Scope: only news published within the last 24 hours by absolute time (UTC cutoff 2026-07-21 21:13 UTC to 2026-07-22 21:13 UTC).
AMD Opens ‘Advancing AI 2026’ — EPYC ‘Venice’ (2nm) in Volume, Instinct MI450, Helios Rack
AMD opened its annual “Advancing AI 2026” conference at Moscone West in San Francisco on July 22. It led with EPYC “Venice” (Zen 6) — the industry’s first x86 server CPU in volume production on TSMC’s 2nm node — alongside the Instinct MI450-series GPU accelerators and the rack-scale Helios system. With adoption commitments from Meta, Microsoft, Oracle, and OpenAI, AMD now fields a complete current-generation AI stack, with no comparable Intel answer expected until mid-2027. One caveat: most 2026 HBM4 supply is already allocated to hyperscalers, so MI455X mass production is pegged around Q2 2027.
Reference: TechTimes · VideoCardz · AMD
South Korea to Launch a Free, Home-Grown National AI Chatbot
South Korea’s government said on July 22 it will launch a free, domestically built AI chatbot service this year to cut reliance on foreign platforms. It is the third government-led AI infrastructure move in July, following Current AI’s $400M public fund (July 13) and the EU Parliament’s EPGenAI Hub (July 20). The pattern is clear: governments increasingly treat AI access as national infrastructure rather than a commercial service to rent from US frontier labs.
Reference: AIToolsRecap
Google Ships Gemini 3.6 Flash, 3.5 Flash-Lite and Flash Cyber — Flagship 3.5 Pro Slips Again, Gemini 4 Pretraining Begins
On July 21 Google shipped three models: Gemini 3.6 Flash ($1.50 / $7.50 per million input / output tokens), the ultra-cheap 3.5 Flash-Lite ($0.30 / $2.50), and a security-tuned 3.5 Flash Cyber restricted to governments and trusted partners. 3.6 Flash cuts output tokens by roughly 17%, pushes its knowledge cutoff to March 2026, and builds in Computer Use. The flagship 3.5 Pro slipped once more, and in the same update Google said it had begun its “most ambitious pretraining run yet” for Gemini 4 — a sign the competitive center of gravity is shifting from top-end performance to the high-volume, low-cost Flash tier.
Reference: MarkTechPost · 9to5Google · AIToolsRecap
Nvidia Publishes Vera Custom Data-Center CPU White Paper and SPEC CPU 2026 Benchmarks
Nvidia published a detailed architecture white paper and SPEC CPU 2026 benchmarks for Vera, its first custom data-center CPU. Vera is built on 88 custom “Olympus” cores (176 SMT threads) that prioritize single-thread performance and memory bandwidth over core density. In an unofficial run it scored 925 on SPECrate integer base, edging a dual-socket EPYC 9755 (898) with fewer threads. The design targets agentic workloads, though the figures come from a reference system since Vera is not yet broadly available.
Reference: Tom’s Hardware · ServeTheHome · NVIDIA Blog
OpenAI Launches ‘ChatGPT for Small Business’ — ChatGPT Work and Codex Pass 10 Million Users
OpenAI on July 21 announced a “ChatGPT for small business” program to help small firms boost productivity and grow. It offers virtual training, in-person U.S. “AI academies,” role-specific webinars and demos across accounting, marketing, and e-commerce, plus access to small-business-focused agents and partners. The company said ChatGPT Work and Codex had passed 10 million users — part of a broader push to extend ChatGPT beyond chat into agentic work.
Today’s Takeaway
On hardware, AMD (EPYC Venice, MI450, Helios) and Nvidia (its first custom CPU, Vera) collided head-on over the data-center AI market. On models, Google pivoted toward a cheaper, more efficient Flash tier and signaled a price war. Meanwhile South Korea’s free national chatbot and OpenAI’s small-business program show AI spreading along two tracks at once — as national infrastructure and as an everyday work tool.
Daily AI Brief 
Leave a comment