The A800: a part designed to the threshold
Late in 2022 NVIDIA began offering Chinese customers the A800, a variant which, in the company's own words in its quarterly report, is "not subject to the new license requirements". For the first time an export-control threshold became an input to the design of a chip.
Why it matters
A numeric threshold has a property a list of companies does not: you can fit yourself under it. The 2022 rule required two quantities to be exceeded at once, so it was enough to lower one - the restriction stops applying and the computing power stays. This is not a breach of the rule; it is a reading of it.
There is one primary source here: NVIDIA's quarterly report for the quarter ended 30 October 2022, filed on 18 November. It says: "We have engaged with customers in China to satisfy their demand with products not subject to the new license requirements, such as our new A800 offering." The same paragraph gives the company's own reading of the rule, and it is conjunctive: the requirements reach circuits achieving "both peak performance and chip-to-chip I/O performance" at A100 level. What this record does not claim. Neither a sale date nor any specification of the A800 could be found in a primary source. What is widely repeated is 7 November 2022 and a cut in NVLink bandwidth from 600 to 400 gigabytes a second; the first comes from the trade press, so does the second, and neither is confirmed by a document from NVIDIA or a regulator. The figure 400 gigabytes a second does appear in an NVIDIA specification, but for the "A800 40GB Active" - a workstation card with an NVLink bridge between two PCIe boards, a different product, and taking it from there would give a server die the specification of a desktop card. So the record sits on the date of the document rather than of the event: the anchor is 18 November 2022, the day the A800's existence was first put in writing. The event itself happened earlier, within the autumn of 2022.