Can Europe's Public Supercomputers Outrun the 1 GW Grid Queue?

If you ask a hyperscaler how to train a frontier AI model, they will point you to a familiar blueprint: buy tens of thousands of top-tier accelerators, pack them into a centralized campus, and hook them up to a gigawatt-scale power grid.

But if you ask an energy utility when that campus will actually get energized, the answer is sobering. The bottleneck for frontier-scale AI has shifted from silicon allocation to the physical realities of high-voltage transmission lines.

An open-source analysis and model called EuroMesh (with data and figures current as of June 2026) suggests an alternative path for Europe. Instead of waiting years for centralized gigawatt datacenters to clear grid queues, Europe could theoretically train a sovereign, frontier-class model by federating the public compute it already owns.

The 7.6-Year Power Bottleneck

The core argument for federated training is not that distributed clusters are computationally superior—they are demonstrably more difficult to manage—but that they already have power.

According to the EuroMesh model, a new 1 GW datacenter campus faces a mean wait time of 7.6 years to connect to the grid. This estimate is anchored by real-world constraints, including statements from AWS citing up to seven-year wait times for major grid connections, and data from the International Energy Agency (IEA) showing grid-queue lead times ranging from two to ten years across major economies. To date, no European operator has successfully energized a single 1 GW point load for AI compute.

In contrast, Europe already operates tens of exaflops of public AI compute. This hardware is distributed across the EuroHPC Joint Undertaking supercomputing flagships and 19 national AI Factories. These sites are already energized, connected, and paid for.

The Topology and the DiLoCo Penalty

Training a model across geographically isolated supercomputers requires throwing out the traditional synchronous training playbook. You cannot run standard All-Reduce operations over wide-area networks (WANs) without turning your training run into a very expensive heater that spends 99% of its time waiting on network packets.

To bypass this, the EuroMesh model relies on low-communication distributed training algorithms, specifically Distributed Low-Communication (DiLoCo) style training. DiLoCo allows local clusters to perform multiple steps of optimization independently before communicating weight updates across the WAN, drastically reducing the required network bandwidth.

To evaluate this trade-off, the EuroMesh model uses a three-layer simulation:

1. Layer 1 (Efficiency): Calculates the per-FLOP efficiency penalty of low-communication training (the "DiLoCo penalty").
2. Layer 2 (Time-to-Availability): Models when different public compute sites energize and how fast cumulative compute accrues.
3. Layer 3 (Regional Scorecard): Evaluates time, cost, carbon, and feasibility across different European regions.

Surprisingly, the sensitivity analysis reveals that the DiLoCo training efficiency penalty is a second-order concern. The race is won or lost almost entirely on Layer 2 (time-to-availability). Because the federated sites are already online, the model estimates that a federated approach could deliver a frontier-class model around 2028. Conversely, waiting for a centralized 1 GW campus to clear grid queues pushes the delivery date of a comparable model out to approximately 2033.

The Pragmatic Caveats

While the math of the EuroMesh model is compelling, translating it into an actual training run introduces massive engineering and political hurdles.

First, the compute is owned but not currently unified. The EuroHPC machines are shared, heterogeneous, and managed via traditional batch schedulers. Carving out a dedicated, synchronized slice of this infrastructure for a single continuous training run is a political challenge, not a hardware one.

Second, DiLoCo-style training remains largely unproven at true frontier scale. While it has shown promise in research, it has not been widely demonstrated on models larger than 10 billion parameters. Attempting to scale this topology to a 100B+ parameter model would be an active research project in its own right, rather than a guaranteed engineering run.

Ultimately, the EuroMesh model serves as a stark reminder of the physical limits of the AI buildout. If Europe wants a sovereign frontier model this decade, its best bet might not be building bigger datacenters, but writing smarter distributed training algorithms to harness the supercomputers it already has.