The article is a mix of several events.

English version above, Chinese version below.

Cloud Revenue Hits New Highs as Chip and Infrastructure Spending Surges

Alibaba Cloud posted its best quarterly results in five and a half years on August 20. External commercial revenue grew 45% year-over-year — the fastest pace in 22 quarters. AI cloud and compute services revenue reached 48.44 billion yuan, with adjusted EBITA up 133% to 5.63 billion yuan, pushing profit margins to 12%. AI-related product revenue grew triple digits for the 12th consecutive quarter, hitting 12.38 billion yuan in Q1 FY2027 (April–June 2026).

On the same earnings call, Alibaba CEO Wu Yongming disclosed that Pingtouge's second-generation chip is expected to tape out in the second half of 2026. The new chip, he said, will deliver computing power and interconnect bandwidth sufficient to "fully replace" large-scale model training. Cloud instances based on the Zhenwu M890 chip are already online and scaling commercially. As of early August, Pingtouge had shipped 560,000 Zhenwu chips to over 650 customers across 20-plus industries.

Separately, on August 21, Bloomberg reported that Broadcom is in talks with lenders to raise over $60 billion in debt financing for AI chip projects supporting Anthropic and others. The deal could include roughly $30 billion in junior debt, with a senior secured tranche potentially reaching $60–70 billion. Total financing could climb to $100 billion if fully executed. Blackstone and Apollo Global Management are among the potential participants. Broadcom already designs custom AI chips for Alphabet and Meta, and has supply agreements with Anthropic and OpenAI.

On the demand side, Meta has become one of Microsoft's largest AI customers, spending hundreds of millions of dollars annually on Azure AI services and consuming trillions of tokens per week, Bloomberg reported on August 20. Meta uses multiple model providers — including OpenAI, Anthropic, DeepSeek, Mistral, and xAI — through Microsoft's Foundry platform, which now hosts 11,604 models from various vendors and serves 100,000 customers. OpenAI alone accounted for roughly 70% of Microsoft's total AI revenue in its most recent fiscal year.

OpenAI Gains Ground on Anthropic in Enterprise; Perplexity Launches Multi-Model Agent API

Ramp data published on August 20 shows Anthropic held nearly 44% market share among paying U.S. business users as of July, versus OpenAI's nearly 40%. But OpenAI is growing faster in Q3 so far — a reversal driven partly by the strong developer reception to GPT-5.6 Sol. Anthropic's Fable 5, meanwhile, saw weaker adoption amid concerns over its price and a 30-day data retention requirement imposed by regulators. The overall enterprise AI market continues expanding: 56% of companies in Ramp's 70,000+ customer base now pay for at least one AI tool, up from 50% in March.

On the model access front, Perplexity officially launched its Agent API on August 21, offering a single endpoint to 41 frontier models from 9 providers, with built-in tools for web search, financial search, and sandboxed code execution. The API is designed for production-grade multi-model agent workflows.

On the open-source side, Ant Group's BaiLing team released Ling-3.0-tiny-base (7.9B total / 1.3B active parameters) and Ling-3.0-flash-base (124B total / 5.1B active parameters) on August 21, alongside six training checkpoints — pre-training, mid-training, and WSM-merged states for each model. The team introduced WSM (Warmup-Stable and Merge), a technique that replaces traditional learning rate decay with checkpoint-weighted merging, making the base models more suitable for continued pre-training and dynamic data expansion. Ling-3.0-flash-base reportedly outperforms base models with 2–3× the total parameter count on code, complex reasoning, and long-context benchmarks.

AI Agents Move Into Operating Systems and Enterprise Data

OpenAI announced on August 20 that its macOS ChatGPT app can now control Apple Messages — reading, composing, and sending texts, searching message history, and generating conversation summaries. The feature requires explicit user consent and full disk access permissions. Bloomberg's Mark Gurman noted this encroaches on territory Apple had reserved for its own Siri AI.

