P049 Bounding unmonitored and undeclared compute
Given stock accounting of accelerators, what is the tightest bound on undeclared compute, and what is the smallest undeclared training run the regime can detect?
Stock accounting of accelerators and the resulting bound on undeclared compute and on what the regime can detect.
Status
Analyzed A desk estimate from published production and ownership-tracing figures: roughly a million chips untraced, so an approximate upper bound on the largest hidden cluster; the authors call the exact lower bound hard to estimate from public information. No regime stock audit has been run. The detectable-run clause has only models: Shavit's chip-sampling analysis (unmonitored compute sets an implicit lower bound on detectable run size), Dean's AI 2040 model estimate of about 0.5% of world compute divertible, and the same paper's 10 MW (about 10,000 H100-equivalent) satellite threshold, all paper-only. This survey's run size of 10^24 FLOP is not addressed.
"As noted earlier, 15.7 million of the 20.2 million chips produced can be traced to the owning company. Much of the remaining 4.5 million can likely also be traced, leaving on the order of a million chips unaccounted for. This yields an approximate bound on the size of the largest hidden cluster."
Posed by
- Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring:
these unmonitored compute sources present an implicit lower bound on the minimum training run size that can be verifiably detected by the proposed system.
(§1.2, p.3, L135–136) (source states the problem but not the quantity the question asks for)
Also raised in: Reuel, Bucknall et al. (2025), Open Problems in Technical AI Governance, compute sections; Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Baker, Kulp, Marks, Brundage & Heim (2025), Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment (RAND WR-A4077-1); Scher & Thiergart (2024), Mechanisms to Verify International Agreements About AI Development (MIRI); Petrie, Aarne, Ammann & Dalrymple (2025), Overview of Flexible Hardware-Enabled Guarantees (flexHEG Part 1); Aarne & Petrie (2025), International Security Applications of Flexible Hardware-Enabled Guarantees (flexHEG Part 3); Scher (2026), Verifying Restrictions on Frontier AI Research (MIRI); Cankaya (2025), Catching Misreporting About ML Hardware Use by Turning Noise into Signal; Cankaya (2025), TSMC Most Definitely Has a Golden Record of All AI Chips It Made; Cankaya (2026), Compute Verification FAQs; Canary Institute research; Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification; Koopmanschap & Barten (2026), How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements; Wasil, Reed, Miller & Barnett (2024), Verification Methods for International AI Agreements; AI 2040 (2026), Get Involved in Verification; Amodo (2026), AI 2040 Plan A — How You Can Help; Dean, Verification Plan (AI 2040); Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P049.1 | For each mechanism, what minimum undeclared-compute fraction is detectable at a stated detection rate (a target to be agreed) above statistical noise? | open |
| P049.2 | What end-to-end latency from a violating workload starting to its detection does a candidate verification stack achieve, versus the hours-to-days needed for agreement stability? | open |
| P049.3 | Given the stock of unmonitored compute (legacy, consumer, non-flexHEG), what is the minimum training-run size detectable by the system within a stated time window? | open |
| P049.4 | What fraction of datacenter-class accelerators are sold to individuals or entities below a reporting threshold (and so could be exempted)? | implied by a source, not stated as open |
| P049.5 | Under projected chip and algorithmic-efficiency trends, how many samples per year are needed to catch a fixed-capability run with at least 90% probability before it completes? | implied by a source, not stated as open |
| P049.6 | Does the result that running on fewer chips for longer does not meaningfully reduce detection still hold when snapshot frequency is low enough that sampled chips often hold no snapshot? | a source suggests an answer (not checked) |
| P049.7 | Does the result that splitting a run across more chips does not reduce detection survive non-Poisson (prover-predictable) snapshot timing? | a source suggests an answer (not checked) |
| P049.8 | Does the chip-sampling framework still apply when a cluster is split across several datacenters? | a source suggests an answer (not checked) |
| P049.9 | What false-positive rate do fingerprinting checks (e.g. random projection) show between honest operating modes such as prefill and generation, given false positives must be effectively impossible? | open |
| P049.10 | What fraction of export-controlled accelerators shipped in a given year can be traced to a declared end user, with downstream use and resale controlled, per a published audit or estimate? | a source suggests an answer (not checked) |
