Europe Has the Compute for Its Own AI — It's Just Spread Across 19 Buildings and 7 Power Grids

🔍 THE BOTTOM LINE

Europe already owns tens of exaflops of public AI compute — and almost none of it is being trained on. A reproducible math model published Sunday on GitHub shows that stitching those idle machines together with low-communication federated training delivers a 405-billion-parameter model in October 2028, fifty-two months sooner than a single 1 GW hyperscale campus can clear Europe’s grid-interconnection queue. The model is sourced, runnable, and has 52 passing tests. The thing standing between Europe and its own frontier AI is not chips, not money, and not talent. It’s the seven-year average wait to plug a gigawatt campus into a continental power grid.

What the Repo Actually Argues

The author calls it the “compute-at-home” thesis: stop trying to out-build OpenAI and Anthropic, and start federating what the EuroHPC Joint Undertaking already operates. The model assumes 40 to 50 medium European datacenters — each housing roughly 28,571 H100/H200-class GPUs — stitched together with DiLoCo-style low-communication training (the technique DeepMind published in 2024 for training across slow links without paying the full all-reduce penalty). Run the math: 28,571 GPUs per site, federated across a portfolio that energises in 2026 and 2027, hits Chinchilla-optimal training FLOPs for a 405B model in month 33 of the model horizon — October 2028. A single 1 GW campus applied for in January 2026 hits the same target in month 85 — February 2033. The headline gap is 52 months. The honest sensitivity band runs 44 months (optimistic: campus lands faster) to 66 months (pessimistic: campus hits the worst case of the EU grid-queue distribution).

Scenario	Mesh delivers 405B	1 GW campus delivers 405B	Gap (mesh ahead)
Optimistic	Dec 2027	Aug 2031	44 months (3.7 yr)
Baseline	Oct 2028	Feb 2033	52 months (4.3 yr)
Pessimistic	Feb 2030	Aug 2035	66 months (5.5 yr)

The band is monotone: the slower the campus, the larger the mesh win. The smaller targets land earlier still — 8B and 70B models come online in month 31 (August 2028) at baseline — but the 405B result is the one that matters editorially because it’s the scale that puts Europe in the same room as Anthropic, OpenAI, and Google DeepMind.

The Grid Bottleneck

The reason the gap is 52 months and not, say, 6, is the European grid-interconnection queue. A 1 GW campus is a point load so large that no existing substation can absorb it. The campus triggers new extra-high-voltage (400/420 kV) transmission build, plus a queue position at the relevant transmission system operator (TSO). The repo’s grid-queue dataset covers seven European regions and produces a single number: the mean wait for a 1 GW point load in 2026 is 7.6 years, anchored by Amazon Web Services’ EMEA energy chief telling an industry audience in February 2026 that securing power in Europe “can take up to seven years,” and the International Energy Agency’s 2-to-10-year range across EU member states, with Frankfurt, London, Amsterdam, Paris, and Dublin averaging 7 to 10 years. The Netherlands is the worst case at 10 years; Germany comes in at 7.5. A 40-to-50 MW medium site, by contrast, usually fits inside an existing substation envelope and waits 1 to 3 years. The gap between hyperscale and medium is, on average, 4 to 5 years — and that gap is exactly what the federation exploits.

What Europe Already Has Sitting Idle

The substrate inventory is the second half of the argument. Europe operates four EuroHPC flagship supercomputers — JUPITER at Jülich (ranked #4 on the June 2026 TOP500), Mare Nostrum 5 in Barcelona, LUMI in Kajaani, Leonardo in Bologna — plus 19 AI Factories selected across three rollout rounds (December 2024, March 2025, October 2025), most of them deploying through 2026 and 2027. Per the EuroHPC substrate document, those machines already aggregate to tens of exaflops at peak. The catch — and the repo is honest about this — is that vendor “AI exaflops” figures use FP4-with-sparsity, which is roughly eight times the dense BF16 number that actually trains models. Sum the headline figures and you wildly overstate federate-able capacity. Per the substrate doc, the realistic dense-BF16 aggregate is materially smaller, and a meaningful fraction of the fleet is consumed by existing science workloads, leaving only a sliver of dedicated AI-training time available to a federation in any given month.

The repo’s model uses a 28,571-GPU figure per medium site, which is roughly 16,000 H200s at the per-GPU BF16 peak — well within the count available on a single large EuroHPC machine or a cluster of mid-tier ones, with headroom to federate across multiple sites. The carbon intensity works out to a power-weighted 188.8 gCO₂/kWh across the seven sampled regions, ranging from Norway at 31 and Sweden at 38 (hydro and nuclear) to Germany at 311, the Netherlands at 294, and Poland at 514 (still coal-heavy) — per lowcarbonpower.org’s country-level data compiled from Ember, IEA, and the Energy Institute. France at 44 is the most attractive single jurisdiction, but no single site clears the model on its own.

The Other Side

Three honest criticisms, each of which the repo flags and the field will need to resolve before this is anything more than a thought experiment.

