Welcome back! It’s been a rather expensive week in AI. Anthropic revealed an eye-watering $895 billion in future computing commitments ahead of its IPO, alongside a $23 billion quarterly loss. Wall Street had a wobble, with the Nasdaq plunging 5.5%, while Marvell announced a $115 billion acquisition and Google committed another $17 billion to a data centre. Meanwhile, American voters overwhelmingly rejected local proposals favouring data-centre development. Apparently, everyone wants AI. Paying for it — or having it built next door — is another matter.

- Anne, and the News Digest AI team

Anthropic disclosed an eye-watering $895.4 billion in future computing commitments in its updated IPO filing on 2 November. The company reported nearly $17 billion in quarterly revenue alongside a $23.3 billion net loss, largely influenced by accounting adjustments. Around $70 billion of its infrastructure commitments fall due within 12 months, highlighting the extraordinary cost of competing at the frontier of AI.

Nasdaq experienced a brutal technology-stock sell-off on 5 November, with its Composite index plunging 5.5% and roughly $2.5 trillion in market value disappearing. Nvidia dropped 9.3% and AMD fell 17.3% as investors grew increasingly nervous about AI infrastructure spending, political opposition to data centres and whether enormous investments will generate sufficient returns.

Ballotpedia recorded overwhelming voter opposition to data-centre development following US elections on 3 November. Of 41 local measures decided that day, 40 produced results favouring restrictions or opposing development. Some were advisory rather than binding, but the message is significant: communities are increasingly challenging the rapid expansion of infrastructure needed to power AI.

Marvell Technology announced a $115 billion acquisition of Lumentum, in one of the largest semiconductor deals on record. The proposed merger would bring together electrical and optical connectivity technologies as AI data centres demand faster communication between processors. The all-stock transaction is subject to shareholder and regulatory approval.

Shield announced a $5.3 billion acquisition of radar manufacturer Echodyne on 5 November, alongside plans to raise $2.5 billion in financing at a $28.8 billion valuation. The agreement expands the defence technology company's presence in radar and counter-drone systems, demonstrating the growing commercial value of AI-enabled autonomous defence infrastructure. Completion is expected in early 2027.

Google announced a $17 billion investment in Wyoming on 5 November to establish its first data centre in the state. The development includes plans for a 702-megawatt fuel-cell installation with carbon capture. Google says it will cover its electricity and related infrastructure costs, as pressure mounts on technology companies to prevent AI expansion from increasing household energy bills.

Drax announced a partnership worth potentially more than £20 billion with Brookfield and Radiant on 3 November. The companies plan to develop more than 1GW of data-centre capacity at Drax Power Station in North Yorkshire, potentially creating one of Britain's largest AI infrastructure sites. The first phase targets 2028, subject to planning, regulatory approvals and investment decisions.

TeraWulf secured Google's financial backing on 4 November for a Kentucky data-centre project supporting approximately $19 billion in existing Anthropic lease commitments. Google's agreement to backstop certain lease obligations is designed to make the project easier to finance. The arrangement highlights how major technology companies are increasingly helping underwrite the enormous infrastructure commitments behind AI expansion.

UK Home Office announced plans for tougher facial-recognition safeguards on 4 November, including an independent watchdog, mandatory bias testing and clearer limits on police surveillance in England and Wales. The watchdog could pursue enforcement action against forces that fail to comply. The proposals reflect growing pressure to balance AI-powered policing with privacy and civil liberties, but are not yet law.

US Department of Justice established a Super Intelligence Task Force on 5 November to address potential misuse of advanced AI by foreign governments, criminal organisations and other malicious actors. The initiative will focus on national security, fraud and critical infrastructure threats, signalling growing US government attention to the legal and security implications of increasingly powerful AI systems.

A I T O O L S

Fireworks Nexus Desktop was launched by Fireworks AI to help businesses manage AI costs without abandoning their existing tools. The Mac application connects services including ChatGPT, Claude, Codex and Cursor, automatically selecting models suited to each task while giving teams greater visibility into spending. It requires no command-line setup, with Pro plans starting at $8 per month.

Sysdig AI Defense became generally available, introducing a security application that monitors what AI agents actually do across employee computers and connected cloud systems. Businesses can identify unauthorised agents, block risky actions before they complete and maintain records showing which employee authorised each activity. It supports macOS and Windows, with a 14-day trial available to request.

Dex was introduced by The Modern Data Company as a conversational assistant that lets employees investigate business performance without navigating multiple dashboards or writing database queries. Users can ask why sales declined, compare regional performance and request charts, with answers linked to underlying data definitions and sources. Dex is currently available to existing DataOS customers rather than as an independent subscription.

Render Sandboxes entered public beta, giving software teams isolated virtual computers where AI agents can safely execute code, install dependencies and complete development tasks. Each environment can suspend when idle, helping developers manage infrastructure costs while keeping experimental agent activity separate from production systems. The service is available through Python and TypeScript SDKs or a command-line interface, making it better suited to technical teams than non-technical business users.

