
Welcome back! An $81.34 billion IPO put Anthropic at the centre of an extraordinary week, with the company pricing the largest initial public offering on record. Nvidia reported $115.8 billion in quarterly revenue, while California sued OpenAI over alleged cybersecurity failures. Washington tightened controls on remote access to advanced computing, and San Francisco moved to restrict new data centres. AI's financial scale is becoming extraordinary. So are the questions about security, oversight and who gets to decide where the infrastructure goes.
- Anne, and the News DigestAI team

📰 Biggest AI Stories This Week
Anthropic priced its initial public offering at $83 per share, valuing the 980 million shares offered at approximately $81.34 billion. The record-sized IPO included shares sold by existing investors, meaning not all proceeds go to Anthropic. Its shares began trading on Nasdaq, marking an extraordinary moment for public-market investment in AI.
Nvidia reported quarterly revenue of $115.8 billion, up 103% year on year. Data-centre revenue alone reached $107.6 billion, while net income climbed to $65.3 billion. Nvidia forecast approximately $130 billion in revenue for its next quarter, excluding any assumed Chinese data-centre compute sales. The figures underline the immense spending supporting the AI infrastructure boom.
California's Attorney General filed a lawsuit against OpenAI, alleging the company failed to adequately safeguard experimental AI systems. The complaint refers to a cyberattack involving Hugging Face and alleges at least 80 cybersecurity incidents. California argues that insufficient oversight endangered the public and violated state law. These are allegations in a newly filed case, not findings established by a court.
Alibaba Group reported a 53% increase in AI Cloud and Compute Services revenue, reaching RMB60.97 billion, approximately $9.09 billion, for the quarter. Revenue from external cloud customers grew 64%, while the division's adjusted operating profit rose 121%. The figures provide fresh evidence of commercial demand for AI computing beyond the US market.
The European Commission confirmed an AI-assisted cyber campaign against its systems during an official briefing. Several applications were temporarily taken offline, with most subsequently restored. Investigators have not publicly identified the attacker, and the Commission said it had found no evidence of a breach affecting its staff as a whole.
The UK Department of Health and Social Care published national guidance on AI in social care, covering children's and adults' services in England. The government highlighted a Swindon pilot reporting a 63% reduction in social workers' administrative time. The guidance addresses accountability, data protection and professional oversight, although it is non-statutory and the pilot's reported gains should not be treated as proven nationwide results.
Blackstone Digital Infrastructure Trust announced an agreement to acquire controlling interests in two Amazon Web Services data-centre joint ventures, involving properties valued at approximately $5 billion. The portfolio includes 13 newly constructed buildings across two campuses, with Amazon retaining a minority interest and leasing the facilities under long-term agreements. The deal shows how investors are seeking long-term returns from the physical infrastructure supporting cloud computing and AI. Completion remains subject to conditions.
Oak Ridge National Laboratory opened its Lux supercomputer to scientific users, reporting 220.79 petaflops of computing performance and 14th place on the global Top500 ranking. Built through a partnership involving AMD and HPE, the system supports US government research into fusion energy, advanced materials and environmental forecasting. The facility provides operational computing capacity for national science programmes, although the expected research benefits are still to be demonstrated.

⚡ New AI Tools Worth Testing
Inform Scan launched a smartphone application that turns property walkthrough videos into floor plans, measurements and structured building records. Estate agents, surveyors and property valuers can capture property details without specialist scanning equipment or repeated manual measurements. The app is available on iPhone and Android in the US, with the first order free and marketing floor plans starting at $12 per property.
Natural Accept launched a payment collection system for businesses using AI agents. It enables agents to send secure checkout links or present embedded payment forms, with proceeds deposited into a Natural wallet for subsequent transactions. The service currently supports verified US businesses processing payments in US dollars and requires integration with Natural's platform.
FlowForma Agent Trust was introduced to help businesses monitor and manage AI agents operating across different software platforms. It identifies agents, tracks their responsibilities, costs and outcomes, and allows managers to pause activity or require human review. The product is aimed at organisations deploying multiple AI workflows, with a free trial available.
Waldo entered public preview as an AI agent that helps businesses analyse geographical information using natural-language instructions. Teams can investigate satellite imagery, location data and geographical patterns to support decisions involving property, logistics, insurance or agriculture, without manually building every analysis. It runs within the Wherobots platform, where free trials are available, although computing usage may incur charges.
Orama Lake was released as open-source software, giving businesses a self-hosted platform for building AI applications and agents using their own data. It combines data storage, search, model development and deployment tools, potentially reducing reliance on separate services when information must remain on company-controlled infrastructure. The software is free under the MIT licence but requires technical installation and a machine with at least 16 GB of memory.

