Analyst: Caroline
QuantumBit Think Tank | WeChat: AI123All
Today is May 12, 2026. Key AI news today: Cerebras IPO oversubscribed 20x; Thinking Machines Lab releases 'interactive model'; OpenAI launches software security project Daybreak and establishes DeployCo; DeepSeek adds chat history search; Kuaishou plans to spin off Kling AI for independent funding and IPO.
Infrastructure
Cerebras IPO Oversubscribed 20x, Plans to Raise Offering Price
Cerebras' IPO has received over 20 times oversubscription, with investor demand far exceeding supply. The company has significantly raised its price range from $115-$125 per share to $150-$160 per share, and increased the offering size from 28 million shares to 30 million shares. At the top of the new range, Cerebras would raise approximately $4.8 billion, with a valuation of around $35 billion. The ticker symbol is CBRS, and the final IPO price is expected to be set on May 13, 2026.
In 2024, Cerebras' IPO was shelved due to concentrated exposure to G42. Through partnerships with OpenAI and AWS, the company reduced G42's revenue share from 87% to 24% in just six months, turning a net loss of $485 million in 2024 into a net profit of $87.9 million in 2025. However, the prospectus still describes a single company (OpenAI) as a 'major source of revenue for the foreseeable future'. OpenAI plans to deploy 750 megawatts of AI computing power from Cerebras between 2026 and 2028, with an estimated total potential value of over $20 billion. OpenAI also provided Cerebras with a $1 billion loan and obtained 33.5 million warrants. AWS subsequently announced that it would integrate Cerebras CS-3 chips into its Bedrock platform. The dual support from two major AI infrastructure buyers has been the core catalyst for this oversubscription round.
Models
Thinking Machines Lab Releases Native Multimodal 'Interactive Model'
Thinking Machines Lab has released a research preview of its first interactive model, TML-Interaction-Small, a 276B-parameter mixture-of-experts model (12B activated parameters) specifically designed for real-time, multimodal human-machine collaboration. 'Interactivity' means the model can generate responses while the user is speaking and initiate speech. The model continuously processes audio, video, and text input streams in 200-millisecond units. The system has two layers: the surface interactive model maintains two-way communication with the user, while the underlying background model handles deep reasoning and long-chain tasks, enabling parallel real-time interaction and background thinking.
Applications
OpenAI Daybreak: AI Security System for Cyber Defenders
OpenAI has launched Daybreak, a software security defense project for enterprises, positioning it as a competitor to Anthropic's Glasswing solution. The project integrates OpenAI models, Codex security agents, and capabilities from partners like Intel, Cisco, and CrowdStrike into the development workflow. Daybreak upgrades from GPT-5.4-Cyber, which has fixed over 3,000 vulnerabilities. Three model versions are available, with GPT-5.5-Cyber targeting authorized professional security workflows. OpenAI also established a 'Trusted Access for Cyber' tiered access system, granting varying levels of AI model access based on user identity and task risk.
For large customers like Apple, Microsoft, and Google that already use Anthropic Glasswing, OpenAI's pricing competitiveness and false positive control will determine whether Daybreak can win market share from rivals.
OpenAI Launches DeployCo to Help Enterprises Put AI Systems into Production
Enterprise deployment has become a competitive focus. OpenAI announced the formation of OpenAI Deployment Company (DeployCo), a commercial entity specialized in enterprise AI system deployment and integration, with an initial investment of $4 billion and a valuation of $10 billion. DeployCo's core model involves embedding 150 Forward Deployed Engineers (FDEs) specialized in cutting-edge AI applications within client organizations to deeply integrate models with existing data, tools, permission systems, and workflows to create stable AI systems. OpenAI acquired these approximately 150 professional deployment talents by acquiring British AI consultancy Tomoro, which has practical experience from AI strategy assessment to deploying custom AI agents within 12 weeks at companies like Tesco and Virgin Atlantic. As of now, enterprise customers contribute over 40% of OpenAI's revenue, expected to match consumer business by year-end. Cisco's 2025 AI Readiness Index showed only about 13% of global enterprises are fully prepared for AI, with Asia-Pacific at 11%. DeployCo aims to bridge the gap between technical demonstrations and business deployment.
DeepSeek Adds Chat History Search in Beta Test
DeepSeek is currently beta-testing a chat history search feature in App version 2.1.0 (213). After updating, a 'Search chat content' box appears at the top of the sidebar, allowing users to enter keywords to precisely find historical conversations containing that word and jump to the specific location. The web version also supports similar search via a magnifying glass icon. This feature is a key experience enhancement for power users. ChatGPT, Claude, and other mainstream products already support historical chat search; DeepSeek, as a leading domestic app, had a notable gap in this basic function. The update catches up with industry standards in product maturity. Currently, the feature only supports exact keyword matching, leaving room for retrieval efficiency improvement.
Kuaishou Plans to Spin Off Kling AI, Seek Independent Funding and IPO
Market sources say Kuaishou plans to spin off its AI video generation business, Kling AI, seeking to raise about $2 billion at a $20 billion valuation. This valuation exceeds half of parent company Kuaishou's market cap of approximately HK$230 billion, highlighting the high premium for pure AI companies. Kling AI had already achieved large-scale commercialization in 2025, with cumulative annual revenue of about 1.04 billion yuan ($144 million). In Q1 2026, revenue reached $75 million, with an estimated ARR of about $500 million by end of April 2026. Although Kuaishou's overall market cap is large, its traditional short video cash cow cannot support the high valuation expectations for cutting-edge AI. After the spin-off, Kling AI could enter the AI track with a 67x price-to-sales ratio, enjoying the valuation premium for high-growth tech companies. Kuaishou responded that the board is evaluating a proposed restructuring that may involve introducing external financing for Kling AI. The company emphasized that no final agreements have been signed and the proposal may not proceed.
Policy
China's Cyberspace Administration Promotes AI Labels for Short Videos, Making AI-Generated Content Tags Mandatory
AI content labeling is shifting from 'platform self-discipline' to 'mandatory publication'. The Cyberspace Administration of China (CAC) recently deployed and promoted standardized content labeling for short videos. In March 2026, it guided 12 leading platforms including Douyin, Kuaishou, Tencent, Xiaohongshu, Bilibili, Weibo, and Taobao to pilot and test content labeling features. Based on the pilot experience, the CAC now directs all Chinese websites and platforms to implement three standards: 1) Platforms must provide users with six 'mandatory labels' including 'Contains fictional/performative content', 'Contains AI-generated content', 'Contains marketing information', 'Content is reposted', 'Content is personal opinion', and 'No label needed'; 2) Content labeling is set as a mandatory pre-publishing step for short videos, requiring publishers to select one label before posting; 3) Enhanced review of newly uploaded short videos for labeling, and batch retrospective labeling of existing videos, aiming for full coverage.
—
Contact Author
—
— End —
[QuantumBit Think Tank] Original content. Unauthorized reproduction prohibited.