According to Bloomberg and other media reports, Anthropic is in talks with Goldman Sachs, JPMorgan Chase, and Morgan Stanley to prepare for an IPO in October 2026, aiming to raise over $60 billion. The company completed a $30 billion funding round in February at a $380 billion valuation. Its annualized revenue has grown from $1 billion to nearly $19 billion within 14 months, primarily driven by the rapid commercialization of Claude Code. Over 500 enterprise customers now spend more than $1 million annually, eight of the top ten Fortune companies use its products, and consumer monthly visits have surged from 16 million to 220 million. Meanwhile, OpenAI is also preparing for an IPO in the second half of 2026 with a target valuation of $1 trillion, setting up direct competition between the two AI giants. Anthropic's business model focuses more on enterprise payments, and the success of its IPO will depend on computational cost control and overall market confidence in AI investments.

Google plans to provide over $5 billion in financing support for Anthropic’s large-scale AI data center campus in Texas. Operated by Nexus Data Centers, the project spans 2,800 acres and is expected to deliver approximately 500 megawatts of power capacity by the end of 2026, with long-term potential to scale up to 7.7 gigawatts. Google is expected to offer construction loans, while multiple banks are competing for subsequent financing roles. This move deepens the collaboration between Google and Anthropic, which previously planned to access up to 1 million TPUs via Google Cloud for training and deploying Claude models. Meanwhile, Anthropic is involved in a legal dispute with the U.S. Department of Defense over its refusal to relax restrictions on AI use in military surveillance and autonomous weapons; a federal judge has ruled to temporarily block the Pentagon from listing it as a 'supply chain risk.'

Chroma has released Context-1, a specialized agentic search model with 20 billion parameters based on the MoE architecture of gpt-oss-20B, designed specifically for multi-hop retrieval and context management. Its core innovation is the 'self-editing context' mechanism, achieving a pruning accuracy of 0.94, enabling active removal of irrelevant information during search to prevent context noise accumulation. The model decomposes complex queries into sub-queries and performs parallel tool calls (averaging 2.56 per query) for iterative retrieval. In performance, Context-1 matches large models like GPT-5 in retrieval effectiveness, offers 10x faster inference speed, and reduces runtime costs by 25x. Chroma has also open-sourced the context-1-data-gen tool for generating multi-hop synthetic datasets across web, SEC filings, patents, and emails. This release advances the development of layered RAG architectures that decouple retrieval from generation.

Amazon confirmed in March 2026 the acquisition of startup Fauna Robotics, founded by former Meta and Google engineers, which developed Sprout—a 42-inch bipedal humanoid robot. Weighing 50 pounds and priced at $50,000, Sprout runs on the Nvidia Jetson Orin platform and features swappable batteries, facial recognition, voice response, and long-term memory capabilities, focusing on social interaction scenarios, especially in households with children and pets. Unlike Tesla Optimus, which targets physical labor, Amazon aims to integrate Sprout with Alexa and Prime ecosystems to build a home social robotics platform. Despite previous setbacks in home robotics (e.g., Astro and iRobot), this acquisition marks a strategic shift toward software-first, emotionally connected robots.

On March 29, 2026, at the Universal Artificial Intelligence Forum during the Zhongguancun Forum Annual Conference, the world's first general intelligence human 'Tongtong' 3.0 was officially launched. This version achieves major upgrades in spatial, cognitive, and social intelligence, enabling precise perception of the physical world, autonomous planning of complex tasks, and recognition of emotions and intentions. It engages in human-like interactions within a 3D simulated 'AI Town' platform hosting multiple intelligent agents. Equipped with consistent values and worldview, its decision-making process is explainable and traceable, effectively mitigating LLM hallucinations. Simultaneously introduced is the embodied AI core engine 'Tongnao,' designed to bridge bidirectional communication between general intelligence agents and physical robots through integration of motion control, reinforcement learning, and world modeling technologies, enabling efficient cross-scenario and cross-task skill acquisition. Additionally, the Tongyu AI education platform was released and has been deployed in over 40 primary and secondary schools.

China's YuanLab has released Yuan 3.0 Ultra, a trillion-parameter AI model utilizing MoE architecture and layer-adaptive expert pruning technology, activating only 68.8 billion parameters during inference—improving efficiency by 49%—and matching international leaders such as GPT-5.2 and DeepSeek V3 in performance. Trained on 824 AI chips across 103 layers, the model demonstrates efficient expert routing and load balancing. Released under an open-source license, its innovative expert routing mechanism lowers the barrier for SMEs to adopt high-end AI. Widely applicable in coding, retrieval, and automation, the model reflects China's strategic push to reduce dependence on foreign technologies and enhance global competitiveness in AI.

Researchers affiliated with Amazon have introduced A-Evolve, a general infrastructure framework for autonomous AI agents that replaces manual tuning with automated state mutation and self-correction. The core design treats agents as collections of files evolved through environmental feedback, using a standardized 'agent workspace' structure (manifest.yaml, prompts, skills, tools, memory). Optimization occurs via a five-stage evolutionary cycle (solve, observe, evolve, gate, reload), with Git integration ensuring full change traceability and automatic rollback. The system achieved strong results across benchmarks: 79.4% on MCP-Atlas (ranked first), 76.8% on SWE-bench Verified, and 76.5% on Terminal-Bench 2.0. The entire evolution process requires no human intervention and can be integrated into existing Python workflows with just three lines of code.

Anthropic researchers used Claude Opus 4.6 to discover the first high-risk vulnerability in Ghost CMS and a 20-year-old stack buffer overflow vulnerability in the Linux kernel—all within 90 minutes—demonstrating AI's exponential growth in zero-day vulnerability discovery. Studies show AI can scan all popular GitHub repositories within hours and automatically identify exploit chains, compressing traditional defense windows from months to mere hours. The doubling period for AI model capabilities in security is only about four months. AI excels most at detecting the very types of vulnerabilities that are hardest for humans to find and most dangerous when exploited. Leaked internal documents also reveal that Anthropic's next-generation model Mythos (codename Capybara), currently under testing, significantly surpasses existing models in cyber offense and defense, and the company plans to prioritize early access for defensive security organizations.

The European Parliament's Committee on Civil Liberties and Internal Market adopted its negotiating position on the 'Artificial Intelligence Omnibus Act' on March 30, 2026, aiming to simplify the AI Act and extend compliance deadlines. The position supports delaying enforcement of high-risk AI rules to December 2027 and August 2028, and calls for a 12-month grace period for generative AI labeling requirements. CCIA Europe emphasized the need for swift agreement to avoid legal uncertainty. Originally, broad compliance for high-risk AI systems was set to fully take effect on August 2, 2026, with violations subject to fines up to €35 million or 7% of global annual turnover. Developer communities are urgently preparing technical compliance solutions, including audit logs at the tool-calling level and pre-execution policy interception.

San Francisco-based enterprise software company Doss announced a $55 million Series B round co-led by Madrona and Premji Invest, with participation from Greyhound, Commerce Ventures, and Intuit Ventures. Doss is building an AI-native 'Operations Cloud' platform for inventory-based businesses, leveraging composable data architecture and AI-driven automation (e.g., Dossbot copilot) to enable real-time dynamic adjustments in procurement, inventory, and order processes, supporting natural language commands. Funds will be used to develop a unified data model, expand intelligent agent systems, and advance adoption as a replacement for traditional ERP. Early customers report significant gains: Verve Coffee Roasters reduced manual batch orders from 30% to 1%, saving over 20 hours weekly; Mezcla doubled its purchase order processing speed.