Newsletter - Beta
Artificial Intelligence
September 22, 2026
September 22, 2026
At its Apsara Conference on September 22, 2026, Alibaba Cloud unveiled the Zhenwu V900 AI processor and a roadmap projecting future Qwen models at 5 to 10 trillion parameters, but the chip is not scheduled for commercial release until Q1 2027 and no independent benchmarks exist for its claimed 3x performance gain. Qwen 4 is currently in training with no disclosed architecture, parameter count, or release date. Decision-makers tracking China's AI hardware and model development should note that the announced figures are company claims and forward-looking projections, not verified results.
A September 16 report from The Information, relayed by Reuters, describes Apple considering a multi-chip M8 Ultra AI inference server with possible Nvidia NVLink Fusion integration, targeting 2029 at the earliest. Neither Apple nor Nvidia has confirmed the project, and no reviewed SEC filing, TSMC earnings presentation, or Foxconn release names it. The available evidence supports a reported internal proposal, not an announced product.
A September 8 Form 8-K reveals that Qualcomm granted an Amazon affiliate a warrant to purchase up to 25 million shares at $161.26 each, with vesting tied to purchase orders and actual payments up to a $60 billion ceiling — not a guaranteed revenue commitment. The filing confirms a multi-generational collaboration covering custom AI inference silicon and optical connectivity but discloses no product names, specifications, delivery timelines, or tranche thresholds. Qualcomm has also not publicly attributed any portion of its stated $15 billion fiscal-2029 data-center revenue target to AWS in reviewed materials.
Reuters confirmed on August 26, 2026, that its news and multimedia content dating back to 1987 is now available on Snowflake Marketplace, structured for enterprise AI use cases including RAG pipelines and internal chatbots. The listing covers Reuters journalism in five languages across 58 cloud regions, but broader Thomson Reuters professional content such as Westlaw, tax, and regulatory datasets is not part of the Marketplace offering. A separate June 2026 Snowflake announcement described Thomson Reuters using Snowflake internally across 37,500 governed tables, a distinct arrangement that does not involve selling professional content through the Marketplace.
GlobalFoundries and RAAAM Memory Technologies taped out a Gain-Cell RAM test chip on GF's FDX fully depleted silicon-on-insulator platform on September 2, 2026, reporting design-stage projections of 40 percent less memory area and up to 60 percent lower power than SRAM, with lead-customer design access planned for early 2027. The figures are company-reported bitcell estimates; no measured silicon results, specific FDX process node, or production qualification have been disclosed. NXP Semiconductors is named as evaluating the technology but has not been announced as a production customer.