Why this paper is being considered
Epoch AI estimates how many H100-equivalent chips were smuggled to China through 2025 (median about 660,000, 90 percent interval 290,000 to 1.6 million), combining diversion evidence from indictments and reporting with resale evidence from China's grey market. It bears on export-control enforcement and on assessments of China's compute. An evaluation would focus on the evidence classes, the extrapolation from detected cases to undetected ones, and the uncertainty treatment.
What an expert evaluation could add: Public explanation and a trade-data follow-up raise useful questions about the estimates. Neither is a fully independent validation. An expert assessment could test uncertainty in diversion estimates and the policy conclusions about export controls. Evidence and commissioning case. AI-assisted judgment, 2026-10-07; separate from ratings and completed evaluations.
AI-generated criterion ratings and reasoning
Scores come from AI prioritization; the accompanying explanations include AI-assisted source checks. These are provisional judgments, separate from human ratings and commissioned evaluations.
Decision relevance
7.4/10
The incremental welfare contribution is a public, probabilistic estimate of smuggled AI compute reaching China, which could improve decisions about export-control design, enforcement resources, liability for intermediaries, chip tracking, cloud-compute controls, and international AI-governance strategy. The welfare pathway is mostly through catastrophic-risk and geopolitical-risk channels rather than direct near-term health or income gains, with large possible stakes but substantial uncertainty about whether public estimates change decisions beyond existing private intelligence. LMIC and animal-welfare mechanisms are indirect; the main VoI is in clarifying whether compute restrictions are functioning well enough to be a credible governance tool.
- Paper claim to check Between 290,000 and 1.6 million H100-equivalents were smuggled to China through 2025, with a median estimate of 660,000. Source description
Value of added scrutiny
8.5/10
The report has circulated on social media and sits among a cluster of policy and press work on chip smuggling, including a CNAS report and a Power Law post. We confirmed the related items exist but did not confirm that they cite this report. No independent methodological review surfaced.
Timing
9.0/10
The report was published on April 29, 2026, roughly five months before this assessment, and appears to be an unpublished policy/research report rather than a peer-reviewed paper. That is early enough for feedback to matter, especially because export-control and compute-governance policy windows are active and the key modeling choices are still contestable. A targeted search found mentions, mirrors, and related estimates, but no clear substantial independent expert review of this specific report; that absence is not proof of no scrutiny, but it supports high marginal evaluation value.
- Source record Publication-stage and date evidence should be checked in the linked paper record. TARGETED_CURATED
Methodological potential
6.8/10
The evidence base combines indictments, investigative reporting, and gray-market observations, so many inputs are hard to verify independently. An evaluation would need to scrutinize detection-rate priors, case-level uncertainty, possible double-counting between diversion and resale evidence, H100-equivalent conversions, the handling of failed or alleged diversions, and whether the resulting estimate materially improves on CNAS, ChinaTalk, private government intelligence, or firm-level information.
- Paper claim to check Between 290,000 and 1.6 million H100-equivalents were smuggled to China through 2025, with a median estimate of 660,000. Source description
Prominence
6.5/10
The scoring model estimated prominence from the paper's venue, authors, institutional setting, and visibility. The model did not supply a criterion-specific explanation. Current public-attention status: Visible circulation within a broader policy conversation; direct citation not confirmed.
- Publisher record Epoch AI published the report with its headline estimate. The publisher is also the author's employer. Epoch AI
Likely influence
7.5/10
The scoring model estimated how far the findings could shape later research or decisions, without supplying a criterion-specific explanation. Current public-attention status: Visible circulation within a broader policy conversation; direct citation not confirmed. See public-attention evidence below.
- Publisher record Epoch AI published the report with its headline estimate. The publisher is also the author's employer. Epoch AI
What the paper says
Source description · Summary and key-takeaways text copied verbatim from the Epoch AI report page; whitespace normalized. Not a formal abstract.
