Today was the single most geopolitically charged day in AI governance this year. WAIC 2026 opened in Shanghai with an unprecedented head-of-state keynote, a new international AI body formally constituted, and China's largest domestic compute cluster on physical display, all while Moonshot AI quietly released what is now the world's largest open-weight model. Against that backdrop, Google missed its third self-imposed deadline for Gemini 3.5 Pro.
Xi Jinping Takes the WAIC Stage for the First Time
Chinese President Xi Jinping delivered the opening keynote at the 2026 World Artificial Intelligence Conference in Shanghai today, the first time a sitting head of state has addressed the annual summit since it launched in 2018. The appearance is not a courtesy visit. Xi's presence signals that Beijing now treats AI governance as a principal arena of geopolitical competition, on a par with trade and security.
The substance of the speech matched its symbolism. Xi called for AI development to be "a symphony of international cooperation rather than a solo performance by a single country," and explicitly opposed "overstretching the national security concept in the field of AI or placing one country's security over that of others." He announced cooperation initiatives with African, Latin American, Asian, and BRICS partners to provide AI capacity-building and prevent what he described as "new historical injustices" in the distribution of technological capability.
For an operator with global exposure: the people writing AI governance norms are no longer primarily in Brussels, Washington, or Geneva. Shanghai is now a third capital in that conversation, and today's speech is the most explicit statement yet of the terms China intends to negotiate on.
Twenty-Nine Nations Sign the WAICO Charter, Choosing Shanghai Over Geneva
One day before the conference opened, on 16 July, representatives of twenty-nine countries signed the founding agreement of the World AI Cooperation Organization (WAICO) in Shanghai. Chinese Foreign Minister Wang Yi signed on behalf of China. Confirmed founding members include Russia, Kazakhstan, Pakistan, Indonesia, Brazil, Cuba, Venezuela, Serbia, and Belarus, alongside a reported ten African and twelve other Asian nations. The organization will be headquartered in Shanghai and structured as an independent intergovernmental body.
The institutional model follows the Shanghai Cooperation Organization playbook: a multilateral body China co-founded and shaped, oriented toward technology governance rather than security, and designed to give developing nations a formal seat in shaping AI rules. The pitch to the Global South is direct. Open-weight models, lower inference costs, and a governance vote are on offer; what the US-EU "trusted partners" framework has provided most of the world so far is exclusion. Twenty-nine signatures in a first round is a meaningful result, not a rounding error.
The practical consequence for businesses operating across emerging markets is that two parallel AI governance frameworks are now in active development. Which framework your counterparts and regulators are aligned with will determine which compliance requirements you face, which vendors you can use, and which model weights you can deploy.
Moonshot AI Opens the Frontier: Kimi K3 Is the World's Largest Open-Weight Model
Released on 16 July, Kimi K3 from Beijing-based Moonshot AI is a 2.8-trillion-parameter mixture-of-experts model built on two proprietary architectural innovations: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, a drop-in replacement for standard residual connections that the company says delivers consistent scaling gains. The result is roughly 2.5 times the scaling efficiency of its predecessor, Kimi K2. Native visual understanding and a one-million-token context window are included from launch.
On the GDPval-AA v2 benchmark, which evaluates real-world tasks across 44 occupations and 9 industries, Kimi K3 scores 1,687, placing third globally behind Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8). API pricing is $3 per million input tokens and $15 per million output tokens; a modified MIT licence covers commercial use. Full model weights are scheduled for release on 27 July. The model is not the cheapest option available, but it is the most capable open-weight model in existence, and its open-weight nature means any operator can deploy it without ongoing vendor dependency.
This is the clearest example yet of Chinese frontier labs converting WAIC into a launch platform with global reach. Kimi K3 lands into a market where Claude Fable 5 Max and GPT-5.6 Sol are the reference points for capability; that a fully open-weight model now sits between them on at least one practical benchmark is a meaningful shift in the build-versus-buy calculus for any team running inference at scale.
Gemini 3.5 Pro Misses Its Third Target Date as Google Eyes a Stopgap
July 17 was the third publicly reported target date for Gemini 3.5 Pro, following a June commitment and an earlier July 12 target, both of which slipped. Reporting from 9to5Google on 16 July confirmed the model has encountered another delay, attributed to unresolved hallucinations and underperformance on coding tasks despite a mid-training data refresh. Google is reportedly testing a stopgap release, provisionally called Gemini 3.6 Flash, to hold its position in the developer market while the Pro rebuild continues.
The competitive context is unforgiving. GPT-5.6 went public eight days ago. Grok 4.5 launched sixteen days ago. Kimi K3 is available today. Google still has Gemini 3.5 Flash as a solid production option, and the company has confirmed it is testing an upgraded Flash variant with enterprise partners. But Gemini 3.5 Pro was positioned as Google's answer to frontier-class reasoning, and each missed date narrows the window in which it can define that positioning.
For a developer or CTO evaluating model selection: plan around what is available today. A model that has missed three internal deadlines is not a production dependency you can build a roadmap around.
The through-line across today's developments is that AI leadership is contested simultaneously on every dimension: model capability, compute hardware, open-weight access, and governance frameworks. An operator who tracks only benchmarks is watching one dimension of a four-dimensional race. The governance question, in particular, is accelerating faster than most enterprise AI strategies have anticipated.