In partnership with
Welcome, AI enthusiasts
Anthropic just pulled one of the first engineers off OpenAI's secretive chip program, and the timing is hard to miss. The company is reportedly building toward custom silicon of its own, and it just hired someone who has done exactly that. Let's dive in!
In today’s insights:
Anthropic ARMS UP on OpenAI's chip brain
Coinbase CEO Predicts 99% Cheaper AI Models
An OpenAI Voice Warms to a Coordinated Pause
Read time: 3.5 minutes
LATEST DEVELOPMENTS
AI CHIP
⚙️ Anthropic ARMS UP on OpenAI's chip brain
Evolving AI: One of the first engineers on OpenAI's chip program just left for Anthropic.
Key Points:
Clive Chan, one of OpenAI's first custom chip hires, has joined Anthropic after about 2.4 years.
His old team built the silicon effort behind OpenAI's 10 gigawatt Broadcom deal.
Anthropic currently runs Claude on Google TPUs and Amazon chips, but reportedly began weighing its own custom silicon around April.
Details:
Chan was one of the first hires on OpenAI's chip program, joining in early 2024 from Tesla Autopilot. That program sits behind the Broadcom partnership announced last October, a 10 gigawatt rollout reportedly worth $18 billion, with shipments starting later in 2026. Anthropic has no chip team of its own. Its April push to design custom silicon was still early, without a dedicated group running it.
Why It Matters:
Anthropic has built its lead on model research, and the names it hired all year reflected that. Chan is the first marquee hire pointed at hardware instead, the silicon that runs these models and sets the cost of every query. As both labs edge toward going public, who owns the chips underneath could end up shaping the race as much as the next model.
Stop switching apps. Your browser can do it all.
Every tab you open, every copy-paste into ChatGPT, every lost train of thought — that's your browser failing you. Norton Neo fixes it. Built-in AI works directly inside your session. Hover to preview. Search everything from one bar. VPN and ad blocking included, free.
Evolving AI:Armstrong expects most AI to get far cheaper, with energy as the real limit.
Key Points:
80% of workloads shift to 99% cheaper models within 12 to 18 months, Armstrong predicts.
The top 20% stays on frontier models for scientific work and high-level agents.
Coinbase has held AI costs flat by routing prompts while token usage grows exponentially.
Details:
Armstrong's routing push comes back to one bet, that cheap intelligence is about to be everywhere. He figures demand for it stays near infinite, so cheap models end up handling the bulk of everyday work and the pricey ones get saved for the hard stuff. For the AI most people use day to day, that should keep prices sliding. He likens it to PC buyers, where hardly anyone pays to max out their specs, and with costs dropping faster than Moore's law, he expects the real squeeze to land on power and chips.
Why It Matters:
Inference keeps getting cheaper, close to tenfold in a year, yet the savings get swallowed by heavier power use and the bills climb anyway. Grid connections now take years, and plenty of data centers sit half-built waiting for electricity. The cheap-AI future Armstrong describes still has to be plugged in somewhere.
Voice dictation that doesn't mangle your syntax.
Most dictation tools choke on technical language. Wispr Flow doesn't. It understands code syntax, framework names, and developer jargon — so you can dictate directly into your IDE and send without fixing.
Use it everywhere: Cursor, VS Code, Warp, Slack, Linear, Notion, your browser. Flow sits at the system level, so there's nothing to install per app. Tap and talk.
Developers use Flow to write documentation 4x faster, give coding agents richer context, and respond to Slack without breaking focus. 89% of messages go out with zero edits. Free on Mac, Windows, and iPhone.
Evolving AI: An OpenAI researcher says support for conditional pauses is rising as Anthropic pushes the idea.
Key Points:
roon, an OpenAI staffer known for accelerationist leanings, posted that more people now back mutual conditional pause agreements.
The framework asks rival labs to slow down together, each verifying the others before easing off.
His comment arrives as Anthropic publicly presses the same idea.
Details:
Anthropic laid out the version people are reacting to, a verifiable slowdown that holds only while rival labs ease off in step. The trigger it names is recursive self-improvement, the stage where a model starts improving itself without much human input. Getting every lab to slow on the same terms, and to prove they have, is the part nobody has solved. Support from the optimist side signals the idea is no longer niche.
Why It Matters:
The labs floating this pause talk are the same ones building the tools millions now lean on for work and daily life, so the speed they choose eventually decides what lands in front of everyone else. A real pause would slow that on purpose, and seeing even the racers weigh it suggests the worry about losing control no longer feels far off.
👀 Click on the image you think is real
QUICK HITS
⚙️ Notion Reinstates Anthropic Models After Brief Sunday Outage.
🛰️ Google Books $920M/Month of SpaceX Compute as Gemini Demand Outstrips Plan.
✉️ ChatGPT Becomes Your Inbox From Inside the Chat Window.
☁️AirTrunk Commits $30B to Build 5GW of AI Data Centers by 2030.
📈 Trending AI Tools
🤖 Lindy – The simplest way for businesses to create, manage, and share agents*
⚔️ Edge Arena - AI agents compete and critique to deliver one execution-ready plan.
🔎 Arobis AI - GEO and AEO tool that boosts SaaS visibility in AI search.
🧪 SciFigureAI - AI figure maker for research papers, posters, and grants.
*partner link