We estimate that between 290,000 and 1.6 million H100-equivalents (H100e) were smuggled to China through 2025. Our median estimate of 660,000 H100e would be roughly a third of China's total compute. Substantial quantities of AI chips have been sent to China in violation of US export controls. Evidence of diverted or missing chips, drawn from indictments and investigative reporting, points to nearly 300,000 Nvidia H100-equivalents by the end of 2025. This would equal roughly a quarter of the compute China acquired through legal channels or domestic production. Because much smuggling goes undetected, the true total is likely higher. We estimate, with 90% confidence, that between 290,000 and 1.6 million H100-equivalents of compute were smuggled through the end of 2025. Our median estimate of 660,000 represents roughly 3% of the global compute stockpile, comparable to what xAI, a leading US AI lab, had at the time. The upper bound of our estimate would mean that, by the end of 2025, the majority of China’s AI compute had been smuggled. We are uncertain about many variables, notably the magnitude of undetected smuggling and the proportion of chips allegedly diverted or missing that ultimately reached China.
Claims to check
- Between 290,000 and 1.6 million H100-equivalents were smuggled to China through 2025, with a median estimate of 660,000.
- Publicly alleged diverted or missing chips amount to nearly 300,000 H100-equivalents by the end of 2025, roughly a quarter of China's legally acquired or domestically produced compute.
- Smuggling may represent a large enough share of China's AI compute stockpile to materially affect assessments of export-control effectiveness.
Methodological or theoretical issues flagged for evaluation
The evidence base combines indictments, investigative reporting, and gray-market observations, so many inputs are hard to verify independently. An evaluation would need to scrutinize detection-rate priors, case-level uncertainty, possible double-counting between diversion and resale evidence, H100-equivalent conversions, the handling of failed or alleged diversions, and whether the resulting estimate materially improves on CNAS, ChinaTalk, private government intelligence, or firm-level information.
Opus 5.5 re-evaluation (experimental)
Read with care. This is an experimental re-evaluation by a different model (Claude Opus 5.5, 2026-10-01). Across a calibration sample Opus scored about 15 points lower than GPT-5.5, so 15 points are added to compare it with GPT-5.5 scores. The calibration is a small sample (n=44) and is not human-validated, and the two models disagree about the order of papers within the top group. The AI score at the top of this page is the usual evaluation-priority score shifted by the difference between the adjusted Opus score and the earlier holistic score; the synthesis uses that shifted value, and the unshifted score is shown beside it. On the dashboard you can switch the Opus scores off. Read the methods note.
Opus rationale (AI-generated)
This is an Epoch AI report that quantitatively estimates how much AI compute has been smuggled to China in violation of US export controls. It builds a detected-cases baseline from indictments and investigative reporting, then models undetected smuggling, reaching a median of 660k H100e (90% CI 290k-1.6M), about a third of China's compute. It is squarely relevant to live AI governance decisions: BIS enforcement resourcing, chip location-verification legislation, controls on transshipment hubs and licensing of downgraded chips. Policymakers, CNAS, IFP, CSET, RAND, GovAI and AI-safety funders such as Open Philanthropy are likely to use and cite it. It is a web-published research report that has not been through independent peer review, and Epoch's numbers tend to travel fast into congressional and media debates, so this is a classic 'consequential but unvetted' case. Concerns: this is not economics or social science in the usual Unjournal sense. It is an assumption-heavy estimation exercise, not causal inference, and the headline rests largely on a weakly identified undetected-smuggling multiplier. Evaluators would need export-control and semiconductor supply-chain expertise, which is a different pool from our usual economists. AI/tech papers have also historically fared less well with our assessors. Recommend monitoring, with a lean toward commissioning if a suitable evaluator pairing can be found (for example an empirical economist with experience estimating hidden or illicit trade flows, plus a compute-governance specialist). The methods of illicit-trade estimation, such as mirror statistics and detection-rate modelling, are a natural angle for an economics evaluation.
