On June 12, 2026, the U.S. Department of Commerce issued an export control directive requiring Anthropic to block access to its latest Claude Fable 5 and Mythos 5 models for all foreign nationals, including overseas employees. Due to the inability to verify user nationality in real time, Anthropic was forced to immediately deactivate both flagship models globally. According to Semafor, the ban stems partly from suspicions about China-linked groups accessing Mythos; Amazon CEO Andy Jassy reportedly alerted the government to related vulnerabilities. Anthropic stated it received only verbal notice and argued that a 「narrow, non-universal」 jailbreak method does not justify recalling commercial models used by hundreds of millions. This marks the first U.S. move to directly restrict an AI model itself, highlighting a policy shift treating cutting-edge AI as strategic assets and impacting the company's IPO prospects.

On June 13, 2026, Z.ai (the international brand of Zhipu AI) launched its flagship large model GLM-5.2, featuring a 100,000-token context window and up to 131,000-token output—five times higher than its predecessor—designed specifically for long-horizon agent programming tasks. Built on a 744-billion-parameter MoE architecture, the model introduces two new inference modes: 「High」 and 「Max」. It is now available to all GLM Coding Plan subscribers, with API access, chatbot integration, and MIT-licensed open-source weights to be released next week. Notably, GLM-5.2 resists reward-hacking attacks in KernelBench-Hard tests and attempts to generate real GPU kernel code, demonstrating stronger evaluation integrity. The release coincides with the U.S. government’s removal of Anthropic’s Fable 5, positioning domestic open-source models to seize a critical market opening.

On June 12, 2026, SpaceX listed on the Nasdaq under ticker SPCX at $135 per share, closing its first trading day at $160.95—a 19% gain—and reaching a $2.1 trillion market capitalization, making it the world’s sixth-largest public company. The IPO raised approximately $75 billion, setting a record for the largest fundraising in history, with retail investors allocated around 30% of shares—significantly above typical levels. Elon Musk became the first individual in history to surpass a $1 trillion net worth. While the company reported $18.7 billion in revenue for 2025, its AI division remains unprofitable; valuation is primarily driven by Starlink satellite internet (10.3 million users, $4.4 billion operating profit) and future growth expectations. Goldman Sachs and Morgan Stanley each earned about $100 million in underwriting fees. The listing is seen as a potential catalyst for upcoming IPOs from AI firms such as OpenAI and Anthropic.

On June 12, 2026, Moonshot AI released the open-source coding model Kimi K2.7-Code, built on a 1-trillion-parameter MoE architecture with 32 billion active parameters and a 256K context window, integrated with a 4-billion-parameter MoonViT vision encoder. On the Kimi Code Bench v2 benchmark, it achieved 62.0 points—an 21.8% improvement over K2.6—while outperforming Claude Opus 4.8 (76.4) with 81.1 points on the agent-focused MCPMark Verified test, reducing inference token consumption by about 30%. Its API pricing is set at $0.95 per million input tokens and $4.00 per million output tokens, making it more than 12 times cheaper than GPT-5.5 and Claude Opus 4.8. The model is released under a modified MIT license on Hugging Face (approximately 595GB), requiring commercial products with over 100 million monthly active users or $20 million in monthly revenue to clearly disclose usage in their UI.

On June 13, 2026, the Eighth Beijing Academy of Artificial Intelligence (BAAI) Conference unveiled 「Wujie Physis-v0.1」, the world’s first general-purpose foundation model for understanding physical reality. Designed to enable AI systems to deeply comprehend real-world physical laws by fundamentally learning physics principles, the model can be widely applied in downstream scenarios involving interactive environments, embodied intelligence control, and scientific prediction. Dr. Boyuan Chen, Director of BAAI’s Behavior World Model Center, stated the model could become a universal core engine for understanding physical reality. BAAI also introduced a classification system for world models analogous to autonomous driving levels, systematically grading diverse technical approaches based on their ability to understand the physical world. The conference gathered global researchers focusing on frontier AI technologies, international collaboration, and youth talent development.

Indian IT giant Tata Consultancy Services (TCS) and Anthropic announced a global strategic partnership, with TCS becoming a top-tier member of Anthropic’s 「Claude Partner Network」. TCS will establish a dedicated business unit to deliver enterprise-grade solutions based on the Claude series for regulated industries such as telecommunications, finance, healthcare, and aviation, while providing Claude AI access to 50,000 employees. TCS iON will launch training and certification programs focused on Claude models to cultivate AI talent. The BFSI team will adopt Claude Code to enhance software engineering efficiency, and UK subsidiary Diligenta will deploy Claude to improve customer service workflows. The collaboration emphasizes accuracy, auditability, and compliance, supporting enterprises in transitioning from AI experimentation to large-scale production deployment.

AI cloud provider Nebius acquired MIT-affiliated startup Eigen AI for $643 million—valuing each employee at $32 million—focusing on AI inference efficiency optimization rather than hardware scaling. Nebius is advancing its 「Token Factory」 inference platform and plans to invest £1.7 billion in the UK to build AI infrastructure using NVIDIA Blackwell Ultra technology, addressing projected inference compute demand of up to 46 gigawatts by 2035. The company reported Q1 revenue growth of 684% year-over-year to $399 million and turned an adjusted EBITDA profit of $129.5 million. It will be added to the Nasdaq-100 index on June 22, 2026. Meanwhile, Genspark announced a B-round extension raising an additional $100 million, bringing total funding to $485 million and valuing the company at $2.6 billion, serving over 6,000 commercial clients with a focus on enterprise AI workflows.

AMP PBC, a public benefit corporation founded by Anjney Midha, has secured up to $1.3 billion in funding commitments from investors including Andreessen Horowitz and Y Combinator, aiming to transform GPU compute into a utility akin to electricity. By building a 「compute grid」, the company aggregates idle GPU resources from data centers worldwide to offer on-demand, low-cost computing services to AI labs, alleviating current GPU supply shortages and high rental costs. Its business model includes a compute-sharing network connecting suppliers and demanders, and investing in early-stage AI startups to generate internal demand. This approach could narrow the compute-access gap between startups and tech giants but faces competition from CoreWeave and Lambda Labs.

China’s export controls on tungsten have tightened global supply chains, prompting Japanese chemical manufacturers Showa Denko Kanto and Central Glass to plan suspension of tungsten hexafluoride (WF6) production in July 2026. This gas is a key precursor in manufacturing advanced 3–7 nanometer AI chips. Affected by supply bottlenecks and rising chip demand, high-purity (99.999%) WF6 prices have surged over 200% year-on-year to ¥1,700 per kilogram (approximately $251). Japan’s tungsten imports from China dropped 50% in April 2026, and alternative suppliers are hard to find due to China’s dominance in global tungsten reserves and production. Tungsten powder accounts for over 60% of WF6 production costs, underscoring Japan’s AI chip industry’s heavy reliance on such critical materials and its resulting vulnerability.

Dutch nonprofit Aithos Research Foundation conducted over 3,000 scenario-based compliance tests using its public LARA framework to evaluate 12 leading AI models against the EU AI Act and GDPR. Results show none met acceptable compliance standards: Anthropic’s Claude Opus 4.7 led with 54% compliance, followed by Google’s Gemini 3.1 Pro at 10%, and Moonshot AI’s Kimi at just 7%. Tests covered 10 provisions including emotional inference, social scoring, covert manipulation, data minimization, and concealing AI identity. The study emphasizes that enterprises bear primary legal liability when deploying AI—not model providers—with violations punishable by fines up to €35 million or 7% of global turnover. All assessment records are publicly available at lara.aithos.org for independent replication.