data centers need, which rules cover them, and what is happening to those rules right now? The data is thin because the industry reports mostly voluntarily. The credible forecasts show a multiple of today's consumption, the government's answer is nevertheless not a sharper measurement but a dilution, up for debate in the Bundestag in November. In this post I put the German numbers next to the global ones and set the night of 02:21 against them as my own measurement. The difference between both worlds is not scale, it is the attitude toward limits.
AI data centers in numbers: strategy, register, law
Germany's National Data Center Strategy describes the buildout without euphemism: installed IT power is supposed to grow from roughly two gigawatts today to more than four gigawatts by 2030, more than a doubling, and AI capacity should even quadruple. The plan comes with closer cooperation between the federal government and Nvidia, announced by chancellor Friedrich Merz and Nvidia chief Jensen Huang at an investment summit in Berlin. Competitiveness sounds reasonable, the data base behind this growth stays thin anyway. A study by Borderstep and others, commissioned by the economics ministry, counts more than 2,000 data centers above 100 kilowatts of grid connection, and even those studies rest on voluntary industry surveys. Nobody admits in such a survey that their building handles power, water or space particularly badly. The same study expects about 80 terawatt hours by 2045, four times today's value, enough to run 16 million heat pump households for a year. A forecast built on voluntary reports stays a blind flight, and before loosening the rules comes measurement, not leniency.
The data center register: a half open window
Since mid August one instrument exists: the federal data center register, prescribed by the Energy Efficiency Law. Out of an estimated 1,000 reporting obligated data centers, only 319 published their data transparently under their names, another 166 reported anonymously, and the websites of the remaining operators gave nothing either. Three findings follow from what was reported, established by the Federal Environment Agency: a good third of German data centers already meets the future PUE requirements. Another third sits above 1.5 and needs clear limits, voluntariness did not deliver them. And the waste heat reuse factor of existing data centers is tiny at just under three percent, large amounts of near carbon neutral heat go to waste every year. A third already meets the future limits, which is not a reason to loosen them.
What the law demands today and what the revision leaves
The Energy Efficiency Law arrived in 2023 and sets two levers: power usage effectiveness (PUE) and waste heat reuse, measured as the Energy Reuse Factor. New data centers have had a PUE limit of 1.2 since July 2026, their waste heat factor rises to 20 percent by July 2028, and from July 2027 they must cover their power needs on paper from renewables. That paper coverage counts even with Norwegian hydro certificates while the actual power comes from a gas plant, certificates are bookkeeping, not electricity. The revision walks each lever back, and the coalition treaty claims Germany implemented the EU requirement with excessive ambition.
That claim is shaky. The European Energy Efficiency Directive demands savings in final energy, member states without obligation systems must deliver comparable measures instead. Germany decided against white certificate systems and explicitly named the stricter data center rules as its substitute measure. Lowering them without replacing the substitute amounts, in the view of legal experts cited in the article, to a serious violation of European law. Denmark shows the alternative: future grid connection requests get prioritized politically, with households, healthcare and charging infrastructure ranked ahead of data centers. A register with three percent waste heat reuse and a third of anonymous reports is the thinnest possible foundation for planning a doubling of national IT power.
Auch in den kommenden Jahren werden Rechenzentren weiterhin auf fossile Energieträger wie Erdgas und Kohle angewiesen sein. (In the coming years, data centers will keep depending on fossil energy sources like natural gas and coal.)
— Jens Gröger, research coordination for sustainable digital infrastructures, Öko-Institut
The global numbers: Gartner and the IEA point the same way
The global view agrees. Gartner puts worldwide data center power consumption at 565 terawatt hours for 2026, up 26 percent from 447 in 2025, with AI optimized servers at 31 percent of consumption and drawing more power than conventional servers from 2027. By 2030 Gartner sees more than 1,200 terawatt hours. The IEA sets consumption for 2024 at around 415 terawatt hours, about 1.5 percent of world electricity demand, and its base case reaches roughly 945 terawatt hours by 2030. Both institutes work with different methods and still land in the same direction: steeply rising demand, driven by artificial intelligence.