Dashboard details and provenance
Full AI dashboard scoring rationale
This seems high value for AI governance decisions around compute export controls, enforcement, chip tracking, cloud access, and the interpretation of compute as a policy trigger. It is a recent Epoch AI report, not an externally peer-reviewed publication, and it makes a consequential empirical claim using a Monte Carlo model built from indictments and investigative reporting, so independent evaluation would add clear value. The main concern is that the evidence base is necessarily thin and partly journalistic, so an evaluation would need to focus on assumptions, double-counting, uncertainty calibration, and whether the estimate is actually decision-changing for policymakers who may have private intelligence.
AI decision-relevance rationale
The incremental welfare contribution is a public, probabilistic estimate of smuggled AI compute reaching China, which could improve decisions about export-control design, enforcement resources, liability for intermediaries, chip tracking, cloud-compute controls, and international AI-governance strategy. The welfare pathway is mostly through catastrophic-risk and geopolitical-risk channels rather than direct near-term health or income gains, with large possible stakes but substantial uncertainty about whether public estimates change decisions beyond existing private intelligence. LMIC and animal-welfare mechanisms are indirect; the main VoI is in clarifying whether compute restrictions are functioning well enough to be a credible governance tool.
AI timing assessment
The report was published on April 29, 2026, roughly five months before this assessment, and appears to be an unpublished policy/research report rather than a peer-reviewed paper. That is early enough for feedback to matter, especially because export-control and compute-governance policy windows are active and the key modeling choices are still contestable. A targeted search found mentions, mirrors, and related estimates, but no clear substantial independent expert review of this specific report; that absence is not proof of no scrutiny, but it supports high marginal evaluation value.
Intake, review, and crux connections
AI governance and the economics of AI policy: priority papers
· 2026-09-30
Papers identified in The Unjournal's September 2026 AI-governance scoping (David Reinstein's internal planning). On integration (2026-09-30), titles, authors, dates and abstracts were re-resolved from canonical sources (arXiv, Crossref, NBER, or the publisher's own page), never from the scoping notes; the papers were deduplicated against the dashboard and scored by the standard Codex GPT-5.5 (medium reasoning) subscription path. Five related papers already on the dashboard but still awaiting a genuine model score were re-scored by the same path and are labeled as surfaced existing records
The papers were identified in The Unjournal's September 2026 AI-governance scoping (David Reinstein's internal planning). Several are central to live policy debates but were missing from the dashboard or not yet scored. Inclusion is not an endorsement or a completed Unjournal team decision.
Public attention and use
The report has circulated on social media and sits among a cluster of policy and press work on chip smuggling, including a CNAS report and a Power Law post. We confirmed the related items exist but did not confirm that they cite this report. No independent methodological review surfaced.
Publisher record listing / discoverability
Epoch AI published the report with its headline estimate. The publisher is also the author's employer.
Source: Epoch AI · Relationship: author institution
Social-media circulation visible circulation
A public post summarised the estimate and compared it with a large AI lab's compute holdings, indicating circulation beyond Epoch's own channels.
Source: X · Relationship: independent social media
Related policy context related context, not uptake
A think-tank report on the same policy problem. We did not confirm that it relies on the Epoch estimate.
Source: CNAS · Relationship: external commentary
What the search did not establish
- No exact-title EA Forum or LessWrong discussion surfaced in the targeted search.
- We did not find a Congressional, BIS or Commerce document citing the estimate by name.
- The most decision-relevant next signal is whether enforcement agencies or other analysts publish a competing estimate.
Targeted search checked 2026-10-01. Search scope: Targeted exact-title searches across the open web, institutional and author pages, news and public social-media results, EA Forum/LessWrong, and policy/white-paper contexts. Evidence records distinguish commissioning or report use from independent discussion, media attention, indexing, and post-publication policy use. A search miss is reported as uncertainty, not proof of absence. Entries with check_depth "quick" rest on roughly one to two searches and should be read as especially uncertain; entries checked on 2026-10-01 were added for the shortlist of papers without human ratings.
Human feedback so far
The privacy-safe aggregate contains 3 current ratings: 3 team and 0 public. The human mean is 50.0/100. The team has not made a final prioritization decision.
No written discussion is public. Private and team-only text is never copied to this page.