AI-assisted working evaluation: Epoch AI compute smuggling to China
Evaluation date: 2026-10-05
Last substantive update: 2026-10-12 (evidence update; no model rerun)
Last attention check recorded: 2026-10-05
Storage reconciliation: 2026-10-12
AI-assistance disclosure. This working assessment supports The Unjournal’s prioritization process. Its judgments and ratings have not been adopted as a commissioned human evaluation or a team decision. A human evaluator should verify the factual claims, methods, and ratings before signing or submitting an official evaluation.
The substantive analysis below preserves the fuller existing draft. The attention date records the earlier public-brief check; storage reconciliation does not imply a new evidence search.
Evidence update, 12 October 2026. A new subsection under “External checks and later evidence” adds a criminal case announced by the US Department of Justice on 1 October 2026. This is an evidence update, not a model rerun: the simulation was not reproduced, and the provisional ratings are unchanged.
Executive summary
Epoch AI's Diversion and resale: estimating compute smuggling to China is a strong candidate for public evaluation because it converts fragmented allegations, indictments, gray-market evidence, and chip-mix assumptions into an explicit probabilistic estimate. Its headline result is a median of about 660,000 H100-equivalents (H100e) smuggled to China through the end of 2025, with a 90% interval of roughly 290,000 to 1.6 million. Epoch interprets the median as about one-third of China's total AI compute at that time.
The report is much better than a simple sum of press allegations. It explicitly models two imperfect views of the same latent quantity: diversion from legitimate supply chains and resale evidence inside China. It also acknowledges major uncertainty, including whether alleged diversions actually occurred, whether chips reached China, and how much smuggling remains undetected.
The central weakness is that the median estimate depends heavily on uncertain priors about detection and reporting accuracy, plus heavy-tailed assumptions about gray-market vendor counts and sales volumes. Those are reasonable objects to model, but the available evidence does not tightly identify them. The report is therefore best read as a transparent uncertainty model rather than a measurement with conventional sampling error.
A second important issue is dependence between the two estimation approaches. Diversion evidence and resale evidence are intended to measure the same underlying flow and can draw on overlapping cases, reporting, and market conditions. Averaging draws from the two approaches narrows the combined interval. If their errors are positively correlated, that narrowing can be too optimistic unless dependence is explicitly modeled.
I also found a concrete internal consistency question worth sending the author. The main text describes the detection-rate assumption as having a 90% interval of 10–80%, while a methodology footnote describes a 90% interval of 10–60% with a median around 24.5%. Those are materially different tails for a parameter that strongly drives the smuggling total. The report should clarify which distribution is actually used in the simulation and update the prose or footnote accordingly.
A September 17 follow-up from Epoch adds a potentially useful external check: China recorded about $3.8 billion of server imports from Malaysia where Malaysia recorded about $0.6 billion of exports to China, a discrepancy Epoch says is consistent with roughly 150,000 H100e of diverted compute under its chip-price assumptions. This does not prove smuggling, but it is a promising independent data stream for updating or validating parts of the model.
Bottom line: the report's main value is transparency about an intrinsically hidden quantity. The 660,000 median should not be treated as a directly observed fact. A high-value Unjournal evaluation would reproduce the simulation, stress-test the detection prior and chip-mix assumptions, model dependence between the diversion and resale estimates, and test the model against trade discrepancies and later enforcement outcomes.
Main claims
- Publicly alleged or missing-chip cases amount to nearly 300,000 H100e by end-2025. This is a relatively concrete lower-level evidence compilation, but allegations are not all adjudicated and some chips may not have reached China.
- The true smuggled total is probably higher because many cases are never detected. Highly plausible qualitatively; the magnitude depends on a weakly identified detection-rate distribution.
- A median around 660,000 H100e, 90% interval 290,000–1.6 million, is a reasonable synthesis of diversion and resale evidence. Reasonable as a model-conditioned estimate; precision is sensitive to priors and dependence assumptions.
- This corresponds to roughly one-third of China's AI compute. Arithmetically follows from Epoch's broader compute-stock estimates; uncertainty in both numerator and denominator should be propagated when used in policy discussion.
- Smuggling was a large enough channel to matter materially for export-control effectiveness. Supported directionally. The report does not by itself establish which enforcement or verification intervention has the highest net benefit.
Evidence and methods audit
Diversion model
The diversion side begins with documented or alleged cases and scales them for uncertainty about whether the report is accurate, whether chips reached China, and how much activity remains undetected.
This is conceptually appropriate. The difficult parameter is the detection fraction. A low detection rate implies that observed cases are only a small visible slice; a high detection rate keeps the total closer to known allegations. The report itself states that the evidence does not allow a confident choice across a wide range.
