What HSE actually counted
The figures come from a scientometric study by the HSE University Scientometrics Center titled "BRICS and AI: The Scale of Countries, Organizations, and Partnerships within Leading AI Conferences". It was finalised and published in September 2026, and reported on around 10–11 September by Quantum Zeitgeist and by Skoltech itself.
The methodology matters more than the headline. The authors used Scopus publication data filtered through the ICORE 2026 conference ranking, counting only A*-rated venues — among them NeurIPS, ICML, ICLR, CVPR and ACL. Ten conferences in total, six publication years. It counts papers: not citations, not benchmark scores, not shipped models, not production deployments.
That distinction is the whole story. Skoltech's research spans machine learning, computer vision, NLP, generative models and industrial AI — all legitimate A* territory. But a paper count is an input metric, not an output.
The arithmetic behind the 20%
The study reports percentages rather than absolute national totals, so the absolute numbers have to be derived. Label this clearly as a calculation, not a quoted figure:
If 300+ Skoltech papers equal approximately 20% of Russia's A* AI output, then the national total across those ten venues and six years is roughly 1,500 papers — call it 250 per year. Skoltech's own share works out to about 50 papers per year, or roughly one A* paper per week from a single institute founded in 2011.
The second figure is sharper still. Skoltech outperforms the nearest Russian university by more than 27%, which puts that competitor at roughly 236 papers over the same period. The gap between first and second place in Russian A* AI publishing is therefore somewhere around 65–70 papers — meaningful, but not the order-of-magnitude lead the percentage framing suggests.
Why the concentration is genuinely unusual
A 20% national share from one institution is a strong result, and it is worth saying so without hedging. Skoltech was deliberately built as an English-language, internationally staffed flagship institute in Skolkovo, and the concentration reflects that design: a small number of well-connected senior researchers publishing consistently at venues where acceptance rates sit in the low twenties or below.
The honest frame, though, is the global one. Russia's entire six-year A* output at these ten conferences is on the order of 1,500 papers. The largest of these venues now accepts main-track papers in the thousands per edition. Twenty percent of a national pie is dominance of a small pie.
The European angle: what this touches and what it doesn't
For European readers there are two practical threads, and neither is about the papers themselves.
Regulation. Since 2 August 2026, the European Commission and the EU AI Office hold direct enforcement powers over general-purpose AI models: binding requests for model access, information demands, forced recalls and financial penalties. The voluntary-grace-period reading of the GPAI Code of Practice is gone, and Article 50 transparency obligations — chatbot notices, deepfake labelling, watermarking of synthetic content — are now mandatory rather than self-regulated. Publishing a paper is not a regulated activity. Placing a general-purpose model on the EU market is, and the documentation burden falls on the provider: training data provenance, evaluation methods, risk assessment. Research literature is part of the sourcing trail regulators can ask for, and the AI Office can now compel access to what sits underneath a model.
Compute. EU and US export controls on advanced accelerators have, since 2022, restricted Russian access to the high-end GPUs that frontier training runs depend on. The current field makes the consequence visible: the frontier list is populated by GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash, Grok 4.6, DeepSeek-V4.1-Flash, GLM-5.3-Flash and Mistral's specialist OCR 4.1. No Russian lab appears on it. Research output and model output have diverged, and that divergence is a compute story as much as anything else.
For developers and research groups inside the EU, the boring advice is the correct one: research-security and sanctions screening policies differ between member states, and joint projects with Russian state-affiliated institutions need a compliance check before anything is signed, not after.
Should you care about Skoltech's papers?
Yes, selectively. A* AI papers from any institution reach the same open repositories and the same code releases, and good methods travel regardless of who wrote them. If a technique published at ICLR by a Skoltech group solves a problem in your pipeline, that is useful information and nothing more needs to be attached to it.
What does not survive scrutiny is the leap from "20% of Russian A* papers" to "Russia is competitive at the frontier". The study is a national comparison built on a BRICS framing. It never claims a global ranking, and the global ranking it would produce would be unflattering next to institutions that publish several hundred A* papers a year in their own right. Read it as what it is: a well-executed measurement of one country's academic AI output, with one institute carrying a fifth of the load.
Is Skoltech's 20% share comparable to a Western university's national share?
Not directly. The denominator here is Russia's total A* AI output across ten conferences over six years. A US or German university's share of its national output uses a completely different denominator, different institutions in the sample, and in most cases a much larger national total. The percentages look alike; the underlying quantities do not.
Do the AI Act obligations apply to research papers from Russian institutions?
No. The AI Act regulates placing systems on the EU market and putting them into service in the EU, not publishing research. Obligations attach to the provider or deployer in the EU. If you build a product on a method described in a paper, your obligations come from your own system, your own risk classification, and your own documentation — not from the paper's origin.
Does the study include models, benchmarks or product releases?
No. It counts publications at ten A*-ranked conferences using Scopus data and the ICORE 2026 ranking. Citations, benchmark performance, model releases and deployment are outside its scope. Anyone using the 20% figure to argue about model quality is reading a metric that was never measured.
