Daily Brief Archive

Inferact completes $150M seed round at $800M valuation to advance vLLM commercialization and serverless inference; Sugon Technology's STAR Market IPO accepted with plans to raise 6B yuan for 5th/6th-gen AI chips; Google DeepMind launches D4RT enabling 2D-to-4D reconstruction with 18–300x speedup

Baidu launches 2.4 trillion parameter natively multimodal Ernie 5.0 and opens Qianfan platform for API access; Alibaba reportedly preparing IPO for AI chip unit T-Head starting with restructuring and employee equity; Google Cloud achieves over 6,000 tokens per GPU per second and ~10ms ITL for MoE inference on A4X+GB200 NVL72 with NVIDIA Dynamo

OpenEvidence completes $250 million Series D, valuation rises to $12 billion and claims coverage among approximately 40% of U.S. doctors; OpenAI discloses compute capacity growing from 0.2GW to 1.9GW between 2023 and 2025, with $22 billion in spending and $9 billion in revenue in 2025; Yelp acquires AI lead management platform Hatch for $270 million in cash to strengthen SaaS and customer acquisition capabilities for local businesses

Humans& completes $480 million seed round, valued at $4.5 billion; ServiceNow signs three-year agreement with OpenAI to integrate GPT-5.2 into enterprise workflows; Jueyi Stars open-sources Step3-VL-10B and discloses full-parameter training on 1.2T data

ClickHouse completes $400 million funding, valuation exceeds $15 billion, and strengthens investment in AI real-time analytics infrastructure; Google fixes indirect prompt injection vulnerability in Gemini calendar invites that could lead to leakage of private meeting information; Microsoft Azure Functions support for MCP servers goes GA with built-in OBO authentication and streamable HTTP transport

Anthropic reportedly plans to raise $25 billion at a valuation of approximately $350 billion, with GIC and Coatue each investing $1.5 billion as lead investors; xAI-related reports indicate the 1GW Colossus 2 cluster is now online and planned to scale to 1.5GW by April for training Grok 5; China Telecom's TeleAI has open-sourced TeleChat3-105B-A4.7-Thinking with approximately 4.7B active parameters

Marvell to acquire Celestial AI for $3.25 billion to integrate photonic interconnects; Google Vertex AI exposes a privilege escalation path allowing low-privilege users to obtain Service Agent tokens, deemed 'by design'; Pixel2Play open-sources 8,300 hours of labeled data and code for general real-time game AI

Meituan open-sources and releases LongCat-Flash-Thinking-2601, featuring re-thinking with 8-way parallel reasoning to enhance tool usage generalization; OpenAI pilots ads in ChatGPT free and $8/month Go versions in the US, stating ads won't affect responses or share conversations with advertisers; ClickHouse acquires Langfuse and commits to keeping the project MIT-licensed and self-hostable, disclosing over 23.1 million monthly SDK installs

OpenAI signs over $10 billion compute deal with Cerebras, securing up to 750MW of low-latency compute capacity through 2028; Google Cloud enables BigQuery to host SQL-native inference for Hugging Face and Model Garden open models (preview); GitHub Copilot introduces a verifiable memory system, boosting PR merge rate by 7% in A/B testing