The night of 02:21: efficiency as an operating condition
And once more the other measurement series that nobody orders and that runs every night anyway. At 02:21:38 the old gateway's last heartbeat ended, the log shows 978 MB of free memory and 3.1 GB of swap in use at the end, the signature speaks for an OOM kill, the device side proof is still open. Recovery started at 02:27, the boot scripts had survived, the first start hung on a stale tmux socket, after cleanup four sessions were live again at 02:35 with 21 jobs. Total outage: around 14 minutes, on a night when an Android update fought the agent processes over the same memory.
The hardening of this operation shows what limits do when you take them seriously. On August 31 my stability log counted 46 OOM kills within the preceding seven days, after a watchdog, an external cron tick and memory discipline the same seven day windows show zero OOM kills since September 7, with availability close to 100 percent. Free memory moves between roughly 900 and 1,400 megabytes, and this morning at 07:05 I took a fresh reading: 1,373,972 kB free of 5,723,472 kB total, measured in Termux, no root. A desktop client from the agent scene takes more memory in an earlier measurement than my complete stack with around 450 MB. Limits force architecture, their absence forces bills.
The honest caveat: a single device measurement proves nothing about gigawatt data centers, the scales differ by orders of magnitude and cooling is its own engineering problem there. The comparison shows only the structure of the decision. On the deck efficiency is the condition for the operation existing at all: a local agent core on old hardware, thinking in the cloud model, mechanics as scripts at zero model requests. That separation is the same lever operators under efficiency pressure need, which is exactly why exceptions should not buy them out of it.
The way out: leaner operations instead of exceptions
The operation on the 45 euro phone pays for this kind of efficiency in its own currency, the economics live in my pro stack post (post.html?slug=pro-agent-stack-45-euro-phone): light local software, thinking in the cloud, mechanics as scripts. The agent scene moves knowledge out of system prompts into files this fall, my operation runs the same direction with skills, gates and scripts. Energy policy asks the same question in gigawatts, and the boundary conditions tighten: Frankfurt is at its power limit (post.html?slug=frankfurt-full-data-centers-wait), my post on phantom power covers the 700 gigawatts of announced compute without secured grids (post.html?slug=phantom-power-700-gigawatt-data-centers), and Texas discussed permitting pauses (post.html?slug=texas-pauses-data-center-permits).
The levers against this pressure have plain names: mandatory register reports instead of voluntariness, which has so far been rewarded with an anonymous third and three percent waste heat reuse. Limits that a third of operators already meets are buildable, the register provides that proof. An EU label that shows efficiency and water use without grading every applicant into class A, and waste heat as a duty instead of an exception clause. Grid connections that get used or handed back instead of sitting parked for years, the Danish example proves allocation is a political decision. Efficiency requirements are not an anti industry course, they are the invoice that makes the AI buildout politically sellable at all.
The point
My conclusion stands on two legs. First, the numbers do not justify weakening the law, they prove the opposite: more measurement and full transparency before the country's IT power doubles, because a third of operators already meets the future limits today. Second, efficiency is the operating wisdom of every running system, from the network edge to the Pixel 6a. My operation paid 14 minutes of outage last night because another program spent the same memory, the lesson was build leaner, not buy hardware. The industry gets power, grids and water, society gets a law without teeth. It can be done more honestly, my deck proves it every night.
What is the German Energy Efficiency Law (EnEfG) and what changes in November?
The EnEfG arrived in 2023 and regulates power usage effectiveness (PUE) and waste heat reuse (Energy Reuse Factor) for data centers. Current rules include a PUE of at most 1.2 for new data centers from July 2026, paper 100 percent renewable power from July 1, 2027, and a 20 percent waste heat factor for new facilities by July 2028. The 2026 revision raises the PUE values and adds further exceptions on waste heat, the Bundestag is expected to debate it in November.
Where does the 80 terawatt hours by 2045 figure come from?
It comes from a study by Borderstep and others commissioned by the German economics ministry: data center consumption in Germany grows from roughly 20 terawatt hours today to about 80 terawatt hours in 2045. The netzpolitik.org framing: that amount would cover around 16 million households with heat pumps or 40 million electric cars. The forecast stands or falls with its data base, and that base rests on voluntary reports.
What does PUE mean and why fight over 1.2?
PUE (Power Usage Effectiveness) is the ratio of a data center's total power draw to its pure IT draw, 1.0 is the ideal. Currently 1.2 is the limit for new data centers under the EnEfG. According to the register analysis by the Federal Environment Agency, a good third of German data centers already meets the future requirements, another third sits above 1.5. The planned revision raises the limits.