On August 21, Microsoft marked five years of Windows 365 by unveiling Microsoft Scout, a new AI agent that accepts instructions through Teams chat and executes them directly on the user's cloud PC — finding files, editing documents, drafting emails — even when the physical device is closed. Windows 365 for Agents provisions dedicated cloud PCs with isolated identities for agent workloads. Microsoft also introduced developer-optimized cloud PCs preloaded with GitHub, VS Code, and Node.js, running Foundry Local for token-free local AI coding with Phi-4-mini.

Tencent Cloud launched TDSQL Nexa, an AI data platform, at the DTCC database conference on August 20. It unifies transactional processing, vector search, AI computation, and large-scale analytics under a single entry point, and provides Agents with business semantic understanding and row/column-level access control. In internal testing, Nexa delivered 10× the compute performance of open-source alternatives at 60% lower cost. Separately, Tencent's Agent Memory open-source project has earned over 20,000 GitHub stars in 80 days, and its DatabaseClaw agent can localize some production faults within three minutes.

Also on August 20, Adobe expanded Firefly with three AI audio services — music generation, speech synthesis, and sound effects — now available to all users. Meta brought its Pocket app (formerly Gizmo) to the U.S. market, allowing users to vibe-code and share interactive mini-games through a scrollable feed. Pocket is the latest in a series of standalone AI-native apps Meta has shipped this year, including Instants, Forum, Seller, and the recently launched Meta AI Mac app.

AI Safety, Content Labeling, and Cyber Threats Converge

Apple informed music suppliers on August 20 that AI Transparency Tags — previously voluntary since their March 2026 introduction — will become mandatory by end of year for any music "substantially" generated by AI. Labels and distributors lose the option to omit the tag. Apple VP Oliver Schusser said in April the company has "technology that can accurately see what music people are providing us," including which AI model was used.

On the same day, Apple filed a new lawsuit against OpenAI alleging trade secret theft, claiming a former employee exploited an authentication vulnerability to download dozens of confidential hardware files, and that OpenAI personnel used recruitment interviews to extract CAD drawings and supplier details.

On August 20, CISA, the FBI, and the NSA issued a joint warning that hackers are using AI to generate exploit scripts and scan for vulnerable Siemens S7 programmable logic controllers in U.S. water and wastewater systems. The agencies said attacks span "all" S7 PLC models, with older or poorly secured devices most at risk. Multiple states — including Minnesota, Michigan, Arkansas, Georgia, and New Jersey — have reported intrusions at water facilities. Iranian-linked hackers are suspected in several recent incidents.

In content enforcement, Qidian, China's largest web novel platform, began removing or de-recommending popular AI-heavy novels — including titles with tens of thousands of average subscriptions — as part of a broader push to identify and moderate AI-generated content.

Gary Marcus published a Substack piece on August 20 warning that OpenAI is developing 24/7 surveillance devices that upload user voice data to the cloud for analysis, and is using vague legal language to decline opt-out requests for government surveillance.

Robotics Goes Public; Robotaxi Fleet Goes Global

Unitree Technology debuted on Shanghai's STAR Market on August 19, closing its first day up 460%, with a market cap exceeding 340 billion yuan. The same day, the 2026 World Robot Conference opened in Beijing with over 300 exhibitors, up 36% from last year. The conference featured a dedicated "Procurement Day" for the first time. Among the standout companies: Yunshenchu (DR02 humanoid robot completing substation trials, 2025 revenue of 337 million yuan, three-year CAGR over 150%), LimX Dynamics (Pre-IPO round at 15 billion yuan valuation), Realman Intelligence (full-stack joint modules + teleoperation network, 2025 operating profit), and Galaxea AI (20 billion yuan valuation, B+ round completed in April).

On August 14, Pony.ai expanded its Uber partnership to deploy over 2,000 Robotaxis across five European cities, starting with Zagreb, Croatia. Combined with existing plans in the Middle East and elsewhere, Pony.ai's total overseas deployment target exceeds 4,000 vehicles. Pony.ai's Q2 2026 Robotaxi revenue reached 81.92 million yuan, up 691% year-over-year, with passenger fare revenue surging 849%. Its seventh-generation Robotaxi achieved unit economic breakeven in Guangzhou and Shenzhen, and the 2027 model targets a total vehicle cost — including the L4 autonomous driving kit — below 230,000 yuan.