🆕 New AI Models
GPT-6.2 Astra was released by OpenAI, targeting complex professional work, scientific research and software development. The model improves how it produces documents, spreadsheets, presentations and code, with OpenAI reporting stronger performance while using fewer output tokens. It is available through the API, ChatGPT Work and Codex, with access depending on subscription and workspace settings.
GPT-6.2 Sol was introduced alongside Astra as a more economical model for everyday professional tasks, coding and multistep agent workflows. OpenAI reports improvements in reasoning, writing and token efficiency compared with GPT-6.1 Sol. It is available through the API, ChatGPT Work and Codex, with standard API pricing starting at $2 per million input tokens.
Gemini 4 Neon was launched by Google, introducing a new frontier model designed for complex coding, scientific analysis and professional workflows. It processes text, documents, charts, audio and video within a one-million-token context window, with particular emphasis on executing demanding multistep tasks. Access began through the Gemini API, Google AI Studio, Antigravity and selected paid Gemini subscriptions.
MAI Code 2 and Code 2 Flash were released by Microsoft as two coding models designed for different workloads. Code 2 handles repository-wide reasoning, debugging and changes spanning multiple files, while Flash focuses on faster code explanations, targeted edits and routine fixes. Both are available through GitHub Copilot, with Code 2 Flash also offered in a limited Microsoft Foundry preview.
MAI Realtime 1 was introduced by Microsoft as its first speech-to-speech AI model. It supports natural conversations, interruptions, background-noise handling and connections to enterprise tools, enabling developers to build voice assistants that can respond and perform tasks during a conversation. The model can be tested in MAI Labs, while Microsoft Foundry access remains in limited preview.
MAI Image 2.7 Flash was made available by Microsoft as a new, lower-cost variant of its existing Image 2.7 model. Designed for high-volume image generation and editing, Flash supports controlled visual changes and transparent backgrounds, with indicative pricing of approximately $10 per 1,000 images. It is accessible through MAI Labs and public-preview access in Microsoft Foundry.
Tencent Hy4 Flash was released by Tencent as a faster, cheaper alternative to its larger Hy4 model. Designed for everyday business and developer workloads, it supports long-context processing while substantially reducing inference costs. Tencent made the model available through Tencent Cloud TokenHub and WorkBuddy, giving organisations another option for cost-sensitive AI deployments.
Liquid AI d1.2 and d1.2-ultra were released by Liquid AI, expanding the company's specialised decision models with audio input and five additional task types. The models can extract values with supporting evidence, identify passages, locate objects in images and produce structured classifications or numerical estimates. Both are available through Liquid's API, targeting applications that require consistent, verifiable outputs rather than open-ended text generation.
Olala v1.0 mini 7A1B had its model weights released by OVHcloud, followed by API availability. The compact reasoning model uses a hybrid architecture with fewer active parameters to reduce computing requirements, and supports tool calling and structured outputs. Released under Apache 2.0, it offers developers a European open-weight alternative that can run on relatively modest hardware.
AssemblyAI-STT-3.7 was released by AssemblyAI as a new family of speech-recognition models covering recorded audio, live conversations and voice agents. The realtime versions introduce immediate speaker attribution and improved handling of short replies, background speech and multilingual conversations. All three models are available through AssemblyAI's APIs, giving developers more specialised transcription options for different audio workloads.
🆕 New AI Agents
Perplexity Computer Multiagent was launched by Perplexity, introducing a substantial new collaboration capability within its existing Computer platform. A lead agent can coordinate up to 16 agents using different AI models, allowing them to exchange findings, challenge assumptions and edit shared files while completing complex research or analytical tasks. The Multiagent effort setting is available to Perplexity Max subscribers on the web.
Replit Missions was launched by Replit, introducing coordinated teams of AI agents that can plan, build, test and refine solutions to complex business or engineering problems. Users specify the desired outcome, spending limit and actions requiring approval, while the system manages the work across multiple agents. Missions extends Replit beyond individual application-building tasks towards larger projects with measurable objectives and human oversight.
Fountain Orchestra was introduced by Fountain as an agent-building platform for frontline business operations. Users can describe a task in natural language, create an agent and run it on demand or according to a schedule, with controls over permissions and approvals. Initial capabilities are available, while unattended event-triggered execution and more advanced hand-offs are scheduled for subsequent releases.
Intershop Order Entry Agent was introduced in Intershop's platform release. The agent converts incoming orders from emails, PDFs, spreadsheets and other formats into digital orders, automatically applying customer-specific pricing and approval rules. Available within Copilot for Merchants, it can also process orders through Microsoft Teams, reducing manual data entry while retaining human approval controls.
That’s a wrap! See you next Friday for more.