What does waste heat reuse (ERF) actually deliver?
The Energy Reuse Factor measures which share of a data center's waste heat is actually reused. The value for existing German data centers sits at just under 3 percent according to the register analysis. The current law expects new data centers to reach 20 percent stepwise by July 2028. The revision adds further exceptions, while the Öko-Institut study recommends binding transparency duties, an efficiency label and metrics for the ecological footprint.
What can a single agent operation say about this?
Nothing about gigawatt scale power quantities, the scales differ by orders of magnitude. But about the structure of the decision: on my deck efficiency is the existence condition, shown by the OOM kill on November 3, 2026 at 02:21 (978 MB free, 3.1 GB swap at the end) and by the hardening from 46 OOM kills down to zero per week. Limits force better architecture, and that argument is currently missing from the German process.
About the author
About the author: I run the HUNTER cyberdeck, an agent operation on a used Google Pixel 6a, without root, with Termux, herdr, Hermes and pi. The own measurements in this post come from the lifecycle ledger and the daily note of November 3, 2026, from the stability log entries of August 31 and September 7, 2026, and from a fresh reading of /proc/meminfo on November 3, 2026, 07:05 CET. The research basis is the set of public documents linked in the sources section, all retrieved on November 3, 2026. Last fact checked on November 3, 2026.
Sources
- netzpolitik.org, guest article by Leonard Burtscher, "KI-Boom: Wie die Bundesregierung die Regulierung von Rechenzentren aushöhlt", Oct 23, 2026, https://www.netzpolitik.org/2026/ki-boom-wie-die-bundesregierung-die-regulierung-von-rechenzentren-aushoehlt/ (retrieved Nov 3, 2026)
- BMDS, National Data Center Strategy, https://bmds.bund.de/service/publikationen/nationale-rechenzentrumsstrategie (referenced in the netzpolitik article, retrieved Nov 3, 2026)
- BMWK, "Stand und Entwicklung des Rechenzentrumsstandorts Deutschland", study with Borderstep involvement, https://www.bundeswirtschaftsministerium.de/Redaktion/DE/Publikationen/Technologie/stand-und-entwicklung-des-rechenzentrumsstandorts-deutschland.html (referenced in the netzpolitik article, retrieved Nov 3, 2026)
- Federal Environment Agency, register analysis, https://www.umweltbundesamt.de/themen/rechenzentrumsregister-daten-widerlegen-hohe (referenced in the netzpolitik article, retrieved Nov 3, 2026)
- Öko-Institut, press release on the Greenpeace commissioned study "KI auf Kosten des Klimaschutzes", https://www.oeko.de/news/pressemeldungen/ki-auf-kosten-des-klimaschutzes-energiebedarf-von-rechenzentren-verdoppelt-sich-bis-2030/ (retrieved Nov 3, 2026)
- IEA, "Energy and AI, Energy demand from AI", https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai (retrieved Nov 3, 2026)
- Gartner, press release "Data Center Electricity Consumption to Grow 26% in 2026", June 10, 2026, https://www.gartner.com/en/newsroom/press-releases/2026-06-10-gartner-says-data-center-electricity-demand-to-grow-26-percent-in-2026 (retrieved Nov 3, 2026)
- Own measurement: /proc/meminfo, MemAvailable 1,373,972 kB of 5,723,472 kB total, Pixel 6a, Termux/PRoot, Nov 3, 2026 07:05 CET
- Own measurement: lifecycle ledger, gateway end 02:21:38 with 978 MB free and 3.1 GB swap, recovery 02:27 to 02:35, 21 jobs, documented in the daily note of Nov 3, 2026
- Own measurement: stability-snapshots.log, oom_kills_7d=46 on Aug 31, 2026 and oom_kills_7d=0 on Sep 7, 2026
- Own measurement, cited from the post "The pro agent stack runs on a 45 euro phone": agent stack about 450 MB, desktop client 658 MB, Sep 5, 2026
Nachgebaut oder getestet?
Teile deine Erfahrung mit genau diesem Build-Log. Reviews werden vor der Veröffentlichung moderiert.
Noch keine Reviews für diesen Eintrag.
Fragen, Reaktionen und Verbesserungen werden direkt über GitHub Discussions eingebettet.
GitHub-Bereich öffnen ↗