Google DeepMind launched two autonomous research agents, Deep Research and Deep Research Max, on April 21, built on Gemini 3.1 Pro. The standard version optimizes low-latency interaction, while Max focuses on asynchronous deep analysis suitable for enterprise scenarios like due diligence. Deep Research Max scored 93.3% on DeepSearchQA, 54.6% on HLE, and 85.9% on BrowseComp, outperforming GPT-5.4 across all benchmarks. New features include MCP protocol support for connecting to enterprise private data sources (already partnered with FactSet, S&P, and PitchBook), native chart generation, multimodal input, and real-time streaming inference output. Both agents are now in public preview via the paid Gemini API, with future expansion planned for Google Cloud.

Amazon and Anthropic have deepened their collaboration, with Amazon committing up to $25 billion in additional investment—$5 billion immediately and the remainder tied to commercial milestones—on top of its prior $8 billion investment. In return, Anthropic has pledged to spend over $100 billion on AWS technologies over the next decade, including multiple generations of chips from Trainium2 to Trainium4 and tens of millions of Graviton cores, and will gain access to up to 5 gigawatts of compute capacity. The two companies will also expand inference capabilities in Asia and Europe. The Claude platform will be integrated into AWS, allowing users to access it directly through their existing accounts. Currently, more than 100,000 customers are running Claude models on Amazon Bedrock.

Meta plans to install tracking software called Model Capability Initiative (MCI) on U.S. employee work computers to collect mouse movements, clicks, keystrokes, and occasional screen snapshots, aiming to train AI agents capable of autonomously performing job tasks. This initiative is part of its Agent Transformation Accelerator (ATA) program. Meta states that the data will only be used for model training, not performance evaluation, and measures have been taken to protect sensitive content. However, the move has raised concerns about privacy and labor rights and may violate GDPR regulations in Europe. Meanwhile, Meta also plans global layoffs of approximately 10% starting May 20 and will reorganize employees into 'AI Builder' roles.

SpaceX has announced a special agreement that could lead to a $60 billion acquisition of AI coding platform Cursor, or alternatively a $10 billion partnership fee. The move aims to strengthen the technical capabilities of Elon Musk’s xAI to compete with rivals like Anthropic. SpaceX owns the Colossus supercomputer, equivalent to a million H100 GPUs, while Cursor is widely used among professional software engineers. Their collaboration targets building the most practical AI models. Earlier reports from CNBC indicated Cursor was seeking a $50 billion valuation to raise $2 billion in funding; if completed, this deal would rank among the largest acquisitions in the AI industry.

Variable Robotics unveiled WALL-B on April 21, the world's first embodied intelligence foundation model based on the Unified World Model (WUM) architecture, enabling unified training of vision, language, action, and physical prediction with native proprioception and understanding of physical laws. Unlike traditional VLA architectures, WALL-B is trained on data from real home environments, enhancing adaptability in complex scenarios, and employs visual anonymization and transparent consent mechanisms to ensure privacy. The company announced that a new generation of robots powered by this model will begin entering real homes in 35 days. Technical details will be revealed at the Shenzhen Artificial Intelligence Conference on April 27.

Shenghong Technology, the global leader in market share for AI and high-performance computing PCBs, listed on the Main Board of the Hong Kong Stock Exchange on April 21 at HK$209.88 per share, achieving a market capitalization of HK$315.44 billion and becoming the largest IPO in Hong Kong this year. The company secured 38 cornerstone investors who collectively subscribed approximately $997 million. In the first half of 2025, its market share in AI and high-performance computing PCBs reached 13.8%, with products covering AI accelerator cards, servers, and data centers. Financially, revenue grew from RMB 7.931 billion in 2023 to RMB 19.292 billion in 2025, while annual profit rose from RMB 671 million to RMB 4.312 billion. Approximately 74% of the proceeds will be allocated to expanding smart manufacturing capacity.

Cast AI released its 2026 Kubernetes Optimization Report, analyzing data from 23,000 enterprise-grade Kubernetes clusters, revealing that enterprise GPU utilization averages only 5%, with CPU utilization at 8%, far below the recommended healthy threshold of 50%. Immature GPU scheduling, data pipeline delays, reliance on Nvidia GPUs, and long-term procurement contracts contribute to severe resource misallocation. Enterprises are over-provisioning GPU resources due to 'insurance-style' fears of falling behind in AI competition, leading to massive capital waste. The report recommends adopting FinOps practices in AI, including intelligent scheduling, elastic scaling, and accelerator diversification, with GPU utilization as a core management metric.

Jeff Bezos’ AI startup 'Project Prometheus' is close to finalizing a $10 billion funding round, with investors including JPMorgan and BlackRock. The company focuses on physical world understanding and engineering applications, extending AI capabilities from digital domains into physical environments. This marks another major AI play by Bezos following his large-scale investments in Anthropic, placing the funding scale among the highest for AI startups.

SoundHound AI announced the acquisition of enterprise conversational AI provider LivePerson for $43 million in equity value (approximately $250 million including debt), integrating its proprietary voice AI technology with LivePerson’s Conversational Cloud platform, which processes nearly 1 billion digital messages monthly. The combined platform will offer an end-to-end omnichannel conversational AI solution serving customers across more than 30 countries, including 25 Fortune 100 companies and 12 of the world’s top 15 banks. SoundHound forecasts 2027 revenue between $350 million and $400 million, with cross-selling potential reaching $500 million. The transaction is expected to close in the second half of 2026. Following the announcement, SoundHound’s stock price dropped 5.65%.

Tesla's in-car voice large model service completed generative artificial intelligence service filing in Shanghai, China, on April 20, under Tesla (Shanghai) Co., Ltd., bringing the total number of filed models in Shanghai to 158. According to regulations, launched generative AI applications must prominently display filing information and include identifiers for synthetic content. Previously, Tesla’s xBots customer service became one of the first approved large model products by foreign enterprises in November 2025. This latest filing signifies a further deepening of Tesla’s intelligent services in the Chinese market.