Shanghai is hosting the World AI Conference this week, and Xi Jinping is delivering the opening keynote – the first time a Chinese president has done so. Meanwhile, China’s Moonshot AI just dropped Kimi K3, the largest open-weight model yet, and TSMC’s Q2 numbers came in so far above expectations that the company added another $100 billion to its US buildout. The supply and demand sides of the AI boom both had a day.

The Geopolitics

Xi Jinping opens the World AI Conference in Shanghai, pitches China as the Global South’s AI partner

The 2026 World AI Conference and High-Level Meeting on Global AI Governance opened today at the Grand Halls in Shanghai, running through July 20. President Xi Jinping attended in person and delivered the opening keynote – the first time a Chinese head of state has done so at the event. His pitch: AI development and governance should be a “global effort” led by international bodies, with explicit space for developing countries and the Global South. He framed US export curbs and chip restrictions as a destabilising force, and offered China’s models, compute, and standards as an alternative stack.

Reuters and CNBC both led with the geopolitical angle. The subtext: China wants a seat at the governance table as an equal to the US, and is using the conference to position the WAIC as a counterweight to the UK AI Safety Summit / US-led Bletchley process. Whether that framing lands in Brussels and Washington is the open question.

The Models

Moonshot AI ships Kimi K3, the largest open-weight model yet

Moonshot AI released Kimi K3 on July 16 (July 17 in China). The model is 2.8 trillion total parameters using a Mixture-of-Experts design that activates 16 of 896 experts per token, ships with a 1,048,576-token (1M) context window, and is multimodal (text and images). It is now live on the Moonshot API, the Kimi app, Kimi Work, Kimi Code, and via OpenRouter. The full open-weights release is promised by July 27.

Axios led the Western coverage: “China’s open-weight Kimi model stuns AI world with frontier-level results.” The performance numbers put it in the same tier as the closed frontier models from OpenAI and Anthropic, which has been the open-weight community’s holy grail for two years. If the weight drop on July 27 lands cleanly, expect a wave of self-hosted enterprise deployments – this is the first model that gives enterprises a real reason to stop renting from a frontier API.

Google expected to launch Gemini 3.5 Pro today

The planned release date for Gemini 3.5 Pro is today, July 17. The expected spec: 2M-token context window, native “Deep Think” reasoning mode, and tighter integration with Vertex AI for enterprise customers. Business Insider reported last month that Google delayed the launch from June to July to “tweak the model,” and several trackers now show a 2M context preview available to enterprise customers through Vertex. Consumer availability through the Gemini app is not confirmed at the time of writing.

If it lands as described, Gemini 3.5 Pro is the first major frontier model with a 2M context window in production – Anthropic’s Sonnet 5 ships 1M, and OpenAI’s GPT-5.6 is at 1.5M. The long-context race is real and Google is trying to leapfrog.

The Infrastructure

TSMC Q2: $40.2B revenue, 77% profit jump, +$100B Arizona

TSMC reported its Q2 2026 numbers this week. The headlines: $40.2B in revenue, 67.7% gross margin, and a 77% year-over-year jump in net profit – all records. The company announced an additional $100 billion in US investment, primarily in Arizona, on top of the existing buildout, and raised full-year capex guidance to $60-64B. CEO C.C. Wei said outsized AI demand forced the increase.

The capex number matters more than the profit print. TSMC is the bottleneck for almost every AI chip made today – NVIDIA, AMD, the hyperscalers’ in-house silicon. A $60B+ capex year from TSMC is what tells you the AI infrastructure build is still ramping, not plateauing. For anyone building on top of these chips, this is the supply signal that matters.

NVIDIA and Japan: Vera Rubin AI factory, 27,500 Rubin GPUs

Jensen Huang was in Tokyo on July 16 for the “Japan AI Ecosystem” event. The big announcement: NVIDIA is working with Japanese firm Noetra Corp to launch a national physical-AI infrastructure built on 13,750 Vera CPUs and 27,500 Rubin GPUs. The deal is positioned as the world’s first national-scale AI factory, with the goal of giving Japanese industry a sovereign compute base for robotics and physical AI workloads.

The Japan push is a template for how NVIDIA is approaching non-US markets: partner with a national champion, sell at scale, lock in CUDA as the platform. Expect similar announcements out of India, Saudi Arabia, and the UAE in the next 12 months.

Meta’s Prometheus supercluster comes online in 2026

Mark Zuckerberg announced Prometheus on July 16 – Meta’s first AI data center supercluster, coming online in 2026. It’s the first of several; Hyperion and other superclusters are planned to follow. Meta is now matching the hyperscaler capex pace (Google, Microsoft, Amazon) on dedicated AI infrastructure, signalling that the in-house Llama training runs are not just for branding anymore.

The Platform Stuff

Google renames NotebookLM to Gemini Notebook

Effective July 16, Google rebranded NotebookLM to Gemini Notebook. The product is unchanged – your notebooks, sources, and audio overviews are all still there. The new name signals that NotebookLM is now part of the Gemini family and will integrate more tightly with other Gemini surfaces over time. The rebrand is also a navigation puzzle: notebooklm.google and gemini.google now point to different products, and Workspace admins are being asked to update their internal links.

Quick Hits

  • Hugging Face security incident (Jul 16) – End-to-end AI-agent-led intrusion disclosed. The same kind of autonomous-agent attack the July 16 daily covered, now with the full postmortem. Link
  • Mistral Robostral Navigate (Jul 14) – 8B-param robotics/navigation model, Mistral’s first physical-AI release. Marks Europe’s official entry into the embodied-AI race. Link
  • Gemma 4 updates (Jul 16) – Flash Attention 4, tool calling improvements, vision support expanded. Still versioning confusion between “Gemma 4” and “Gemma 4.1”. Link
  • Anthropic: Government of Alberta uses Claude to find and fix cybersecurity vulnerabilities. Link
  • World AI Conference Shanghai runs July 17-20. Expect more announcements through the weekend. Conference site

Compiled July 17, 2026. Sources: Reuters, CNBC, Axios, Xinhua, NVIDIA, Google, Moonshot AI, Microsoft, Mistral, Hugging Face, Anthropic, buildfastwithai.com, jls42.org.