AI chipmaker Cerebras Systems has priced its IPO at $185 per share, exceeding the previously guided range of $150–$160, issuing 30 million Class A ordinary shares to raise $5.55 billion. The IPO values the company at a fully diluted valuation of $56.4 billion, making it the largest tech IPO so far this year. Investor demand reached 20 times the available supply. Cerebras reported 2025 revenue of $510 million, up 76% year-on-year. The company has signed a compute supply agreement with OpenAI worth over $20 billion. Shares will begin trading on May 14 on the Nasdaq under the ticker CBRS. Morgan Stanley, Citigroup, Barclays, and UBS are serving as joint bookrunners.

Recursive Superintelligence, founded by former Salesforce chief scientist Richard Socher, has announced $650 million in funding, valuing the company at $4.65 billion. GV and Greycroft led the round, with participation from Nvidia and AMD's venture arms. Founded in early 2026, the team now exceeds 25 members across San Francisco and London. Its goal is to build recursive self-improving superintelligence systems that autonomously optimize their own code and training infrastructure through automated experiment design, testing, and validation. Team members hail from top institutions including OpenAI, DeepMind, and Meta, with plans for public launch by mid-2026.

Meta CEO Zuckerberg announced the launch of Incognito Chat, a feature based on end-to-end encryption where conversations with Meta AI are not stored on servers and messages are permanently deleted upon leaving the session—unaccessible even to Meta itself. In comparison, Google Gemini’s temporary chat retains data for up to 72 hours, ChatGPT for 30 days, and Claude for at least 30 days. Built on last year’s Private Processing technology, the feature will roll out gradually across WhatsApp and Meta AI apps in the coming months, responding to recent legal suits over AI conversation data retention.

OpenAI has launched GPT-5.3-Codex-Spark, its first AI product not reliant on Nvidia hardware, running on Cerebras WSE-3 chips. It achieves generation speeds exceeding 1,000 tokens per second—15 times faster than standard GPT-5.3-Codex—and is optimized for real-time code collaboration, supporting interruption and response during generation. Currently available only to ChatGPT Pro users via desktop app, command line, and VS Code plugin. Cerebras’ wafer-scale processor eliminates multi-GPU communication overhead, a key factor in achieving low latency. OpenAI plans to integrate Spark’s real-time capabilities with Codex’s full long-term planning functionality in the future.

Nvidia has announced an engineering-level partnership with UK-based AI startup Ineffable Intelligence to co-develop next-generation AI systems based on reinforcement learning. Founded by David Silver, former head of reinforcement learning at Google DeepMind, in late 2025, the company focuses on building 'super learners' capable of autonomous learning and discovering new knowledge through experience. The two parties will leverage Grace Blackwell chips and Vera Rubin platform to build scalable reinforcement learning infrastructure. The startup raised a record-setting $1.1 billion seed round in April 2026, led by Sequoia Capital and Lightspeed, with participation from Nvidia, Google, and others.

Anthropic has launched Claude for Small Business, integrating popular tools such as QuickBooks, PayPal, HubSpot, and Canva, offering 15 preset workflows covering finance, operations, sales, marketing, HR, and customer service, enabling automation of tasks like payroll planning, monthly reconciliation, and invoice follow-ups. In parallel, Anthropic has partnered with PayPal to launch free AI training courses and a nationwide training tour. Additionally, Anthropic announced that starting June 15, paid users will receive a dedicated monthly Agent SDK quota, with weekly usage limits for Claude Code temporarily increased by 50%, and free users seeing a fivefold increase.

He Kaiming’s team has introduced ELF (Embedded Language Flows), the first language model using continuous diffusion, differing from mainstream autoregressive or discrete diffusion models. ELF maintains the entire generation process within a continuous embedding space, mapping to discrete tokens only at the final step. With just 105M parameters, 45B training tokens, and 32-step sampling, it reduces generation perplexity to 24 on OpenWebText—significantly outperforming mainstream discrete diffusion models requiring 1024 steps. The model surpasses existing diffusion models in text generation, machine translation, and summarization, validating the efficiency of continuous diffusion paths in language generation for the first time.

METiS (Jitai Tech) listed on the Hong Kong Stock Exchange on May 13, becoming the world’s first publicly traded AI drug delivery company. The IPO raised over HK$2.1 billion, with the public tranche oversubscribed by more than 6,900 times, freezing over HK$730 billion in funds—the highest of the year. The international placement was 82x oversubscribed, also a record. The company specializes in AI-driven nanodelivery technologies, featuring its core NanoForge platform and a library of millions of ionizable lipids. It currently has over 10 R&D pipelines, with MTS-004 being the first AI-powered formulation drug in China to complete Phase III trials. In 2025, it generated revenue of RMB 105 million and has collaborated with over 30 pharmaceutical companies.

Tencent reported Q1 2026 revenue of RMB 196.46 billion, up 9% year-on-year; non-IFRS operating profit was RMB 75.63 billion, up 9% year-on-year, or 17% growth excluding the impact of AI products. Enterprise services revenue grew 20%, driven by rising demand for AI services. R&D investment reached RMB 22.54 billion, up 19% year-on-year; capital expenditures totaled RMB 31.94 billion, primarily for AI infrastructure. Ma Huateng responded to concerns about AI lag, stating the company is advancing steadily. Executives revealed plans to deploy more domestic chips in the second half, with further increases in AI-related capital spending. The Hunyuan Hy3 preview model ranked first on OpenRouter for token consumption for three consecutive weeks.

At Baidu’s AI Developer Conference, Li Yanhong stated that AI has entered an 'execution era' centered on agents, introducing DAA (Daily Active Agents) as a new metric to replace token consumption in measuring AI value. Baidu launched DuMate, a mobile agent product integrating AI search and Mianda capabilities to enable automatic execution from idea to outcome. Meanwhile, 'Baidu Yi Jing' has been upgraded into the world’s first full-scenario multi-agent digital human platform. Li elaborated on the concept of 'Super Individuals': one person plus an agent squad forming the smallest productivity unit. Native AI marketing is expected to generate RMB 2.8 billion in revenue in 2026, up 65% year-on-year.