P035 Capture volume, storage and recomputation infrastructure
What capture bandwidth, storage and recomputation infrastructure does tap-based verification require, and can sampling cut it to a feasible level?
Capture volume, storage and recomputation infrastructure.
Status
Analyzed Desk estimates of the quantities asked: sampling 0.01% of a scalable unit's 256 400G scale-out links through a 1x256 optical circuit switch to one FPGA, about $25,000 per unit or about 0.2% of an inferred $12M unit price; storage-fabric capture sized from bursty traffic (about 40 GB/s peak, 3 GB/s average); Amodo's SITREP gives a worst case of one recomputation NIC per inference front-end NIC and tags sampling taps 'Not started'; the author reports that typical north-south edge traffic composition was unavailable to him. Nothing built or measured; this survey's cost goal of 1% of facility cost is met only on paper.
A sampling rate of 0.01% of the compute fabric traffic would mean about 20Gb/s of traffic to timestamp and hash (and maybe parse).
Posed by
- AI 2040 (2026), Get Involved in Verification:
Capturing all frontend traffic can work without new R&D, but needs a lot of infrastructure plus work on network topologies and sampling.
(State of Play, network taps, L45) (source states the problem but not the quantity the question asks for)
Also raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Cankaya (2026), A System Overview for Near-Term, Low-Trust AI Compute Verification (v0.2); Cankaya (2026), Research Note: The Fundamentals and Feasibility of Secure Network Taps for Verifying AI Datacenter Use; AI Security Forum, Requests for Discussion; Amodo (2026), AI 2040 Plan A — Verification SITREP; Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P035.1 | How many recomputation-server NICs are needed per 1,000 inference frontend NICs for full capture versus a given sampling rate? | partly answered |
| P035.2 | What capture-bandwidth reduction can sampling or packet-hashing taps achieve while still letting the verifier recompute sampled packets? | open |
| P035.3 | What storage (TB/day) does retaining an inference cluster's full tapped stream require, and is it feasible? | open |
| P035.4 | What is the ratio of total north-south edge traffic to inference-API traffic (and its composition) in a production AI datacenter? | open |
| P035.5 | What egress bandwidth overhead (%) does verification logging (network exhaustion) add? | open |
| P035.6 | For off-site checkpoint verification during a frontier training run, what egress bandwidth is needed and what slowdown results? | open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P035.7 | For batched delivery of captured data, what trusted storage medium and doubly-observed retrieval procedure suffices? |
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 P035.2: What capture-bandwidth reduction can sampling or packet-hashing taps achieve while still letting the verifier recompute sampled packets?
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