Because the headline result is sensitive to this distribution, the evaluation should show the posterior/combined estimate under several defensible alternatives rather than only the authors' chosen prior. In particular:
- a conservative high-detection prior;
- the authors' baseline;
- a low-detection prior;
- a model where detection improves over time as enforcement and reporting intensify.
The report should also distinguish clearly between alleged diversion, indicted conduct, adjudicated conduct, and physical arrival in China.
Resale model
The resale side estimates gray-market volume using uncertain vendor counts and uncertain per-vendor annual volume. Epoch uses broad lognormal distributions and allows a heavy right tail, especially during periods when the H20 ban increased incentives to source restricted chips.
This is a sensible response to sparse data, but products of two heavy-tailed uncertain quantities can generate very wide and prior-sensitive outcomes. The evaluation should inspect:
- the empirical basis for vendor counts;
- whether sampled vendors are representative;
- whether vendor listings are duplicates or intermediaries selling the same inventory;
- whether advertised quantities correspond to completed transactions;
- the chosen upper-tail behavior for per-vendor volume.
The resale approach also assumes a large share of smuggled compute passes through intermediaries/gray-market resellers. That scaling parameter deserves the same sensitivity treatment as the detection rate.
Combining the two approaches
Epoch's diversion estimate is about 530,000 H100e at the median, with a 90% interval around 180,000–1.6 million; the resale estimate is about 700,000, with a wider interval around 230,000–2.3 million. The published combined estimate averages draws and ends up around 660,000 with a narrower 290,000–1.6 million interval.
This combination is intuitive if the two estimators contain substantially independent noise. But they are not obviously independent:
- the same smuggling episodes can appear in both investigative reporting and gray-market evidence;
- both respond to the same policy shocks;
- later resale evidence can be generated by the same diverted shipments used to anchor the diversion model;
- CNAS/ChinaTalk comparison estimates share some methodological ancestry with Epoch's resale approach.
A public evaluation should estimate or bound the correlation between errors. At minimum, report combined intervals under correlations of 0, 0.25, 0.5, 0.75, and 1, rather than letting the averaging step mechanically imply extra precision.
H100-equivalent normalization and chip mix
Epoch converts many generations of Nvidia accelerators into H100-equivalents. This is necessary for policy comparison but introduces assumptions about relative AI-training usefulness. Epoch's estimate exceeds some earlier estimates partly because it assigns a higher share of newer Hopper and Blackwell chips.
The evaluation should reproduce the estimate in both physical chip counts and several H100e normalizations. It should show how much of the difference from CNAS and ChinaTalk comes from inferred smuggling volume versus chip-mix conversion.
A concrete discrepancy to resolve
The report's main text says the modeled detection rate spans roughly 10–80% at the 90% interval. A methodology footnote gives roughly 10–60%, with a median near 24.5%.
Since the detection rate is one of the most influential assumptions, this is not a cosmetic wording issue. The replication should identify the actual distribution used in code/calculation and trace the headline estimate to it.
External checks and later evidence
Epoch's September 17 data insight provides a useful prospective validation route. It reports that between April 2024 and June 2025, China recorded about $3.8 billion of server imports from Malaysia while Malaysia recorded about $0.6 billion of exports to China. Unit counts were similar, but reported values differed sharply. Epoch argues the gap is consistent with high-end AI servers and could represent about 150,000 H100e.
This is not proof: classification, valuation, origin rules, re-exports, and chip composition can create trade discrepancies. But it is exactly the kind of independent data stream an evaluation should use rather than relying only on the original case/resale evidence.
Other useful validation signals would include:
- later indictments resolving earlier allegations;
- seizure data;
- vendor disappearance or enforcement actions;
- customs microdata where accessible;
- changes in observed gray-market prices after control changes;
- chip-location or server telemetry if any auditable source becomes available.
Evidence update, 12 October 2026: a newly charged diversion case
On 1 October 2026 the US Department of Justice announced the arrest of a California technology-company owner on a three-count federal indictment (returned 29 September): conspiracy to violate export controls, outbound smuggling, and money-laundering conspiracy. Prosecutors allege that more than $300 million of export-controlled servers containing US-made Nvidia GPUs were shipped to Singapore and Malaysia and then forwarded to customers in China. Reports of the indictment place the alleged conduct in 2023–2024, including more than $176 million received from two Malaysia-based shipping companies between January and October 2024 (Interesting Engineering).
How this bears on Epoch's estimate:
- Allegations, not adjudicated conduct. The defendant is presumed innocent. In this report's classification (alleged, indicted, adjudicated, confirmed arrival in China), the case is indicted, without adjudication or confirmed arrival.
- No H100e conversion yet. A dollar value cannot be converted into H100-equivalents without the server configurations, GPU models and counts, and prices. Press coverage cites at least one order for H100 servers, but the full hardware mix is not public.
