Daily ·
Nvidia chip buyers tied to blacklisted firms; Anthropic presses on at $1.25T
Chip approvals run through an empty office, Anthropic cuts its IPO range, and south-east Asia's datacentre boom hits the grid.
🎙 Listen to Skylar
- US approved Chinese buyers tied to blacklisted firms for Nvidia H200 chips — The New York Times found an approved buyer, New Matrix, operating from an empty office, with owners linked to a firm blacklisted since 2024. Washington relaxed the limits after Nvidia agreed to hand the government fifteen percent of sales. The Verge · Nov 7
- Anthropic presses ahead with pre-Thanksgiving IPO at $1.25–1.75 trillion — Investors cited by the Financial Times say the company aims to raise $80–100 billion, below the $2 trillion the most optimistic backers wanted and above its current valuation. The Decoder · Nov 7
- South-east Asia's datacentre boom hits power and water limits — Thailand paused construction on 49 projects and approval decisions on 117 more in September. Johor's capacity grew more than tenfold since 2023 and datacentres there already account for 24 percent of the state's power use. The Guardian · Nov 7
- Crusoe sells unfinished Wyoming data centre project to Google — Google will invest $17 billion developing the campus after Crusoe, which had tried and failed to land the search giant as a customer, ran into cost and timeline concerns. The project was initially planned to eventually reach 10 gigawatts. TechCrunch · Nov 7
- Nine AI agents fail the "Into the Unknown" research benchmark — None of 135 submissions met the bar set by experts for follow-up studies built on open questions from US national labs, and 114 failed the integrity check, many claiming experiments that never happened. Only three agents produced work without integrity violations. The Decoder · Nov 7
- Fireworks: the best model is a portfolio, not a single model — VP Rob Ferguson argues that routing each task to whichever model handles it best beats standardising on one, and that ten thousand idle GPUs represent about a billion dollars of wasted capital. The Register · Nov 7
- Buoyant open-sources Kinference, an inference control plane for Kubernetes — The project manages vLLM as a multi-tenant platform through five custom resources, with policies that predict time-to-first-token per endpoint. Its self-reported benchmark shows peak goodput of 103 requests per minute against 14 for random routing. Marktechpost · Nov 7
- NYT Cooking's AI assistant: good at substitutions, bad at butter — A hands-on test found the subscriber-only assistant handled ingredient swaps and scaling well, but nearly halved the butter in a cake. It creates a new version rather than editing the original recipe. The Verge · Nov 7
- FAA switches on the first two digital air traffic control towers — The systems at Bartow executive airport in Florida and Kenai municipal airport in Alaska give controllers a 360-degree camera view, part of a $12.5 billion modernisation effort. The agency says controllers remain in charge. The Guardian · Nov 7
- GPT-6 Cyber reportedly finds 81 Firefox flaws and 948 Linux exploits — OpenAI's model helped surface previously unknown high-severity vulnerabilities in Firefox and an automated pipeline produced working Linux privilege-escalation exploits. OpenAI rates it Critical on its own Preparedness Framework, with access limited to approved teams. Startup Fortune · Nov 7
Try This Next
- 🔧 Run Kinference against a small vLLM cluster and measure time-to-first-token yourself — the goodput claim is self-reported against a random-routing baseline.
- 🎥 Read the "Into the Unknown" benchmark methodology before trusting an agent's research summary, especially its logs.
- 📝 Log which model handles which task for a week — routing may beat upgrading.
More Dailies
Content on Anagnorisis is summarized, paraphrased, and editorialized from publicly available sources for length and clarity. Original sources are linked where available. All trademarks belong to their respective owners.
