Huawei officially introduced the semiconductor 'Tao (τ) Law' at the 2026 IEEE International Symposium on Circuits and Systems, proposing signal propagation delay constant τ as a unified optimization goal for chip evolution, replacing traditional transistor geometric scaling. The core technology, 'LogicFolding', achieves a 55% increase in transistor density and 41% improvement in energy efficiency at the same process node. Huawei disclosed it has already designed and mass-produced 381 chips based on this law, with Kirin 2026 set to debut LogicFolding. The company projects high-end chips will reach an equivalent 1.4nm process level by 2031. The Tao Law also includes system-level optimizations such as unified bus architecture, Hi-ONE optical interconnect (8Tb/s bandwidth), and 3D stacking, forming a four-layer technical framework from devices to data centers. Following the announcement, A-share semiconductor stocks surged across the board, with SMIC hitting its daily limit and reaching a record high.

Anthropic is testing a new dual-mode memory system for Claude, including Classic Memory and Memory Files, the latter using a wiki-like structured format for infinite scalability and on-demand retrieval. It also launched 'Dreams', a background memory integration mechanism inspired by human REM sleep, which automatically merges duplicates, updates outdated entries, and resolves contradictions during conversation gaps. This feature has been implemented in Claude Code as Auto Dream, reducing error rates by 97% and accelerating document validation by 30%. These upgrades support the next-generation persistent AI agent platform, Conway, which will operate in an independent environment capable of search, dialogue, system operations, event listening, task triggering, and code execution, enabling 24/7 uninterrupted service. Through 'Memory Files + Dreams + Conway', Claude establishes a complete闭环 from storage to maintenance to action.

From May 18 to 24, 2026, DeepSeek V4-Flash topped OpenRouter's global LLM API usage leaderboard with 3.43 trillion weekly token calls, a 66% increase week-on-week, surpassing OpenAI, Gemini, and Claude. Overseas users account for 94% of its traffic, marking broad global developer adoption of Chinese large models. Total global LLM token calls reached 28.9 trillion last week, with China leading at 9.223 trillion for the fourth consecutive week, followed by the US at 4.93 trillion. Three DeepSeek models made the top ten, and the company ranked first in total API usage at 5.74 trillion tokens. Its success stems from extreme engineering optimization and ultra-low pricing—V4-Flash input costs are just 1/180th of GPT-5.5. The company recently announced permanent price reductions across all APIs and plans to raise 70 billion RMB to advance AGI research.

Microsoft has revoked Claude Code access for many engineers, shifting focus to its own GitHub Copilot CLI. Internal feedback indicates that enthusiastic use of Claude Code led to token consumption far exceeding projections, with 'compute costs surpassing employee salaries'. Analysts suggest multiple motivations: controlling expenses at fiscal year-end, avoiding lock-in to a competitor’s ecosystem, and mitigating risk as Claude Code holds 54% of the AI coding market and outperforms Copilot on benchmarks like SWE-bench—threatening Microsoft’s grip on developer entry points. This incident reveals the structural cost paradox of 'copilot mode': fixed employee salaries versus unlimited token bills, making ROI for AI-assisted coding difficult to quantify.

Elon Musk announced on social media that xAI will open-source a 500-billion-parameter model by the end of the year, calling it 'should still be quite useful'. Meanwhile, xAI's latest Grok V9-Medium base model (1.5 trillion parameters) has completed pretraining and entered supervised fine-tuning, with reinforcement learning training set to begin in days and public release expected in 2–3 weeks. Musk highlighted significant improvements over the previous v8-small (500B parameters), particularly in programming tasks, enhanced by supplementary training on Cursor programming tool data. V9-Medium has been optimized for NVIDIA Blackwell architecture GPUs, signaling xAI’s shift of Grok from chatbot to developer-focused AI tools.

MiniMax, in collaboration with Tsinghua University and the OpenBMB community, released BitCPM-CANN—the first trinary (1.58-bit) large model series fully trained end-to-end on Huawei’s domestic Ascend computing platform. The series includes models ranging from 0.5B to 8B parameters. Each parameter is compressed into one of three values: -1, 0, or +1, reducing inference memory requirements to 1/6 of BF16 while preserving 90%–97.2% of model capability. The 8B model can run on mainstream flagship smartphones, and combined with MoE architecture, enables deployment of 50B–100B parameter models on devices with only 8GB memory. Training efficiency reaches 95% of conventional baselines. Previously, all publicly known trinary model training relied on NVIDIA GPUs; BitCPM-CANN demonstrates that domestic hardware can achieve ultra-low-bit training.

Enterprise AI agent platform Dust announced a $40 million Series B round led by Abstract and Sequoia, bringing total funding to over $60 million. The platform aims to overcome the limitations of AI's 'single-user mode' by enabling AI agents and employees to collaborate within shared context, tools, and goals, facilitating organizational-scale capability reuse. Dust integrates over 100 data sources and features built-in memory, feedback mechanisms, and enterprise-grade governance. Adopted by more than 3,000 organizations, it reports over 70% weekly active usage, a 2025 net revenue retention rate of 240%, and zero churn. Customers include Vanta, Clay, Persona, and Doctolib, using Dust to automate sales, operations, and customer success workflows.

Chinese humanoid robot startup Unitree Technology is accelerating its IPO on the Shanghai Stock Exchange's STAR Market, targeting a valuation of approximately 4.2 billion RMB ($620 million). The exchange's listing committee formally accepted its application on March 20, 2026, and scheduled the review for June 1—just 73 days later, indicating a significantly expedited process. Financial reports show the company generated 11.7 billion RMB in revenue and 105 million RMB in profit during the first nine months of 2025, with humanoid robots contributing 51.5% of revenue. In 2024, annual revenue was 392 million RMB with 94.5 million RMB in profit, confirming profitability. Unitree's humanoid robots have appeared on China Central Television's Spring Festival Gala in both 2025 and 2026.