The frontier and its rival institution landed the same morning. Google DeepMind shipped Gemini 3.5 Pro on Jul 17 — a flagship it scrapped and rebuilt from the base up, now carrying a 2M-token window and a Deep Think tier gated behind a $250-a-month plan — days behind OpenAI’s GPT-5.6 family and Mira Murati’s open-weight Inkling. In Shanghai the same day, Xi Jinping delivered his first World AI Conference keynote, calling AI “a symphony of international cooperation, not a solo performance by a single country” and standing up WAICO, a 29-nation bloc headquartered in China with 5,000 training slots pledged to developing states. Huawei answered US export controls in silicon, wiring 8,192 Ascend chips into one machine. Anthropic lined up its IPO banks. The build-out keeps compounding; the safety grades still top out at C-plus.

• Google ships the rebuilt Gemini 3.5 Pro. DeepMind released its flagship Jul 17 after scrapping the earlier base model over failures in recursive tool-calling and SVG generation, then rebuilding from scratch. The reported specs — a 2M-token context window, a Deep Think reasoning layer on the $250-a-month Ultra tier, and gains in coding and long-horizon work — still trace to third-party leaks rather than an official model card, which Google had not posted at launch. It arrives days behind OpenAI’s GPT-5.6 family and Murati’s open-weight Inkling.
• Xi opens WAIC and stands up a rival governance bloc. In his first World AI Conference keynote since the event launched in 2018, Xi Jinping called AI “a symphony of international cooperation, not a solo performance by a single country” and pledged 5,000 training slots for developing nations. Twenty-nine countries — among them Russia, Brazil, Indonesia and Pakistan — signed the founding charter of the World AI Cooperation Organization, a China-headquartered intergovernmental body Beijing casts as a counterweight to Western-led standards.
• Huawei answers export controls in silicon. At WAIC, Huawei unveiled the Atlas 950 SuperPoD, lashing 8,192 Ascend 950DT processors into a single logical machine over its UnifiedBus 2.0 memory fabric. Huawei self-reports 8 exaflops of FP8 compute and 6.7 times the throughput of Nvidia’s NVL144, with a SuperCluster tier it says scales past 500,000 NPUs — none of it independently benchmarked. The pitch is frontier-scale compute built on chips China can fabricate without US sign-off.
• Anthropic lines up its IPO banks. Goldman Sachs, Morgan Stanley and JPMorgan are scheduling investor meetings ahead of a possible public listing later this year, per reporting Jul 15 — the firmest signal yet that the lab means to test public markets. It closed a $65B Series H at a $965B valuation in May, overtaking OpenAI as the most valuable AI startup, and now claims a $47B revenue run rate, up from $30B earlier in the year. No filing date is set.
• The bill compounds; the guardrails don’t. TSMC turned the build-out into receipts last week — $40.2B in Q2 revenue, a 67.7% gross margin, and a 2026 growth outlook raised past 40% — as the five largest US clouds guide to $710B-plus in 2026 capex, now underwritten by roughly 9.8 GW of contracted nuclear power. Against that spend, the Future of Life Institute’s Summer 2026 Safety Index still tops out at C-plus, with several labs graded near failing.