Inference is processed energy. Electricity goes into a data centre and comes out as tokens, and agents, assistants and automated workflows are all priced in tokens. The electricity grid moves raw energy from where it is generated to where it is used. A compute grid would move the processed kind, and Europe should build one.
Europe is paying for the plants, AI Factories now and up to seven AI Gigafactories next, and wants them on clean power. It has no plan to connect them. Electricity became a continental economy because of the grid, which gave it a common unit and a cross-border market that lets a factory in Lyon run on wind from the North Sea. Compute needs the same, and the public role should stop there.
The American clouds already run one
The large American clouds have run compute grids inside their own walls for years. Google's cluster manager, Borg, does not count chips. It counts Google Compute Units, defined so that one unit "should provide about the same amount of computational power on any machine in the fleet", and it prices work by tier: production jobs pay full price, batch jobs pay less, and free-tier jobs pay nothing and get evicted when production needs the room. Microsoft's Singularity scheduler can pause any AI job and resume it in another data centre or region. Amazon sells its spare capacity at a discount as Spot instances, which an electricity trader would recognise as a balancing market.
No one in Europe can build that one company at a time. Alphabet, Amazon, Meta and Microsoft have guided to more than $700 billion of capital spending for 2026 alone. The public money in the Gigafactories programme is up to €10 billion. No European firm or member state closes that gap by itself, so the spending belongs at European level, and only for compute time. Public money should not go into models, applications or an attempt to build a European champion. I have argued before that Europe should stop trying to win the frontier-training race and put its effort into shared compute and the layer above the model. Electricity already splits the work this way: private companies own the power stations and sell to customers, and the public side sets the unit, runs or regulates the network, and couples national markets so power crosses borders under one set of rules.
A common unit
A kilowatt-hour is the same everywhere. A GPU-hour is not. Chips differ by generation, memory and interconnect, and a job has to travel with its data. The market has not solved this: the compute futures CME starts listing on 5 October have one contract per chip model, one for the H100 and another for the B200. CME files them under energy, next to power.
Europe already runs a compute federation with a common unit. The Worldwide LHC Computing Grid, which processes CERN's data across hundreds of national sites, pledges and accounts for capacity in one benchmark-normalised unit, HEPScore23, chained to a reference machine so that old pledges stay comparable as the hardware changes. AI compute needs that plus grades. Start with an accelerator-hour weighted by benchmark against a reference chip, then trade what it cannot absorb (memory, interconnect class, jurisdiction, whether the work can be interrupted) as grades at a price difference, the way oil is traded. It will be messier than the kilowatt-hour, and it only has to be good enough that a buyer in Brussels can purchase an hour of capacity without caring which member state it runs in.
The Gigafactories are heading the right way
The AI Gigafactories call closes on 12 November. Private consortia will build and run the sites, and the EU and member states will act as anchor customer, buying guaranteed shares of compute access time. Buying compute time rather than buildings is the right model.
But each contract is bilateral. Nothing in the published terms defines a common unit, a way to pool the public's shares across sites, or a link to the existing AI Factories, each of which hands out GPU-hours on its own machine through its own calls. The EuroHPC platform meant to give them one front door shipped in April with the AI Factories "planned for future integration". Selection is due in early 2027. Unless the contracts change, Europe will get seven more separate sites.
Build where the power is
The first reason to connect the sites is location. Most of the cost of AI compute is the chips, not the electricity; what holds sites back is getting a grid connection at all. The traditional data-centre hubs face connection queues of seven to ten years. Meanwhile clean power is thrown away where the users are not. Germany curtailed about 9.6 TWh of wind and solar in 2025, and Great Britain about 10 TWh, most of it Scottish wind that could not be moved south.
Bringing that power to Frankfurt means new transmission lines: SuedLink, Germany's main north–south line, is due in 2028 after roughly a decade of planning and permitting. Bringing the tokens means fibre, and a round trip across the continent takes tens of milliseconds, nothing next to the time a model takes to answer. With a grid, an operator can build next to curtailed wind and sell to a user a thousand kilometres away. Without one, operators wait in the queue or skip it: German operators are planning their own gas plants.
Training runs, fine-tuning and batch jobs can also wait or move, so a scheduler that can pause them turns data centres into demand that network operators can call on. Google has a gigawatt of this written into its utility contracts. A European grid could offer it from every connected site, which gives network operators a reason to connect those sites sooner.
Who runs it
The second reason is sovereignty. The compute grids that work today belong to American companies, and each decides its own prices and whose job gets evicted when it needs the room. A European customer rents from them and has no say in either. A European grid would settle those questions in public rules: who may sell into it, where data may go, and whose law applies. Only the public side can write those rules.
That starts with the Gigafactory contracts, which are being drafted now. Denominate the public's Gigafactory shares in a common, benchmark-normalised unit that can be scheduled across sites; that alone creates the grid without buying a single extra chip. Pool those shares with the AI Factories' public capacity, and allocate or sell them through one mechanism instead of one proposal queue per site. Give dispatch to an operator that owns none of the machines, as network operators own no power stations. Sell interruptible capacity for less and offer it to network operators as flexibility. And let any operator that meets the sovereignty rules sell in, whether it is a private European cloud, a national centre or a Gigafactory.
Jerome Leclanche
CEO, Ingram Technologies