A federal jury in California unanimously ruled on May 18 that Elon Musk's lawsuit against OpenAI and CEO Sam Altman would not be heard because it exceeded the three-year statute of limitations. The jury found that Musk was aware of OpenAI's shift to for-profit operations as early as 2021 but delayed filing suit until 2024. As a result, the court did not conduct a substantive review of Musk's core allegation that OpenAI abandoned its nonprofit mission. The trial revealed several sensitive details, including Brockman holding nearly $300 billion in equity and Microsoft's stake valued at $1350 billion. The ruling clears a legal hurdle for OpenAI's potential trillion-dollar IPO. Musk's legal team plans to appeal to the Ninth Circuit Court, while other antitrust and trade secret lawsuits remain ongoing.

Cursor AI has launched its most powerful model, Composer 2.5, dedicating 85% of its compute to additional training on the Kimi base model, resulting in significant improvements in long-task handling, complex instruction following, and intelligence, achieving 10x efficiency compared to similar products. CEO Michael Truell stated this marks the beginning of a collaboration with SpaceXAI to train a significantly larger model from scratch using the equivalent of one million H100 GPUs on the Colossus 2 supercomputer, increasing total compute investment tenfold. Cursor also announced it will double model usage quotas for all users within the next week. Elon Musk personally promoted the release, noting that part of the model was trained on Colossus 2.

OpenAI has released two major products: Workspace Agents allow users to create reusable workflow agents via natural language, supporting team sharing and background execution, applicable to IT, product, and operations use cases, now available in preview for business, enterprise, and education users. Concurrently, Mac Codex has evolved from a code assistant into a full-fledged intelligent agent capable of directly operating Mac applications through GUI interactions, featuring multimodal vision recognition, screen UI understanding, and actions such as clicking and typing. Codex integrates over 90 office and development plugins, enabling cross-platform information retrieval across Slack, Gmail, Notion, etc., and runs securely via macOS Accessibility APIs.

OpenAI has partnered with Dell Technologies to integrate Codex coding and agent AI tools into Dell's AI data platform and AI Factory, enabling enterprises to securely run AI agents in hybrid cloud and on-premise environments. Codex is used weekly by over 4 million developers, with applications expanding to report generation, customer lead screening, and feedback routing in enterprise workflows. This partnership allows Codex to operate closer to internal codebases and sensitive data, meeting data residency and compliance requirements. On the same day, Dell also launched an on-premise AI operating system with Palantir and partnered with SpaceXAI to bring Grok models into enterprise on-premise settings.

AI research lab Decart has secured $300 million in a new funding round led by Radical Ventures, with participation from Nvidia, Sequoia, Benchmark, Adobe, and Toyota, bringing total funding to over $450 million and a valuation nearing $4 billion. The company unveiled three products: DOS inference and training platform (achieving 1600 tokens/sec), Lucy world model (for immersive experiences with sub-30ms latency), and Oasis physical AI model for robotics and autonomous driving. Decart is among the first companies to deploy large-scale real-time AI models on AWS Trainium and plans to launch DOS 2.0, Lucy 2.5, and Oasis 3 soon.

Anthropic announced the acquisition of Stainless, an SDK and MCP server platform that has provided SDK support since the early days of the Anthropic API. This acquisition strengthens Anthropic's capabilities in developer tools and infrastructure. On the same day, Anthropic also doubled token limits for all Claude Design plans. Additionally, Hitachi announced a strategic partnership with Anthropic to deploy Claude across its global workforce of approximately 290,000 employees, train around 100,000 AI professionals, and establish a frontier AI deployment center spanning North America, Europe, and Asia.

Sapient Intelligence has released and open-sourced HRM-Text, a 1-billion-parameter brain-inspired hierarchical reasoning language model. Using a dual-stack nested recurrent architecture, it performs multi-step internal reasoning in continuous latent spaces, trained on just 40 billion tokens—about 1,000 times less than mainstream models—at a cost of approximately $1,000 and completed within a day. It achieved strong benchmark results: MATH 56.2%, ARC-Challenge 81.9%, DROP 82.2%, MMLU 60.7%. When quantized, it occupies only 0.6 GiB of memory, enabling offline execution on smartphones. The model has been integrated into the Hugging Face Transformers library, and the 1B checkpoint is publicly available.

Salesforce CEO Marc Benioff revealed the company expects to spend nearly $300 million in 2026 purchasing Anthropic's AI model tokens, primarily for AI-driven software development and coding assistance. Approximately 15,000 of the company's engineers are currently collaborating with AI systems such as Anthropic, OpenAI Codex, and Cursor, improving engineering efficiency by over 30%. Agentforce has achieved $800 million in annual recurring revenue, and AI technologies now handle 30%-50% of the company's workload. Although Salesforce paused software engineer hiring in 2025, it plans to hire 1,000–2,000 additional sales personnel to promote AI products and recruit 1,000 graduates and interns in the near term.

ByteDance has released Lance, a 3-billion-parameter multimodal AI model under the Apache 2.0 open-source license, supporting image and video understanding, generation, and editing. Designed for efficient inference and rapid product integration, the model was trained using no more than 128 A100 GPUs, making it suitable for commercial experimentation by startups. Its shared multimodal sequence architecture combines dedicated expert modules for understanding and generation, performing well on benchmarks like VBench. The Apache 2.0 license reduces legal barriers, allowing developers to freely use, modify, and distribute the model.

On May 18, the UAE Cabinet approved the 'UAE Government 4.0' initiative, aiming to train 80,000 federal government employees in Agentic AI technologies, with the goal of becoming the world's first nation to deploy Agentic AI in 50% of government services and operations. The training program covers all levels from ministers to new hires, delivered via a personalized learning platform across five categories: leadership, technical, and professional tracks. The cabinet also approved the first public service package utilizing Agentic AI for citizens, residents, businesses, and investors, and passed a national policy promoting AI adoption in healthcare.