EuroMesh Model Shows Federated Compute Beats Waiting for Gigawatt Datacenters

A new gigawatt AI campus in Europe waits a mean of 7.6 years for grid power. The continent already operates tens of exaflops of public compute across EuroHPC supercomputers and the national AI Factories. Federating that existing hardware with low-communication training could deliver a frontier-class model by 2028, not 2033.

The EuroMesh model, released as a fully reproducible open-source analysis on GitHub, puts hard numbers on a question that usually gets hand-wavy policy answers. Its headline result comes down to one inequality: the federation wins if its sites are online before a gigawatt campus is.

The Grid Queue: 7.6 Years vs. 2028

Layer 2 of the three-layer model tracks time-to-availability per region. Grid-connection lead times are sourced from seven regions, anchored by AWS's "up to seven years" statement and the IEA's 2-to-10-year range. The central estimate of 7.6 years for a 1 GW point load means no European operator has yet energized one; the grid simply can't keep up.

Meanwhile, the EuroHPC flagships and 19 AI Factories already exist. They are batch-scheduled, heterogeneous, and shared, but the capacity is measurable. EuroMesh models that compute as a stopgap, not a permanent solution. The training efficiency penalty from low-communication (DiLoCo-style) training is second-order, confirmed by the sensitivity tornado.

Three Layers: Why Availability Trumps Efficiency

The model is built in three layers. Layer 1 calculates per-FLOP efficiency of distributed training with communication constraints. Layer 2, the crucial one, determines when each compute site energizes and how fast cumulative compute accrues. Layer 3 scores each region on time, cost, carbon, and feasibility.

The repo includes a full parameter set with confidence tags, a pytest suite of 52 tests, and a run.py that regenerates every CSV and figure from a clean tree. The model's results, scenarios, and caveats are in model/RESULTS.md. All data sources are cited per region.

Honest Caveats and Real Constraints

The author doesn't sugarcoat the limitations. Grid-queue lead times are sourced central estimates, not observed figures, because no European operator has yet energized a 1 GW point load. The addressable fraction of EuroHPC machines for a single coordinated run is a political decision, not a hardware fact. Frontier-scale distributed training above about 10B parameters is unproven, so the target is a credible frontier-class model rather than a guaranteed 405B.

Those caveats are spelled out in model/RESULTS.md and the report's caveats section. The analysis is dated June 2026 and is not peer-reviewed. But the model is auditable, reproducible, and its core thesis is blunt: waiting for gigawatt datacenters means waiting a decade, while federating what Europe already owns buys roughly five years.

Source: Show HN: Can Europe train a frontier AI model on the compute it owns?
Domain: github.com