Today China's AI companion regulation enters force, closing the era of persistent persona customisation for the world's largest mobile-AI market. Elsewhere, Anthropic pushes into the classroom the day after new details emerged on its pre-IPO hardware strategy, and Thursday shapes up as the densest day of AI launches this year.
China ends the AI companion era — for now
The Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services entered force today, 15 July 2026. Co-issued in April by the Cyberspace Administration of China alongside four partner agencies — the NDRC, MIIT, Ministry of Public Security, and SAMR — the rules require anti-addiction systems, mandatory usage-session notifications, and instant-exit mechanisms on any AI service that adopts a persona or simulates an emotional bond with users.
ByteDance's Doubao (the most used AI chatbot in China), Alibaba's Qwen, and Tencent's Yuanbao all began disabling or removing persistent-persona customisation before the deadline. The compliance posture reveals the structural problem: anti-addiction controls and instant-exit interrupts are architecturally incompatible with how persistent-memory agents are built. Platforms could not satisfy the rules by flipping a configuration switch — they had to remove the feature.
The data situation is unresolved for some users. Doubao users retain access to their agent configurations until 15 October 2026. Qwen issued no comparable grace period, and agent data there is already scheduled for permanent deletion. For any operator building companion or persistent-memory services globally, today's enforcement date sets the first national template: anti-addiction constraints and user-exit rights must be handled at the architectural level, not the policy layer.
Claude for Teachers — Anthropic bids for the institutional trust layer
On 14 July, Anthropic launched Claude for Teachers, giving verified US K-12 educators free premium access through 30 June 2027. The package includes Claude Code, Claude Cowork, and a Learning Commons connector mapped to academic standards in all 50 states. A FERPA-aligned data processing agreement governs the arrangement; content submitted by educators is not used to train Anthropic's models.
Education was one of OpenAI's first institutional beachheads; Google holds the broadest K-12 penetration through Google Classroom. Anthropic is now competing directly with both, at zero cost to the individual teacher. The compliance architecture — a separate data processing agreement and an explicit no-training commitment — mirrors the model Anthropic used for its California government deal and the Bristol Myers Squibb partnership. Institutions that accept these terms at a low or zero price now have, in effect, a vendor of record for the next procurement cycle.
Thursday: WAIC Shanghai, 300-plus debuts, and the Gemini question
The 2026 World Artificial Intelligence Conference and High-Level Meeting on Global AI Governance opens in Shanghai on Thursday 17 July, running through 20 July. Over 1,400 participants and 1,000 exhibitors are confirmed; more than 300 AI products are scheduled for their global debut across a floor space exceeding 100,000 square metres for the first time. Xi Jinping is attending in person — the first time at WAIC — a geopolitical signal about how seriously Beijing is treating this moment.
The conference's stated narrative is a deliberate shift from model competition to industrial delivery: Chinese AI is now measuring itself by deployment at scale, not benchmark scores. That framing lands on the same day as Google's widely anticipated Gemini 3.5 Pro release, for which no official specifications have yet been confirmed. No single day this year has carried this density of potential announcements. Operators with imminent model-selection decisions should set their evaluation criteria before the noise arrives.
Anthropic's IPO approach: silicon, classrooms, and October
Anthropic confidentially submitted a draft S-1 to the SEC on 1 June 2026, targeting a Nasdaq listing in October 2026. At its May 2026 Series H-1 valuation of $965 billion, the Goldman Sachs-, JPMorgan-, and Morgan Stanley-led offering is expected to raise more than $60 billion — the largest AI company debut by some distance, and the first among the frontier labs, with OpenAI having deferred its own listing to 2027.
This week's moves fit the pre-IPO logic. Claude for Teachers builds the institutional-trust narrative that enterprise buyers and public-market investors will evaluate. Separately, Anthropic's talks with Samsung over a 2-nanometer custom inference chip signal that it is following the OpenAI-Broadcom playbook: own the compute layer before it becomes a commodity. Anthropic recently hired Clive Chan, an early member of OpenAI's chip team, to lead the effort.
For enterprise buyers, the IPO window is now open. Vendors about to trade publicly typically offer their most favourable contract terms to anchor customers during the roadshow period. Intensified commercial outreach and expanded compliance commitments from Anthropic's enterprise team are likely over the next three months.
The thread across all four items today is the same: AI is moving from capability demonstrations to institutional control — who holds the regulatory licence to operate, who owns the hardware, who signed the compliance agreement first. In education, government, and China's consumer market alike, those contests are being settled now, quietly, and largely out of public view.