One: cross-border coordination. The EuroHPC flagships are funded and scheduled at the member-state level, with the EuroHPC JU acting as a coordinating layer above them. A federation that needs 40 sites to honour a single training run for months is asking a consortium of national procurement offices, science ministries, and vendor contracts to do something none of them were designed to do. The political question — “can Europe organise itself?” — is genuinely harder than the math question. The June 4 EU sovereignty package ([Europe Just Declared War on US Tech Dominance — €320 billion, chip emergency powers, cloud sovereignty tiers) was the policy side of this. The mesh is the engineering side. Neither works without the other.

Two: EuroHPC machines are shared, batch-scheduled science rigs. Federating them for a months-long frontier training run means displacing atmospheric-modelling, materials-science, and physics jobs that have their own published allocation calendars and user communities. The repo acknowledges the fraction of federate-able capacity in any given month is a fraction, not a majority, of nominal peak. The model handles this with a time-averaging assumption that the field will need to validate.

Three: the BF16 vs FP4 precision problem. When NVIDIA says an H200 delivers “1979 TFLOPs of AI performance,” that’s FP4-sparse — a number optimised for inference, not training. The dense BF16 peak on the same chip is roughly 989 TFLOPs. Summing the headline “AI exaflop” figures from EuroHPC press releases and the substrate’s own denominator don’t talk to each other. The repo flags this explicitly, treats it as a calibration issue rather than a fatal flaw, and asks the model reader to read every FLOP figure with the unit suffix attached. That kind of transparency is rare in the AI-compute discourse and worth naming.

NZ Angle

New Zealand has no equivalent to EuroHPC and no federation to join. The entire New Zealand sovereign-compute conversation — the [XeroForce and Kererū moment | the government’s AI strategy work, the Māori data-sovereignty effort, the call for a domestic frontier option — is, structurally, a 100% foreign-stack proposition. There is no NZ analogue to the Jülich or Barcelona machines, no shared EU procurement layer, no pooled grid-connection queue to fast-track. Whatever sovereign AI New Zealand builds, it builds on hardware and software sourced from the same three or four countries that the EUROMESH model is trying to make less dependent on. That’s not an argument against doing it — it’s an argument for being honest about the dependency shape.

❓ FAQ

Q: How confident is the 52-month headline? Is it a guess? A: It’s the baseline of a three-scenario sensitivity band. The baseline assumes a typical EuroHPC energisation calendar; the optimistic case assumes sites land 12 months earlier than scheduled; the pessimistic case pushes a 1 GW campus into a worst-case EU grid-queue (Netherlands-style, 10 years). The full math, all three scenarios, and 52 pytest tests are in the repo’s RESULTS.md. The model itself runs to zero from a clean tree.

Q: Does this actually use existing GPUs, or is it a proposal to buy new ones? A: Existing. The 28,571-GPU figure per medium site is a per-site assumption about the size of a site that fits inside an existing substation envelope. The federation is of machines already in the EuroHPC fleet and the 19 AI Factories, plus existing commercial colocation. The repo explicitly excludes the 1 GW campus path because that requires building new transmission.

Q: What is DiLoCo, and can it actually scale to 405B? A: DiLoCo is a low-communication distributed training method published by DeepMind in 2024. The basic idea: do most of the optimisation locally, synchronise less often, accept a small efficiency penalty in exchange for tolerating slow cross-site links. It was originally demonstrated on smaller models (hundreds of millions of parameters); scaling it to 405B is one of the open engineering questions. The repo treats the efficiency penalty as a layer in the model and runs a sensitivity tornado — the result is dominated by Layer 2 (time-to-availability), not Layer 1 (per-FLOP efficiency).

Q: What about the carbon? Is federated training greener? A: Roughly neutral, with high regional variance. The power-weighted mean is 188.8 gCO₂/kWh across the seven sampled regions, with France (44) and the Nordics (53 composite) the cleanest options and Poland (514) the dirtiest. A single new 1 GW campus in a clean-grid jurisdiction (France, Nordics) would beat the mesh on carbon; in a dirty-grid jurisdiction, the mesh can be designed to route around it. The repo treats carbon as a per-region scorecard input, not a hard constraint.

Q: Is this an official EU policy proposal? A: No. It’s a third-party GitHub repo posted to Hacker News on Sunday 14 June 2026 by a developer (handle smashini), framed as a sourced model and a short report. It has 10 stars and 73 HN comments at time of writing. The relevant official EU programme is InvestAI — €20 billion for up to five AI gigafactories inside a €200 billion target — which takes the centralised build-out path the repo argues against.

🔍 THE BOTTOM LINE

The most interesting thing about EUROMESH isn’t the conclusion. It’s the data: a sourced, runnable, three-layer model with a 52-month baseline, a 44-to-66-month honest band, 52 tests, and a per-region grid-queue dataset that nobody else has published. Whether Europe takes the mesh path, the gigafactory path, or neither, the dataset itself is now public and reusable. The Anthropic Fable 5 and Mythos 5 ban of 12 June 2026 — the sharpest reminder yet that frontier AI is foreign-controlled and access can be revoked on a Friday night — has made the sovereignty question unavoidable. The math is no longer the bottleneck. The politics is. And the day-2 escalation in Washington (the Jassy trigger, the directive’s broadening) shows the pressure is going up, not down.