Anthropic’s IPO roadshow reveals its chip strategy; half a billion dollars in training data flows to Chinese labs

Compact Conversations for 2026-11-11: 5 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: Anthropic’s IPO roadshow reveals multi-chip Claude strategy

Anthropic’s IPO roadshow video includes previously unreleased details on how the company trains and serves Claude across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs. Co-founder Tom Brown describes a multi-chip strategy and says a Claude model got Claude running on AMD GPUs in three days with almost no human intervention.

Why it matters: Anthropic is disclosing how it runs Claude across four major chip families as it prepares for a public offering.

Source: Simon Willison’s Weblog

Number to Know: Half a billion dollars in AI training data sold to Chinese labs

A ChinaTalk investigation reported by The Decoder finds US data brokers including Mercor, Surge AI, AfterQuery, and Turing are selling roughly half a billion dollars worth of AI training data and reinforcement learning environments to Chinese AI labs. At least nine firms are involved, and export controls barely apply.

Why it matters: At least nine US firms are selling about $500 million in training data to Chinese AI labs, including RL environments previously built for OpenAI and Anthropic.

Source: The Decoder

The Feed

Claude Meetings joins Google Meet calls in beta

Claude Meetings is rolling out in beta on Team and Enterprise plans. Claude can join Google Meet calls as a visible participant to take notes, answer questions, and build artifacts like docs, whiteboards, and slides during the call.

Why it matters: Claude Meetings is available in beta on Claude Team and Enterprise plans, including Claude Desktop.

Source: Claude Release Notes

United Airlines moves to block AI agents over security concerns

Reuters reports United Airlines is moving to block AI agents from its platform over security concerns.

Why it matters: United Airlines is taking steps to block AI agents from its platform because of security concerns.

Source: Reuters - AI

Radical Numerics launches AI pathogen detection model

Startup Radical Numerics announced Omnii Defense, a model that says it can identify natural and AI-engineered pathogens without an internet connection. It runs on a laptop with a USB gene sequencer and assigns sequences a threat score. Access is limited to vetted testers while national labs evaluate it.

Why it matters: Radical Numerics says Omnii Defense can flag natural and AI-engineered pathogens offline using a laptop and USB gene sequencer.

Source: Wired

One Thing to Try

Smitline is an open-source MCP tool that lets Claude Code make actual phone calls. In a demo, Claude Code ordered dinner by phoning a restaurant, introduced itself as AI, noted the call was recorded, and completed the call in 89 seconds for 11 cents. It is a quick add if you already run Claude Code with MCP servers.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host A: [conversational] Anthropic kicked off its IPO roadshow yesterday, and the video it released for investors includes some previously unreleased detail about how the company actually runs its business. Simon Willison flagged it on his blog, and even went to the trouble of transcribing the video himself since there’s no official transcript. Host B: The most interesting section is co-founder Tom Brown describing what Anthropic calls its approach to compute. They train and serve Claude across all four major chip families: AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs. Brown says that’s only possible because of an early bet on a multi-chip strategy, when many people thought running AI models on different chips wasn’t worth the effort. Host A: [thoughtful] There’s one detail in there that’s getting a lot of attention. This year, when Anthropic brought AMD GPUs on for the first time, they had a Claude model try it. It got Claude running on AMD in three days with basically no human intervention, writing almost all of the kernels itself. Brown also describes compute multipliers, software efficiency gains that let the same hardware train better models, and says the company expects Claude to eventually discover new multipliers on its own.

Host B: [with emphasis] Today’s number is half a billion dollars. That’s roughly the worth of AI training data that American data brokers are selling to Chinese AI labs, according to a ChinaTalk investigation reported by The Decoder. Firms like Mercor, Surge AI, AfterQuery, and Turing are involved, at least nine companies in all. Some of it is reinforcement learning environments, simulated training worlds for tasks like coding and financial analysis, and Chinese buyers are getting the same products previously built for OpenAI and Anthropic. The report notes export controls barely apply here, and the authors argue dual-use areas like cybersecurity and biology should be brought under existing rules.

Host A: From Anthropic’s own release notes, Claude Meetings is rolling out in beta on Team and Enterprise plans. You can add Claude to a Google Meet call as a visible participant that takes notes, answers questions, and builds artifacts like docs, whiteboards, and slides during the call. [with a small lift] It works in the Claude apps on desktop too. Host B: It’s also worth noting their November release notes included new Enterprise security features earlier this month, like a security center and inference hooks, for teams tracking this space. Host A: [quick aside] In other news, Reuters reports United Airlines is moving to block AI agents from its platform over security concerns. Host B: [skeptical] That headline is about all that’s confirmed so far, but an airline explicitly pushing back on agentic access is notable on its own. Host A: And finally, Wired profiles a startup called Radical Numerics, which announced a model called Omnii Defense that it says can flag natural and AI-engineered pathogens. It runs on a laptop with a USB gene sequencer, no internet needed, and assigns sequences a threat score. The company built it with cofounders from Stanford’s Evo DNA model work, and access is currently limited to vetted testers while national labs evaluate it.

Host B: [conversational] One Thing to Try today comes from a project that went viral on Reddit. A developer shared Smitline, an open-source MCP tool that lets Claude Code make actual phone calls. The demo video is real: they asked Claude Code to order dinner, and it phoned a nearby restaurant. The agent introduces itself as an AI upfront, says the call is recorded, and the whole call took 89 seconds at a cost of 11 cents. It grew out of a hackathon project that won in Montreal. [lighter] The setup is an MCP server wired into Claude Code, so if you’ve got that running, it’s a quick add. Even if you don’t want your AI phoning restaurants, it’s a useful reminder that MCP tools aren’t limited to text and code, they can reach the phone system too. Link’s in the show notes.

Host A: That’s Compact Conversations for Wednesday. More AI news tomorrow. Until then, happy prompting.