Frankfurt hosts one of the world's largest internet exchanges and Europe's densest concentration of data centres, built there because of network interconnection rather than because of power or land. Artificial intelligence workloads have changed the requirement fundamentally: the constraint is now grid connection capacity, and the industry is being pushed toward locations that have power rather than locations that have networks.
Data centres followed fibre for twenty years and now follow electricity. That reversal is reshaping where digital infrastructure gets built in Germany, and it collides directly with the grid constraints that already limit industrial investment. This case study closes the media and telecom pillar of the Germany Company Stories hub.
Why Frankfurt?
A major internet exchange created dense interconnection, and proximity to it reduced latency and transit costs, which attracted more operators in a self-reinforcing cycle.
What changed?
Artificial intelligence training and inference require power densities several times higher than conventional hosting, making grid capacity the binding constraint.
Where is capacity moving?
Toward regions with available grid connection and generation, including northern and eastern Germany, rather than toward established network hubs.
Why did the industry concentrate in Frankfurt?
Because of interconnection. A large internet exchange allows networks to exchange traffic directly rather than paying transit providers, which reduces cost and latency, and the value of being present rises with the number of other networks present.
That is a network effect in the literal sense, and it produced the same winner-takes-most dynamic seen in marketplaces. Once a critical mass of networks connected in one location, every new network had a strong reason to connect there too.
Data centres followed the exchange because customers wanted short, cheap paths to many networks. Financial trading, content delivery and enterprise connectivity all valued proximity in milliseconds.
The secondary factors were legal and commercial: a central European location, a stable jurisdiction, strong data protection law and proximity to a major financial centre with substantial enterprise demand.
Why do artificial intelligence workloads change the requirement?
Power density. A conventional server rack draws several kilowatts; racks packed with accelerators for training or inference draw many times that, and cooling a high-density rack requires liquid systems rather than air.
The consequence is that a facility designed for conventional hosting cannot simply be filled with accelerators. Electrical distribution, cooling capacity and floor loading all become constraints, and retrofitting is frequently more expensive than building new.
The scale of new facilities has grown correspondingly. Campuses measured in hundreds of megawatts are now planned, which is comparable to the consumption of a small city and requires transmission-level grid connection rather than distribution-level supply.
That single change moves the siting decision from network planning into energy planning, and it puts data centre developers into the same connection queues as industrial projects, competing for the same scarce grid capacity described in the grid bottleneck analysis.
Where is new capacity actually being built?
Where power is available. Regions with substantial renewable generation and lower local demand can offer connection capacity that saturated metropolitan networks cannot, which favours northern and eastern Germany over the traditional Frankfurt corridor.
Several large projects have been announced on exactly this logic, including facilities in the low hundreds of megawatts sited for grid access and renewable proximity, such as the campus described in the Schwarz Group case study.
The trade-off is latency and interconnection. A facility distant from the exchange requires dedicated fibre back to the interconnection point, which adds cost and a few milliseconds, and for training workloads that is entirely acceptable because training is not latency-sensitive.
The emerging pattern is therefore a split architecture: latency-sensitive inference and enterprise workloads near the interconnection hubs, and training and bulk compute where power is abundant, connected by high-capacity fibre.
What about waste heat and cooling regulation?
Germany has legislated on both, requiring new facilities to meet efficiency standards and to make waste heat available for reuse where feasible, alongside requirements on renewable electricity procurement.
Waste heat reuse is technically straightforward and commercially awkward. A data centre produces large volumes of low-grade heat, and using it requires a district heating network nearby with a compatible temperature regime and a willingness to depend on a private facility for supply.
Where the conditions exist, the arrangement is genuinely valuable: the data centre reduces cooling cost and the heating network reduces fuel consumption. Where they do not, the requirement adds cost without benefit, which is why siting near existing district heating has become a planning consideration.
Cooling technology is changing independently. Liquid cooling raises the temperature at which heat is rejected, which makes reuse considerably more practical than with air-cooled facilities, so the regulatory requirement and the technical trend are converging.
What does this mean for German industrial policy?
That data centres and manufacturing now compete for the same scarce resource. Every hundred megawatts allocated to computing is a hundred megawatts unavailable to an electrolyser, a steel plant or a chemical facility in the same region.
That is an uncomfortable trade because the employment profiles differ enormously. A large data centre employs a few hundred people directly; an industrial facility consuming comparable power employs many times that number.
The counter-argument is that computing capacity is infrastructure for the entire economy and that hosting it domestically supports the sovereignty objectives described in the digital sovereignty analysis.
The resolution most likely lies in flexibility. Data centres can shift non-urgent workloads in time and location far more easily than an industrial process can, which makes them valuable grid participants if the price signals and contracts are designed to reward it.
