AI is beginning to shape both the research process and the infrastructure that supports it. Anthropic’s use of Claude in its own R&D, Crusoe’s funding and higher GPU-cloud prices show different sides of the effort to scale these systems.

The edition also follows conversational advertising and the software supply chain forming around agents, two areas where new interfaces bring new operational questions.

In this edition — jump to a story

🤖 Claude is becoming part of Anthropic’s R&D loop

Anthropic says Claude now leads 26% of its model research and development work under human supervision, and participates in roughly 90% of that work in some capacity. The company also reported that about 30,000 agents were doing research and engineering work on its most-used internal platform as of August.

Why it matters: A frontier-model lab is now using its own models at scale to help produce the next generation of those models. That creates a consequential feedback loop between model capability and research productivity — and makes the quality of evaluation, attribution of agent work, and human oversight increasingly important to understanding how quickly AI R&D itself can accelerate.

☁️ Crusoe raises $3.9 billion as AI infrastructure attracts megadeals

Crusoe announced the initial close of a $3.9 billion Series F at a $30.9 billion post-money valuation. The company plans to expand its vertically integrated platform, spanning energy, data-centre capacity, and cloud services, including both large campuses and modular AI-factory units.

Why it matters: The scale of the round shows how capital-intensive the new generation of AI-native infrastructure companies has become. Crusoe is not only buying GPUs or selling cloud instances: it is trying to integrate energy, physical data-centre capacity and cloud services into one business. Building a startup at that layer increasingly resembles financing industrial infrastructure as much as scaling a conventional software company.

🛡️ GPU cloud pricing is still moving upward

Nebius has notified customers of higher on-demand rates for selected GPU, CPU, and memory resources from 1 October. It is the company’s second pricing adjustment in a short period, showing that the economics of on-demand accelerated compute remain unsettled even as providers continue to add capacity.

Why it matters: The price of AI workloads is shaped by more than model efficiency. Hardware generation, availability, contract terms, and the choice between reserved and on-demand capacity are becoming both architecture and FinOps decisions. For teams operating sustained AI workloads, capacity planning can matter as much as the headline hourly GPU price.

🚀 OpenAI is testing advertising that continues as a conversation

OpenAI introduced new advertising features for ChatGPT, including tests of Sponsored Agents. After selecting an ad, a person can choose to start a clearly labelled conversation with a business-sponsored agent; OpenAI is also adding AI creative tools and integrations with HubSpot and Shopify.

Why it matters: This could change the advertising funnel from impression → click → landing page into impression → sponsored agent → conversation → conversion. If that model works, attribution, creative, product information and disclosure all change with it: the commercial message is no longer just an ad that sends the user elsewhere, but an interactive system participating directly in the buying journey.

📣 AI agents are creating a new software supply chain

AI agents increasingly depend on skills, plugins, MCP servers and external services that can change what an agent is able to access and do. AIR, which emerged from stealth with $50 million, is building around that problem: discovering agents and continuously evaluating the components and tools connected to them.

Why it matters: This extends software supply-chain security beyond packages shipped with an application and into dependencies an agent can discover or invoke at runtime. Organisations will need a dynamic inventory of agent capabilities, provenance and permission controls, change monitoring, and visibility into the external tools autonomous workflows can call — a security model closer to a continuously changing supply chain than a static SBOM.

Sources