Nvidia unveiled the Vera Rubin computing platform (successor to Blackwell) at GTC 2026 on March 16, positioning it as a rack-scale system for MoE and agent-based inference. External reports indicate it contains 336 billion transistors, uses HBM4 memory, and delivers approximately 3.3–5 times higher performance on MoE workloads compared to the previous generation. Jensen Huang stated that inference has reached a 'tipping point,' with the company seeing a potential $1 trillion in orders; Nvidia also launched NemoClaw, an open-source enterprise agent platform focused on privacy and secure deployment, expected to be available on major cloud platforms within 6–12 months.

Nvidia released an open-source reference architecture called the 'Physical AI Data Factory Blueprint' to standardize training data pipelines for robotics, vision AI agents, and autonomous vehicles: automating data generation, augmentation, and evaluation by combining synthetic data, reinforcement learning, and model evaluation, leveraging the Cosmos foundation model and coding agents to expand limited real-world data into diverse datasets covering rare edge cases. Cloud providers including Microsoft Azure and Nebius have already integrated the blueprint; Nvidia also open-sourced the orchestration framework OSMO for automated resource management of agents such as Claude Code and OpenAI Codex. The blueprint is scheduled to launch on GitHub in April.

Meta has signed a five-year AI compute agreement with Nebius, totaling up to $27 billion: Nebius will deliver $12 billion in dedicated computing capacity starting in early 2027, based on NVIDIA's Vera Rubin platform; Meta may additionally purchase up to $15 billion in extra capacity and holds first rights on unsold clusters. Following the announcement, Nebius shares rose approximately 13%–16%. This contract highlights how hyperscalers are using long-term contracts to secure next-generation GPU cluster supply, bringing the delivery capabilities and capital demands of 'new cloud/AI-native cloud' providers into focus.

Encyclopaedia Britannica and Merriam-Webster filed a lawsuit against OpenAI in New York federal court on March 16, alleging unauthorized copying and use of their encyclopedic and dictionary content to train GPT/ChatGPT, involving approximately 100,000 articles. The complaint states that ChatGPT can output passages nearly verbatim from original texts and diverts subscription and advertising traffic through generative responses, creating direct competition with content providers. Plaintiffs also allege improper trademark usage and seek injunctive relief and undisclosed damages. OpenAI responded that its training relies on public data and asserts 'fair use'.

Alibaba announced the creation of a new AI business group, 'Alibaba Token Hub (ATH)', integrating previously fragmented teams including Tongyi Lab, MaaS business lines, and Qwen, Wukong, and others, covering the full chain from foundational model development to ToB/ToC applications, directly led by CEO Eddie Wu. Tongyi Lab remains under Alibaba Cloud CTO Jingren Zhou, with a new foundational model support team established. Reports suggest this move comes amid key personnel departures and stock pullbacks, aiming to reduce redundant investment and improve delivery efficiency; Qwen's consumer-facing services have surpassed 100 million monthly active users.

South Korea's Ministry of Science and ICT launched a high-performance GPU rental program on March 16, partnering with cloud providers to offer over 2,000 GPUs to domestic enterprises for AI projects, alleviating global GPU shortages affecting training and inference. The government allocated a 2.08 trillion KRW budget for 2026 to support access to computing resources, having secured 13,000 GPUs in 2025. The program is open to industry, academia, and research institutions, aiming to enhance compute accessibility while promoting local model training, inference deployment, and application services, lowering the barrier for small and medium-sized teams.

Accenture announced the completion of its acquisition of UK-based AI firm Faculty. Following the deal, Faculty co-founder and CEO Marc Warner has been appointed Chief Technology Officer of Accenture and joined the Global Management Committee. More than 400 data scientists and AI engineers from Faculty have been integrated into Accenture, strengthening its talent and engineering capabilities in 'secure applied AI' and scalable delivery. Faculty’s decision intelligence product Frontier will be incorporated into Accenture’s portfolio to help public sector and enterprise clients build scalable AI solutions; both parties have already collaborated with clients such as Novartis in life sciences. Financial terms of the acquisition were not disclosed.

A team from the Institute of Computing Technology, Chinese Academy of Sciences (ICTCAS), proposed the DeepRead framework for Agentic RAG, addressing the 'structural blindness' issue in long documents by extracting chapter hierarchies via OCR and establishing a coordinate system for paragraphs (doc_id/sec_id/para_idx), enabling agents to locate evidence according to native document structure. Through collaboration between Retrieve and ReadSection tools, it enables a 'locate first, then read continuously' retrieval-reading workflow. Experimental results cited show accuracy on ContextBench long-document reasoning rising from 74.5% to 91.5% (+17.0%), reducing fragmented retrieval and redundant calls in complex tasks like financial report and paper analysis.

Singapore-based data center operator DayOne Data Centers is reportedly close to submitting confidential IPO documents to the U.S. SEC, potentially raising around $5 billion, with underwriters including Bank of America, Citigroup, JPMorgan Chase, and Morgan Stanley. Formerly known as GDS International, the company completed a C-round funding exceeding $2 billion in January 2026 for international expansion. DayOne operates data centers in Singapore, Malaysia, Indonesia, Thailand, Hong Kong, Tokyo, and Finland. If the listing proceeds, it could provide a new capital market valuation benchmark for the AI compute and data center sector in this cycle and intensify competition for regional facilities and power resources.

Nature Machine Intelligence published a framework integrating large language models (LLMs) with the Robot Operating System (ROS), aiming to reliably convert natural language outputs into executable physical actions. The framework supports compiling LLM outputs into executable forms such as inline code or behavior trees, acquires new skills via imitation learning, and uses human or environmental feedback for automated optimization and iterative 'reflection' to improve robustness in long-horizon tasks. The paper demonstrates scalability in scenarios such as tabletop object rearrangement, dynamic task optimization, and remote supervised control; all experiments were conducted using open-source pre-trained LLMs, with code publicly released on GitHub.