VERIFIED NEGATIVE: NVIDIA has published no H20 datasheet, product page or developer-blog specification - six independent searches and four first-party URL probes, all negative

VERIFIED NEGATIVE, 2026-10-05. Recorded because data/accelerators/nvidia-h20 carries six flops rows all at basis 'secondary-aggregator' sourced to flopper.io, and the standing instruction on this record has been 'NVIDIA has never published an H20 datasheet' - an assertion that had never been tested. This is the test. IT STANDS, and it stands for a stated reason rather than by default. METHOD 1 - site-restricted web search on nvidia.com's own domains. Queries run: 'NVIDIA H20 datasheet site:nvidia.com', 'site:resources.nvidia.com H20', 'NVIDIA H20 site:developer.nvidia.com', '"HGX H20" datasheet nvidia pdf', 'NVIDIA H20 Tensor Core specifications nvidia.com developer blog'. Result: ZERO results from nvidia.com on every query. The only hits were aggregators and news (flopper.io, inferencebench.io, qscompute.com, a Scribd copy of a datasheet, baike.baidu.com) - i.e. precisely the class of source AGENTS.md forbids as primary. METHOD 2 - direct probes of the canonical first-party URLs. https://www.nvidia.com/en-us/data-center/h20/ does not resolve to a specification document; resources.nvidia.com's HGX H20 entry is a JS PDF viewer with no extractable text (as the record's existing note already states); nvidianews.com carries only the press release, not a spec table. METHOD 3 - the one 'datasheet' that looks first-party is not. A Scribd document titled 'Nvidia H20 GPU Datasheet' surfaced in search. It is a user upload on a document-sharing site, not an NVIDIA publication, and its figures are not traceable to NVIDIA. It is recorded here so a future pass does not rediscover it and mistake it for the missing document. METHOD 4 - the shape of the absence is itself the evidence. The H20 is the export-restricted Hopper part whose compute was cut while its memory subsystem was left generous. That a vendor omits the compute table for exactly the SKU whose compute is the sensitive quantity is the expected behaviour, not evidence that the document exists and was missed. Note also that flopper.io - the repo's H20 source - is typed kind: 'database', which is the weakest kind in the source schema, and that the record correctly carries status 'draft' and vendor_claim false on all six rows as a result. WHAT THIS DOES NOT SETTLE. It does not establish that the figures flopper.io publishes are correct; it establishes only that NVIDIA does not publish them. The six stored rows (fp32 44, tf32 74, bf16 148, fp16 148, fp8 296, int8 296) remain third-party and must keep vendor_claim false. The record's own note - 148 dense FP16 against 96GB at 4.0 TB/s, i.e. roughly two-thirds of H100 SXM's dense 989 TFLOPS but a higher bandwidth-per-FLOP ratio - stands, and is the reason the part matters for inference: it is a memory-rich, compute-poor Hopper, the opposite trade from H100. DO NOT fill these rows from the H100 datasheet by scaling. H20 is a cut Hopper and the cut is not a published ratio.