Germany Plans to Quadruple AI Computing Power by 2030
Germany's Digital Minister Karsten Wildberger told Bloomberg TV that the reality is that computing power is already insufficient, and demand will continue to rise. He also presented a set strategy: to quadruple computing power and double data center capacity by 2030. The location is Berlin, and the discussion follows the national data center strategy approved by the cabinet in March, not new legislation.
The cabinet document states that based on 2025 as a baseline: nationwide data center IT access power will at least double, high-performance computing and AI access power will at least quadruple, with the benchmark being grid load rather than cabinet numbers. The Digital Ministry mentioned 28 supporting measures to make Germany a European sovereign data hub. Economic Minister Katherina Reiche co-sponsored the proposal. By 2025, data center electricity consumption is expected to be about 21 terawatt-hours, accounting for about four percent of national electricity consumption.
Industry organization Bitkom estimates that by 2025, Germany's AI-related capacity will be about 530 megawatts, accounting for about 15 percent of total installed capacity; by 2030, this figure is expected to reach 2,020 megawatts, accounting for about 40 percent of the total. Nationwide data center installed capacity is projected to be about 2,980 megawatts in 2025 and exceed 5,000 megawatts by 2030. Frankfurt alone has already surpassed 1,100 megawatts. Consulting materials indicate that Germany accounts for less than two percent of global AI computing power. The Jupiter supercomputer in Jülich has about 24,000 H100 equivalents, while the largest system in the U.S. can reach about 275,000.
Implementation measures include converting retired lignite power plants into data centers to utilize existing power grids, changing business taxes to be paid to the municipality where the facility is located to improve acceptance, prioritizing the reuse of existing sites in planning, aiming for 100 percent renewable electricity, and promoting tax-free input of waste heat into municipal heating at the EU level. Data centers are currently not on the EU's list of industries eligible for subsidized electricity prices. Since April, grid operators have been queuing large users based on "completion maturity"; those who can build first will connect first.
Buyers are local manufacturers needing inference and industrial models, telecom and retail groups; sellers are operators who can secure land, electricity, and permits, as well as chip manufacturers like NVIDIA. Funding flows from cloud subscriptions to transformers, grid queues, and the renovation of old power plants. Deutsche Telekom and NVIDIA have launched an industrial AI cloud in Munich's Tucherpark, claiming nearly 10,000 Blackwell chips, with a peak performance of 50 exaflops. The Schwarz Group does not plan to wait for EU subsidies and is building a site south of Berlin independently. The EU's €20 billion and five gigawatt-level factory tenders have been repeatedly delayed, with the bulk of subsidies pushed to 2028 and 2030. Beneficiaries are landowners and equipment vendors who can secure electricity; pressures include electricity prices, public acceptance, and the gap between U.S. cloud vendors, which still account for about one-third of new builds, and "sovereign" goals. The event is driven by demand pressure, and planning itself cannot deploy servers.
Source: Public Information
ABAB AI Insight
Germany framing the computing power shortage as an industrial issue acknowledges that the last round of digitization only purchased applications, not generators. Frankfurt has long relied on low-latency financial servers, while AI requires continuous high power, not just a series of network points. The leap from 530 megawatts to 2,020 megawatts in four years is written in access power, equating to treating transformers and old coal plants as industrial policy tools. Wildberger's statement "not enough, hence demand will rise" treats the shortage itself as an investment rationale: first acknowledging the gap, then writing it into the 2030 grid contract.
The capital path is electricity—land—tax, not model laboratories. Tax reform to the town where the data center is located buys the consent of local councils; renovating old power plants buys ready high-voltage entry; tax-free waste heat buys political cover for municipal heating networks. Chips still need to pay NVIDIA, and the industrial cloud in Munich turns "European sovereignty" into leasing U.S. accelerator cards. The EU's gigawatt factories pushing subsidies to after the construction peak effectively gives the time window to retailers and telecoms who pay out of pocket. Those who can prove "can build first" in the maturity queue will be able to take on the load first.
Comparisons include U.S. utilities planning trillion-dollar grid spending by 2030, UAE and South Korea with millions of H100 equivalents in parks, and France securing large-scale investment independently. Germany remains the leader in Europe for the number of data centers, but ranks mid-tier by GDP share. The industry phase is control: shifting from "welcoming all clouds" to "prioritizing new EU capacity," while still about one-third of new builds are U.S.-based. The sovereign narrative and capital structure have yet to align.
Structural changes indicate a reconfiguration of the supply chain. Pricing power shifts from software licenses to grid queues and municipal taxes. The mechanism is: model training can be completed abroad, while inference and industrial data are increasingly required to stay within jurisdiction; whoever controls transformers and permit timelines grants entry visas for models. Quadrupling capacity without increasing controllable power will shift the gap from chips to electricity bills. Planning can change the map's color but cannot alter the fact that U.S. hyperscale clouds are already established in Frankfurt.
ABAB News · Cognitive Laws
Writing the computing power shortage into planning is necessary for transformers to start construction.
Sovereign clouds buy grid queues, not slogans.
Old power plants can become industrial policy faster than new models.