- Likely overlap, not an independent increment. The alleged shipments fall within the window of Epoch's April 2026 estimate (cumulative smuggling through end-2025) and overlap the April 2024 – June 2025 window and Malaysian route of its September trade-data analysis. If the case was not public when Epoch published in April, its conduct belongs to the undetected share that the diversion model's detection prior is meant to scale up for; the same chips could also appear in resale evidence and in the trade discrepancy. Adding the case on top of the April estimate or the September trade figure would double-count.
- Evidence about detection lags. A case concerning conduct inside the estimation window that became public only after the estimate was published is relevant to the detection-rate parameter. Recording such later-revealed cases systematically, with dates of conduct, disclosure and resolution, could help calibrate a time-varying detection model.
- The detection-interval discrepancy still matters. This update did not re-read Epoch's full report, so the discrepancy noted above (main text 10–80% versus footnote 10–60% at the 90% interval) is treated as unresolved. Because new cases bear directly on that parameter, the distribution actually used should be identified first.
Before testing how this case shifts the estimate, a case-level crosswalk is needed: for each case in Epoch's dataset and each later case, the dates of conduct and disclosure, evidence type, legal status, hardware (GPU model, count and server configuration), route, and possible overlaps with other cases, resale evidence and the trade discrepancy.
Current attention and decision relevance
The headline estimate continues to be repeated in export-control discussion, and Epoch has continued publishing new trade-data evidence in September 2026. That increases the value of clarifying exactly what the estimate does and does not establish.
The decision relevance is straightforward: estimates of leakage can inform debates over enforcement resources, chip-location verification, intermediary controls, server-level controls, and the relative role of onshore/cloud access. But the paper estimates how much leakage may have occurred, not the comparative effectiveness or cost of those interventions. Policy conclusions require additional evidence.
Highest-value robustness work
- Reproduce the full Monte Carlo model from public inputs.
- Publish a tornado plot or variance decomposition showing which priors drive uncertainty.
- Run alternative detection-rate priors and time-varying detection.
- Resolve the 10–80% versus 10–60% detection-interval discrepancy.
- Separate allegations, indictments, adjudicated cases, and confirmed arrival.
- Model positive error correlation between diversion and resale estimates.
- Deduplicate evidence that enters both sides of the model.
- Recompute in raw chip counts and alternative H100e conversions.
- Stress-test the vendor-count and per-vendor-volume heavy tails.
- Use the Malaysia trade discrepancy and later cases, such as the October 2026 indictment, as out-of-sample checks, after a case-level crosswalk that prevents double-counting.
Questions for the author
- Which detection-rate distribution is actually used in the simulation: the main-text 10–80% 90% interval or the footnote's 10–60% interval?
- Is the simulation code and underlying case-level dataset public or available for independent reproduction?
- How much of the variance in the 660,000 estimate is attributable to detection, reporting accuracy, arrival probability, vendor count, vendor volume, reseller share, and chip mix?
- How are duplicate allegations or cases that also appear in resale evidence handled?
- What correlation between errors in the diversion and resale models do you think is realistic?
- How would the combined interval change under substantial positive dependence?
- How much of Epoch's higher H100e estimate relative to CNAS/ChinaTalk comes from volume versus newer-chip composition?
- Can cases be classified by allegation / indictment / conviction / confirmed physical arrival?
- How would you incorporate the September Malaysia trade discrepancy into an updated estimate without double-counting evidence?
- Which future observable would most strongly update your belief about the detection rate?
Evaluation ratings
- Importance / decision relevance: 4.5 / 5
- Transparency of conceptual method: 4 / 5
- Quality of direct empirical anchors: 3.5 / 5
- Identification of key latent parameters: 2.5 / 5
- Robustness shown: 3 / 5
- Reproducibility: 3 / 5 pending public code/data verification
- Interpretability of headline uncertainty: 3 / 5, with dependence and detection-prior questions
- Value of further public evaluation: 4.5 / 5
Overall assessment
Epoch does a valuable thing: it makes an unobservable policy quantity explicit enough to argue about. The report is notably careful about uncertainty and about the distinction between observed allegations and latent total smuggling.
The main risk is false precision. The 660,000 median is generated by a model whose most consequential parameters are only weakly identified, and combining two related estimators can make the final interval look more precise than the evidence warrants.
The right Unjournal contribution is therefore not to decide whether "one-third" sounds high or low. It is to reproduce the uncertainty model, expose the influence of each assumption, correct any internal inconsistency, and show how much the conclusion changes when dependence and alternative priors are treated explicitly.
Sources
- Epoch AI, Diversion and resale: estimating compute smuggling to China
- Epoch AI, Trade data is consistent with more than $3 billion of chips smuggled into China via Malaysia, 17 Sep 2026
- US Department of Justice, California man arrested for smuggling more than $300 million of export-controlled computer servers to China, 1 Oct 2026
- Secondary coverage of the indictment: Interesting Engineering