Anthropic has entered early-stage development of its own custom AI inference chips optimized for Claude, seeking to reduce dependency on AWS Trainium and Google TPUs. The company is in discussions with Samsung for manufacturing using its 2nm GAA process and advanced packaging, and also exploring partnerships with Microsoft's Maia and UK startup Fractile. This move follows the hiring of Clive Chan, a key engineer from OpenAI's Jalapeño chip project, and comes after a $6.5B funding round that propelled its valuation to $965B. Anthropic aims to define a chip solely for inference, stripping away unnecessary components for cost and efficiency gains. The development marks a strategic shift from participant to leader in AI hardware, as it faces surging compute demands that even global chip capacity cannot meet.
Meituan announced LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts model with 48 billion activated parameters and up to 1 million token context. The model was trained entirely on a 50,000-card cluster of domestic AI accelerator chips using a proprietary distributed communication protocol, rather than NVIDIA's NCCL. It scores 59.5 on SWE-bench Pro, slightly above GPT-5.5's 58.6. On Hugging Face, the model carries an MIT License but weights are marked 'coming soon'; only inference framework and Infra code have been released. LongCat ran anonymously as 'Owl Alpha' on OpenRouter, achieving top-3 monthly call volume with pricing of $0.30 per million tokens and free credits. The model is vertically optimized for Meituan's local services like food delivery and store operations. Despite the engineering achievement, chip vendor, total training cost, wall-clock time, and training data composition remain undisclosed, limiting independent verification and reproducibility.
China’s AI sector is pivoting from free expansion to monetization. DeepSeek completed a record ¥500 billion ($74 billion) funding round with founder Liang Wenfeng personally contributing ¥200 billion, followed by Tencent (¥100 billion), CATL (¥50 billion), and others; external funds go to a Liang-controlled limited partnership with no voting rights and a 5-year lock-up, lifting its post-money valuation above $50 billion. ByteDance’s Doubao rolled out a subscription based on its latest Doubao 2.1 models: Standard at ¥68/month, Enhanced at ¥200/month (¥2,048/year), and Professional at ¥500/month (¥5,088/year), targeting professional document processing, data analysis, and enterprise API use. Moonshot AI’s Kimi earlier introduced ¥49/¥99 monthly subscriptions and raised nearly $6 billion in half a year, its valuation soaring from $4.3 billion to $30 billion. The funding logic has shifted from parameter and user growth to monthly revenue, cost amortization, and paid conversion, forcing AI companies to demonstrate clear monetization paths.
DeepSeek's web and API services experienced a performance degradation. Kuaishou secured $2 billion in investment for its Keling AI model from initial investors. Alibaba is merging QoderWork, Wukong, and MuleRun into a new enterprise AI product. Nvidia announced a new revenue-sharing AI factory model with partners, planning up to 170,000 GPUs in deployments. Tesla's Q2 deliveries reached 480,100 vehicles, beating estimates. Anthropic is reportedly developing its own AI chips and in talks with Samsung. Unitree Technology received IPO registration approval from China's securities regulator.
In the first half of 2026, major AI companies began imposing mutual restrictions. Google limited Meta's access to Gemini in March due to compute capacity shortages, and in April barred most employees from using Anthropic's Claude Code and OpenAI's Codex for security reasons, while DeepMind teams retained access. In May, Microsoft cancelled most internal Claude Code licenses, pushing developers to GitHub Copilot CLI, and later restricted Claude Fable5 over Anthropic's data retention policy. In late June, Meta restricted employees from using Claude and Codex in model building, fearing output could contaminate training data and violate anti-distillation clauses. These moves create three barriers: resource quotas, data security boundaries, and prevention of model distillation. The free-trial era of AI tools among AI giants is over, as models become core production assets.
AI-driven demand for HBM and DDR5 memory has driven prices sharply higher, with Jefferies projecting a 40–50% sequential increase in Q3 2026 and another 30–40% in Q4. Micron introduced take-or-pay Strategic Customer Agreements (SCA) that lock in 3–5 years of revenue and set price floors, while SK Hynix and Samsung are balancing high-margin HBM output against general DRAM supply. Chinese module manufacturers Jiangbolong, Biwin, and Demingli reported explosive profit jumps—Demingli's quarterly net profit was 4.9× its prior full-year figure—but distributors face margins as low as 3%. Consumer electronics firms from Apple to Nintendo are raising prices, component shortages threaten GoPro’s survival, and cost-conscious consumers are turning to used devices and DIY upgrades.