{"slug": "show-hn-can-europe-train-a-frontier-ai-model-on-the-compute-it-owns", "title": "Show HN: Can Europe train a frontier AI model on the compute it owns?", "summary": "A new report and model from a developer argues that Europe can train a frontier-class AI model by federating its existing public compute, including EuroHPC supercomputers and AI Factories, achieving a model by 2028 versus 2033 for a new gigawatt datacenter. The analysis relies on grid-connection lead times and low-communication training techniques, concluding that federation is a viable stopgap while larger campuses face years-long delays.", "body_md": "A sourced model and short report on a single question:\n\nCan Europe stand up a sovereign frontier-class AI model now, by federating the public compute it already owns, while the gigawatt datacenters it is planning take years to connect to the grid?\n\nThe answer the model gives is yes, as a stopgap. Europe already operates tens of exaflops of public AI compute across the EuroHPC supercomputers and the national AI Factories. A 1 GW campus, by contrast, waits a mean of 7.6 years for grid power. Federated with low-communication (DiLoCo-style) training, the compute Europe already has can deliver a frontier-class model around 2028, against around 2033 for a new gigawatt campus.\n\nThe report is (built\nfrom\n\n`paper/compute-at-home.pdf`\n\n[). It is a short, sourced read aimed at a general audience. Title: \"Do We Need OpenAI or Anthropic? Europe Has Tens of Exaflops at Home.\"](/sammysltd/euromesh/blob/master/paper/compute-at-home.md)\n\n`paper/compute-at-home.md`\n\n```\neuromesh/\n├── README.md\n├── requirements.txt\n├── paper/\n│   ├── compute-at-home.md / .pdf   the report\n│   ├── grid_queue_dataset.md       sourced 1 GW vs 40 MW grid-connection lead times\n│   ├── eurohpc_substrate.md        sourced EU public-compute inventory + \"is it enough\" math\n│   ├── build_pdf.sh, _report.typ   PDF build (pandoc + typst)\n│   └── figures/                    generated charts (PNG + SVG)\n└── model/\n    ├── MODEL_SPEC.md               the model specification (equations, params, invariants)\n    ├── RESULTS.md                  full results, scenarios, sensitivity, caveats\n    ├── run.py                      regenerates every CSV and figure\n    ├── src/                        the three-layer model (efficiency, ramp, regions)\n    ├── params/                     hardware.yaml, training.yaml, regions.csv + SOURCES\n    ├── results/                    generated CSVs (do not hand-edit)\n    └── tests/                      pytest suite (52 tests) + invariant self-checks\n```\n\nThree layers. Layer 1 is the per-FLOP efficiency of low-communication training (how much the DiLoCo penalty costs). Layer 2 is time-to-availability (when sites energize and how fast cumulative compute accrues). Layer 3 is a per-region scorecard on time, cost, carbon, and feasibility. The headline result is set almost entirely by Layer 2: it reduces to one inequality, the federation wins if its sites are online before a gigawatt campus is. The training efficiency penalty is second-order, confirmed by the sensitivity tornado.\n\n```\npython3 -m venv .venv\n.venv/bin/pip install -r requirements.txt\n.venv/bin/python -m model.run          # regenerates all CSVs in model/results and figures in paper/figures\n.venv/bin/python -m pytest model/tests/ # 52 passed\nbash paper/build_pdf.sh                 # rebuilds paper/compute-at-home.pdf (needs pandoc + typst)\n```\n\nThe run is reproducible from a clean tree: deleting every output and re-running exits 0 and regenerates everything.\n\n**Grid-connection lead times:**`paper/grid_queue_dataset.md`\n\n, seven regions, per-region primary sources, anchored by the AWS \"up to seven years\" statement and the IEA 2-to-10-year range, with limitations stated.**EU public compute:**`paper/eurohpc_substrate.md`\n\n, the EuroHPC flagships and the 19 AI Factories, accelerator counts and the training-time math.**Model parameters:**`model/params/SOURCES.md`\n\nand`model/params/SOURCES_hardware_training.md`\n\n, with confidence tags.\n\nThe point of this repo is clarity, not novelty. The thesis rests on grid-queue\nlead times, which are sourced central estimates rather than observed figures (no\nEuropean operator has yet energized a 1 GW point load). The compute is owned but\nnot yet usable for one coordinated run: the EuroHPC machines are shared,\nbatch-scheduled, and heterogeneous, so the addressable fraction is a political\ndecision rather than a hardware fact. Frontier-scale distributed training is\nunproven above about 10B parameters today, so the target is a credible\nfrontier-class model rather than a guaranteed 405B. All of this is in\n`model/RESULTS.md`\n\nand the report's caveats section. Figures and dated events are\nas of June 2026. This is an independent model and analysis, not peer-reviewed.", "url": "https://wpnews.pro/news/show-hn-can-europe-train-a-frontier-ai-model-on-the-compute-it-owns", "canonical_source": "https://github.com/sammysltd/euromesh", "published_at": "2026-06-15 13:31:09+00:00", "updated_at": "2026-06-15 13:39:10.032173+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "ai-infrastructure", "large-language-models", "ai-policy"], "entities": ["EuroHPC", "AI Factories", "DiLoCo", "AWS", "IEA", "OpenAI", "Anthropic"], "alternates": {"html": "https://wpnews.pro/news/show-hn-can-europe-train-a-frontier-ai-model-on-the-compute-it-owns", "markdown": "https://wpnews.pro/news/show-hn-can-europe-train-a-frontier-ai-model-on-the-compute-it-owns.md", "text": "https://wpnews.pro/news/show-hn-can-europe-train-a-frontier-ai-model-on-the-compute-it-owns.txt", "jsonld": "https://wpnews.pro/news/show-hn-can-europe-train-a-frontier-ai-model-on-the-compute-it-owns.jsonld"}}