Qualcomm OnQ blog: 'Qualcomm AI200 Accelerator for Generative AI' (Feb 2023)

Qualcomm's own OnQ engineering blog introducing AI200, read via browser on 2026-10-03 (qualcomm.com serves almost nothing to web_extract - it gates on a Qualcomm ID - so the browser is required). Qualifies as a vendor primary source for the AI200's positioning but NOT for a spec table: it gives no TOPS figure, no memory capacity, no bus width and no TDP. Its substantive content is architectural and commercial: AI200 is described as supporting generative AI and LLMs at the data centre edge with a cloud-inference pricing model, and as building on the Hexagon AI/AI Engine lineage. Because it publishes no throughput number, it cannot be used to fill a flops row - which is exactly why the AI200 record's compute field stays null rather than being back-computed from the AI100's TOPS figure. NOTE the naming continuity trap: the chip is 'Qualcomm AI200' while the inference card is sold as the 'Qualcomm Cloud AI 100' family - Qualcomm kept the Cloud AI 100 card name for the AI200-based card, so 'Cloud AI 100 Ultra' and 'AI200' can refer to the same product generation and the repo should not treat them as two unrelated parts. | DEAD-LINK VERIFIED 2026-10-04. THREE independent GET probes return HTTP 404 (7.6KB body, empty <title>, no AI200 content) - Qualcomm's news CMS has retired the post. The existing note already establishes the important limit: this blog gives positioning, not a spec table (no TOPS, no memory, no bus width, no TDP), so its loss costs no number. The substantive architectural and commercial claims survive only in this record and in the Wayback Machine; archiving was not done, so archived_url correctly stays null. | 2026-10-05 REPAIR: DEATH CONFIRMED AND NOW FINAL - NO REPLACEMENT EXISTS, AND NO FIGURE IS LOST. Re-probed on 2026-10-05: HTTP 404 again, body is an empty <title> React shell with no AI200 content, so it is the CMS retiring the post rather than a bot-block. I then looked for a live replacement on purpose and did not find one that carries this blog's content: Qualcomm's newsroom now carries a different, much later post on the same chip ('Building AI inference that scales: Inside Qualcomm AI200 Rack ...', 2026-03) and a launch PR ('Qualcomm Unveils AI200 and AI250 - Redefining Rack-Scale Data Center ...', 2025-10), but neither is this 2023 OnQ article and neither substitutes for it. Wayback Machine checked explicitly and the check came back EMPTY: the availability API returns archived_snapshots: {} for this exact URL, so there is no archive to point archived_url at and null remains the truthful value. WHAT THIS COSTS: nothing numeric, and that was already the record's own established limit - this blog carries positioning, not a spec table (no TOPS, no memory capacity, no bus width, no TDP), so the AI200's null flops field is unaffected. The record is kept as the citation for the naming-continuity trap (chip 'Qualcomm AI200' vs card family 'Cloud AI 100') precisely because that is a naming fact this record holds and no live page currently states it as plainly. A URL that no longer resolves is a reason to distrust the FIGURE, not to discard a naming observation; readers should treat the architectural/commercial prose as historical.