A I A G E N T S & M O D E L S

Gemini Omni 1.2 Flash — Google released an updated video-generation model on 5 November, extending generated clips to 15 seconds and improving visual consistency, reference-image adherence and conversational editing. Google also reports a 30% reduction in generation costs, making iterative video production more economical. The model is generally available through the Gemini API and Google Cloud.

GPT-6 Cyber — OpenAI introduced a specialised cybersecurity model on 5 November, designed to identify vulnerabilities, validate potential exploits and support authorised security testing within defined boundaries. It extends AI-assisted vulnerability research while introducing safeguards intended to prevent activity outside approved scopes. Access is initially restricted to selected, verified Daybreak Red customers through the API and Codex.

Inkling 2 Small — Thinking Machines Lab released its second-generation model on 5 November, combining image understanding, adjustable reasoning effort and improvements in coding, research and complex knowledge work. The mixture-of-experts model has 594 billion total parameters, with 17 billion active during inference. It is available through Tinker for experimentation, sampling and fine-tuning; downloadable weights are planned but have not yet been released.

GPT-Image-2.6 Sunburst — OpenAI launched its Images 2.6 generation on 4 November, introducing Sunburst for higher-quality images and greater consistency across editing rounds, alongside the faster Flare variant. The release also supports creating animated GIFs, expanding its usefulness for marketing assets, product imagery and creative production. Both models are available through the API, with Images 2.6 also introduced in Codex.

SWE-2 Flash — Cognition released a smaller coding model on 4 November, developed from Thinking Machines’ Inkling-Small using additional training focused on software engineering. Cognition reports approximately three times the generation speed of its larger SWE-2 model at one-fifth the price, potentially making routine coding, documentation and development automation less expensive. SWE-2 Flash is available in Devin Desktop and CLI, while model-weight distribution remains restricted.

Tencent Hy4 — Tencent released a substantial upgrade to its Hy4 model on 4 November, adding visual understanding and computer-use capabilities not present in its earlier preview. The model can interpret images and charts and interact with command-line and graphical interfaces, extending its usefulness for coding, design and multi-step tasks. It is available through Tencent products and cloud services, although downloadable model weights remain forthcoming.

Sarvam Prism — Sarvam AI released a specialised decision model on 3 November that returns calibrated probabilities for predefined choices instead of generating conventional text responses. Built on Sarvam 105B, it targets business processes such as enquiry classification, workflow routing, moderation and agent evaluation, where consistent structured decisions matter. The model is available through Sarvam’s API and with open weights for customised deployment.

Soniox STT v6 — Soniox introduced its sixth-generation speech-recognition models on 3 November, adding two capabilities specifically designed for conversational AI. Developers can provide the spoken agent’s text as context, while voice-echo filtering helps distinguish the user’s speech from the agent’s own audio during interruptions. Available through real-time and asynchronous APIs, the models support transcription and multilingual applications across more than 60 languages.

Perplexity Teammates — Perplexity launched persistent AI work agents on 5 November that can manage ongoing projects, monitor changes and coordinate tasks across marketing, sales, operations and other business functions. Unlike one-off Computer sessions, Teammates retain project context, collaborate with other agents and can communicate through Slack, with users able to review or redirect their work. Initial access is limited to Perplexity Max subscribers on the web.

Kiro Crew Bots — Kiro released persistent, customisable agents on 5 November that manage continuing responsibilities rather than isolated tasks. Each bot has its own memory, workspace, objectives and configurable permissions, allowing developers to delegate recurring research, code maintenance, monitoring and other workflows while retaining oversight. The capability is generally available within the open-source Kiro Crew development workspace.

Mistral Studio Managed Agents — Mistral made Managed Agents available in public preview on 5 November, enabling developers to deploy tool-using agents that continue running in the background for hours or days. Agents can recover from interrupted execution, retain workflow state and pause for human approval before resuming. The release reduces the infrastructure developers must build themselves to operate longer-running enterprise AI workflows.

Cursor Live App Agents — Cursor introduced a substantial new cloud-agent capability on 5 November that takes applications beyond code generation into deployment. Users can describe an application, and the agent can build it, provision a PostgreSQL database, configure authentication and prepare hosting before producing a shareable link. Security checks and explicit publishing approval remain required, making the capability particularly relevant to teams building internal business tools.

Wealthbox AI Agents — Wealthbox launched autonomous workflow agents on 4 November for financial advisers and wealth-management firms. The release includes 24 prebuilt agents alongside customisable options that can monitor client activity, prepare follow-ups, identify neglected relationships and initiate recurring administrative workflows. Agents operate on schedules or business events but require approval before modifying records, and are included across all Wealthbox subscription plans.

That’s a wrap! See you next Friday for more.