Multiple media outlets citing an OpenAI announcement report that the company will discontinue its standalone Sora video generation application and developer API, approximately 15 months after its launch in late 2024. OpenAI stated it will consolidate research and operational resources into more core commercial and productivity product lines, but did not disclose migration plans for existing user projects or data. Reports also note that this move casts uncertainty on prior ecosystem partnerships built around Sora, including integrations with major content providers.

Alibaba's DAMO Academy unveiled the server-grade CPU 'XuanTie C950' at the XuanTie RISC-V Ecosystem Conference, operating at 3.2GHz with a SPECint2006 single-core performance breaking 70 for the first time. It natively supports CoVE confidential computing and RVA23.1 extensions. DAMO Academy also introduced two RISC-V native AI engines—Vector and Matrix—claiming native support on CPU for billion-parameter large models (e.g., Qwen3-235B-A22B, DeepSeek V3-671B), achieving over 30% performance improvement in cloud networking and storage tasks compared to mainstream solutions.

Simon Willison disclosed that LiteLLM v1.82.8 on PyPI was compromised with a malicious `litellm_init.pth` file that executes automatically during installation, without requiring an import in code. The stealer collects credentials and configurations from SSH, AWS, Kubernetes, Docker, Git, shell history, and cryptocurrency wallet-related files, posing risks across common development and cloud environments. PyPI has isolated the version; users who installed during the affected period are advised to immediately rotate relevant secrets and audit their systems.

Cloudflare released Dynamic Workers (Dynamic Worker Loader), replacing containers and micro VMs with V8 isolates to execute AI agent-generated code, claiming startup speed improvements of ~100x and memory efficiency gains of 10–100x, enabling 'a new sandbox per request'. Its 'Code Mode' exposes tool capabilities as TypeScript-typed interfaces, reportedly reducing token consumption for tool calls by up to 81%. It also provides mechanisms like `globalOutbound` to intercept outbound requests, enabling credential injection and egress control without exposing raw keys.

Databricks introduced Lakewatch, positioned as an 'open, agentic SIEM', leveraging the Lakehouse architecture to unify security, IT, and business data, addressing high costs and data silos in traditional SIEM systems. The product emphasizes embedding AI agents directly into security operations: supporting natural language queries, automated data ingestion, and detection rule creation, enabling responses at 'machine speed' to machine-scale attacks. Built on open standards such as OCSF and Delta Lake, it emphasizes data ownership and pluggable ecosystems, reducing vendor lock-in risks, targeting use cases requiring long-term full-data retention and multimodal analysis.

Oracle launched Fusion Agentic Applications and expanded AI Agent Studio capabilities for Fusion users, focusing on natively embedding goal-oriented AI agents into ERP/HCM/SCM/CX transaction systems, executing workflows within existing approval, permission, and audit frameworks. Oracle claims 22 agent applications are now available, covering shift scheduling and payroll optimization, procurement cost reduction, cross-selling, and collections. It also offers no-code building, workflow orchestration, context memory, monitoring observability, and ROI dashboards, aiming to drive enterprise adoption from plugin-style assistants to governable production-grade automation.

Ai2 released MolmoWeb, an open-weights visual Web Agent that performs web navigation and understanding directly from browser screenshots, without relying on HTML or accessibility trees, thereby reducing coupling with site structure and browser implementation. Alongside the model, Ai2 released the MolmoWebMix training dataset and pipeline, containing 30,000 human task trajectories and approximately 2.2 million screenshot-question-answer pairs, enabling reproducibility, auditing, and retraining. This release aims to address the unverifiable and non-reproducible nature of closed-source 'computer-use' agents, providing researchers with auditable data and end-to-end training infrastructure.

Reuters reports Broadcom stating that surging AI chip demand has pushed its foundry partner TSMC to near-capacity limits, making it one of the 2026 supply chain bottlenecks. Strain is not limited to chips but extends to laser components and printed circuit boards (PCBs); delivery cycles for PCBs used in optical transceivers have reportedly extended from about six weeks to six months. To secure supply, many customers are signing 3–4 year long-term agreements; Samsung is also pushing for 3–5 year contracts to hedge against capacity uncertainty.

Xinhua News cites OpenRouter data showing global LLM weekly token usage reached 20.4 trillion during March 16–22, a 20.7% increase week-on-week; Chinese models accounted for 7.359 trillion tokens, exceeding US models (3.536 trillion) for three consecutive weeks. According to a press conference by China's National Data Bureau, the country's average daily token usage has exceeded 140 trillion. By the end of 2025, over 100,000 high-quality datasets had been built nationwide, totaling 890PB. The report attributes the growth in call volume to low-cost APIs, open-source ecosystems, and large applications such as WeChat and DingTalk, positioning token count as a core metric for measuring application scale.