| P049.11 | In a pilot audit, what fraction of installed data-centre accelerators (by shipment records) can be located to a named facility and owner, including chips located only retroactively? | implied by a source, not stated as open |
| P049.12 | Using production, trade and ownership data, what is the upper bound on untraced accelerators (H100-eq), and can unknown compute be held to <=~1% of the world total? | partly answered |
| P049.13 | In a given regime design, what fraction of the relevant accelerator stock (by FLOP/s) is located and monitored, and can a declared-fleet audit bound unmonitored chips to a stated fraction? | implied by a source, not stated as open |
| P049.14 | What is the confidence interval on accelerators smuggled or diverted (e.g. into China, or from post-deal production), in H100e? | a source suggests an answer (not checked) |
| P049.15 | What is the historical detection rate of large hardware diversion cases? | a source suggests an answer (not checked) |
| P049.16 | With a fab golden record, what fraction of produced AI dies can a declared-inventory audit reconcile, and with what error bar (% of H100-equivalents)? | open |
| P049.17 | Can customs / HS-code data attribute accelerator shipments to authorised vs unauthorised end uses with stated precision, given chips are not passively detectable at borders? | implied by a source, not stated as open |
| P049.18 | What is the annual failure / write-off / maintenance rate (with CI) of H100/B200-class accelerators in large fleets, and what false-alarm rate does that imply for liveness pings and declared write-offs? | open |
| P049.19 | Has a receiving party verified a batch of delivered accelerators as correct in quantity and untampered (attested identity plus seal inspection) with a documented protocol? | implied by a source, not stated as open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P049.20 | What is the shape of the assurance curve (coverage vs confidence), and where do handoff risk and covert-project risk cross over? |
| P049.21 | Does a given verification solution leave anything meaningful out of scope (completeness)? |
Possibly relevant work found by automated search
Found by AI agents searching for this problem's sub-questions, mostly by keyword and citation in the OpenAlex scholarly database and on the web; a second AI pass reviewed each one, and we spot-checked some by hand. They seemed reasonably related, as far as one can tell from outside each paper's field. Some come from neighbouring fields, such as arms control and financial audit, and bear on the problem by analogy. Not reviewed against the status above; a paper listed here does not change that status. 928 of 1052 quotes checked word for word against the paper's text.
For P049.1: For each mechanism, what minimum undeclared-compute fraction is detectable at a stated detection rate (a target to be agreed) above statistical noise?
For P049.4: What fraction of datacenter-class accelerators are sold to individuals or entities below a reporting threshold (and so could be exempted)?
For P049.6: Does the result that running on fewer chips for longer does not meaningfully reduce detection still hold when snapshot frequency is low enough that sampled chips often hold no snapshot?
For P049.8: Does the chip-sampling framework still apply when a cluster is split across several datacenters?
For P049.9: What false-positive rate do fingerprinting checks (e.g. random projection) show between honest operating modes such as prefill and generation, given false positives must be effectively impossible?
For P049.10: What fraction of export-controlled accelerators shipped in a given year can be traced to a declared end user, with downstream use and resale controlled, per a published audit or estimate?
For P049.12: Using production, trade and ownership data, what is the upper bound on untraced accelerators (H100-eq), and can unknown compute be held to <=~1% of the world total?
For P049.13: In a given regime design, what fraction of the relevant accelerator stock (by FLOP/s) is located and monitored, and can a declared-fleet audit bound unmonitored chips to a stated fraction?
For P049.14: What is the confidence interval on accelerators smuggled or diverted (e.g. into China, or from post-deal production), in H100e?
For P049.18: What is the annual failure / write-off / maintenance rate (with CI) of H100/B200-class accelerators in large fleets, and what false-alarm rate does that imply for liveness pings and declared write-offs?
For P049.19: Has a receiving party verified a batch of delivered accelerators as correct in quantity and untampered (attested identity plus seal inspection) with a documented protocol?
* Full text not reviewed; more likely to contain errors.
Data
- problems.csv — one row per problem
- problem_status.csv — status, scope and evidence per problem
- problem_sources.csv — the posed-by quote and other documents per problem
- leaves.csv — all sub-questions, including those without a checkable form
- leaf_sources.csv — every source row behind each sub-question