Fable 5 offline, OpenAI & Anthropic both IPO-filed, Gemini 3.5 Pro imminent.

Tencent & National IC Fund lead one of the largest startup financings in Chinese history. DeepSeek-V3.1 update pushed June 9 with faster inference and stronger agent capability. R2 remains unreleased; V4 (April) is still the flagship.

Multimodal model handling images and video with 1M-token context at ~1/6 the per-token cost of Qwen 3.7 Max. Exposed via OpenAI-compatible APIs across Beijing, Singapore, and US-Virginia endpoints — squarely targeted at enterprise agent workloads.

Hunyuan as model base, Yuanbao as consumer assistant, WeChat Agent as in-ecosystem execution layer. Tencent Cloud documentation also confirms legacy Hunyuan 2.0 endpoints sunset June 26, pushing developers toward Hy3-preview (~40% reasoning-efficiency gain).

744B-parameter MoE architecture with usable 1M-token context (output up to 131K tokens), dual High/Max thinking-effort system, no regional usage restrictions. Released two days after the US Commerce order to Anthropic; HK parent shares jumped 48% intraday.

Switzerland's Federal Roads Office (FEDRO) granted AmiGo — the joint service from Apollo Go and PostBus — a special permit for Level 4 operations, with open-road trials in an ~80 km² Bern-area zone starting June 1. First European autonomous-driving approval for Baidu. On the model side, ERNIE 5.1 (May) remains flagship.

Following ~4x post-IPO run since the January HK debut, MiniMax disclosed it has entered the formal tutoring agreement required before a STAR Market filing — moving toward A+H structure. Product side: Hailuo 2.3 Fast targets batch creation (native 1080p, improved facial realism).

See story #10. ChatGPT now auto-routes intent across coding, agents, image, and external services. GPT-5.2 removed June 12; GPT-5.5 Instant got quality/style refresh with new in-chat writing and coding blocks.

Released June 9 via Claude API + AWS Bedrock at $10/$50 per million input/output tokens. Mythos 5 was Project Glasswing-limited. US Commerce export-control order on June 12 forced suspension. Same week: $965B IPO filing + $47B annualized revenue disclosure.

DeepMind named a core partner on Genesis, a DOE-led initiative to accelerate scientific discovery using frontier AI on national-lab compute. Gemini 3.x family continues to roll out; Gemini 3.5 Pro release imminent (see story #4).

No major Meta release in June 9–16. Next-generation "Avocado" model (Llama successor) timelines continue slipping; internal restructuring underway with a reported shift from open-source to closed-source frontier development. Llama 4 Behemoth, originally slated for June, remains unreleased.

Following Build 2026 (June 2–3), Microsoft pushed Claude Sonnet 4.5, Haiku 4.5, Opus 4.1 into public preview on Microsoft Foundry with Entra auth and Azure billing (MACC-eligible). Hosted Agents (hypervisor-isolated, per-agent Entra ID, built-in content safety) slated for GA by end of June. MAI-Thinking-1 / MAI-Code-1 rolling out.

On June 9, Bloomberg reported xAI moved Starlink's Jack Garabedian to run the Grok human-data/training team, replacing Diego Pasini. On June 10, ex-engineer Devin Kim sued xAI in California state court alleging he was fired for pushing safety guardrails on Grok. Grok V9-Medium (1.5T parameters) teased for mid-June release; Grok Imagine expanding from images into video generation.

Vera Rubin NVL72 rollout to Fairwater AI factories continues through 2026. Jensen Huang's RTX Spark "superchip" announcement (laptops/desktops shipping this fall via Dell, HP, Lenovo, Microsoft Surface, MSI, ASUS) sent AMD, Intel, Qualcomm shares lower as Nvidia signals it intends to own every layer of the AI stack.

MI355 targets B200/GB200 in volume. MI400X claims 10x MI300X performance, 432 GB HBM4, 19.6 TB/s. 2026 "Helios" rack pairs EPYC "Venice" CPUs, MI400 GPUs, Pensando "Vulcano" NICs, ROCm 7 — 72 GPUs/rack at 260 Tb/s scale-up.

Gaudi 3 still shipping via Dell/HPE/Lenovo/Supermicro (claimed ~50% better inference vs H100). Next-gen Jaguar Shores (HBM4E) slipped to 2H 2027. Intel AI cadence remains the weakest of the five vendors covered this week.

Leading-node demand running 25–30% above capacity in 2026, no relief expected until 2027. Apple has locked >50% of initial 2nm wafers (A20/A20 Pro), squeezing Qualcomm and MediaTek. 2026 2nm tape-outs projected at 4x the 5nm node level.

Total Ascend die shipments could reach 1.6M in 2026 (SMIC enhanced 7nm). Training-focused Ascend 920C delivers >900 TFLOPS BF16 with HBM3 (4 TB/s, up from 910C's 3.2 TB/s). Ascend 950PR ships Q1 2026; 950DT lands Q4 2026 with HiZQ 2.0 HBM (144GB, 4 TB/s) — credible export-control workaround stack.

The Cyberspace Administration of China's Reporting Center launched a specialized channel in June 2026 for citizens to flag AI-related abuses (deepfakes, misinformation, algorithmic harms). Part of the expanded "Qinglang · Rectifying AI Application Chaos" enforcement sweep — signals tougher operational oversight of GenAI service providers already licensed under the 2023 Interim Measures.

(1) Trump AI EO (June 2): 30-day voluntary pre-release safety review framework + federal cyber defense upgrades + DOJ resources for AI crime. (2) Great American AI Act draft (June 4): bipartisan 269-page draft with 3-year preemption of state laws regulating frontier development — direct challenge to CA SB 53 and Colorado AI Act. (3) BIS clarifies (June 1): AI chip export ban extends to Chinese firms' overseas subsidiaries — closes January loophole.

Italy (June 10): Two implementing decrees under Law 132/2025 designating national AI authorities, rules for employment/education/law-enforcement (incl. biometric surveillance), civil liability provisions, and new criminal offenses for AI misuse. Commission (June 1): 60 independent experts will advise the AI Office on GPAI model classification and systemic-risk evaluation ahead of August 2, 2026 GPAI compliance date.

Diplomats, AI labs, and civil society gather at the Palais des Nations to address AI's implications for international peace and security — the most significant multilateral AI governance event of June 2026 and a key venue for shaping post-Paris-Summit norms on autonomous weapons and frontier model risks.

The Multilateral Alignment of Technology Controls on Hardware Act would coordinate US semiconductor manufacturing equipment restrictions with allied jurisdictions, with ASML (Netherlands) and Nikon/Canon (Japan) as primary targets. Adds pressure on EU and Japan to harmonize with US extraterritorial controls.