What should an enterprise consider when choosing a facility?
Power contract structure alongside the usual factors. Electricity is the dominant operating cost, and whether the facility passes through market prices, holds long-term renewable agreements or has fixed-price arrangements determines cost stability over the contract life.
The second consideration is capacity headroom. A facility operating near its electrical capacity cannot accommodate growth, so a customer expecting to expand should verify the site's available connection headroom rather than its current utilisation.
The third is interconnection. For workloads that need low latency to many networks, an established hub facility remains superior, and for bulk compute a remote site with cheaper power is preferable. Most organisations need both and should not force a single answer.
The fourth is regulatory: efficiency and heat reuse obligations affect operating cost and are being tightened, so a facility meeting only current minimum standards may face retrofit costs that are passed through during the contract.
How does colocation differ from hyperscale?
In who owns the equipment and who bears the utilisation risk. A colocation provider builds and operates the facility, selling space, power and cooling to customers who install their own hardware. A hyperscale operator builds facilities for its own use.
The economics differ accordingly. Colocation earns a return on the building and the electrical infrastructure with contracted revenue per kilowatt, which resembles industrial property with a services layer. Hyperscale is a cost centre supporting a much larger business.
The convergence point is build-to-suit, where a developer constructs a facility to a hyperscale operator's specification under a long lease, which combines infrastructure capital with a single creditworthy tenant.
For enterprises the practical distinction is control and flexibility. Colocation permits any hardware and any network provider; cloud hosting removes the hardware question entirely and creates the dependency examined in the sovereignty analysis.
What does sovereignty mean at the facility level?
Ownership and operational control of the building, the staff and the encryption keys, rather than merely the physical location of the servers.
A facility located in Germany but operated by an entity subject to another jurisdiction does not resolve the legal question described in the sovereignty analysis, which is why European operators emphasise ownership structure in their marketing rather than geography alone.
For an enterprise assessing a provider, the practical questions are who holds administrative access, who holds the keys, which entity employs the operational staff and which law governs the operating company. Those four answers determine the actual sovereignty position.
How should organisations think about the AI compute question?
By separating training from inference. Training is episodic, latency-insensitive and enormously power-hungry, which makes it suitable for remote low-cost capacity or for rented cloud capacity purchased when needed.
Inference is continuous, latency-sensitive and closer to a production workload, which favours capacity near users or near the data being processed, and it is where sovereignty and data residency requirements usually apply.
For most enterprises the practical conclusion is that owning training infrastructure rarely makes sense, while controlling inference infrastructure sometimes does, particularly for regulated data. That split is the pragmatic version of the sovereignty debate, expressed as an architecture decision rather than a policy position.
What happens if the AI capacity build proves excessive?
The facilities remain useful, which distinguishes this from purely speculative infrastructure. A data centre built for accelerator workloads can host conventional computing at lower revenue per rack, so the downside is compressed returns rather than a stranded asset.
The more exposed party is whoever committed to specialised equipment and long-term power contracts at peak prices. Hardware depreciates on a short cycle and power contracts run for years, so an operator with excess capacity carries fixed obligations against falling revenue.
The historical parallel is the fibre build of the early 2000s, where enormous capacity was constructed, valuations collapsed, and the physical asset was subsequently used for two decades of internet growth. The infrastructure survived; the investors who funded it largely did not.
What should a German industrial region weigh?
Employment per megawatt against the strategic value of hosting compute capacity. A large data centre delivers substantial construction employment, modest permanent employment, significant local tax revenue and considerable grid consumption.
The strongest local case combines waste heat supply to a district network, a commitment to renewable procurement, and grid flexibility services that improve rather than degrade local network stability.
Regions that negotiate these terms at the planning stage capture real value; those that compete purely on speed of approval capture the consumption without the benefits.
Frequently Asked Questions
Why did data centres concentrate in Frankfurt?
Because a major internet exchange created dense interconnection between networks, and proximity reduced transit cost and latency, which attracted further operators in a reinforcing cycle.
What changed with artificial intelligence workloads?
Power density per rack rose several-fold, requiring liquid cooling and transmission-level grid connections, which made electricity capacity rather than network access the binding constraint.
Where is new capacity going?
Toward regions with available grid connection and renewable generation, particularly northern and eastern Germany, connected back to interconnection hubs by high-capacity fibre.
Are announced projects reliable indicators?
No. Many announced facilities lack secured grid connections, and queues in constrained regions extend for years, so pipeline figures overstate near-term delivery substantially.
Discover more from Kurums | Business Intelligence
Subscribe to get the latest posts sent to your email.
