Second-largest private funding round in history, behind only OpenAI. And that valuation now puts Anthropic ahead of OpenAI as the most valuable AI company in the world, public or private. Quick caveat before I get into it. I'm not a finance guy and I've got no inside line on Anthropic. This is me thinking out loud. Here's what I see happening. Most companies raise pre-seed, seed, then a Series A through D. After that they IPO, get acquired, or run out of road. Anthropic just hit Series H. That on its own doesn't tell you much. Stripe and SpaceX stayed private for years on purpose to dodge public-market scrutiny. Slack and Lyft hit Series H and IPO'd inside a year. Everyone takes a different path. What it does tell you is that Anthropic chose to stay private through eight rounds. Bloomberg reckons an IPO could land as soon as October. Their annual revenue went from $1 billion in December 2024 to $47 billion this month, widely called the fastest revenue ramp of any software company ever. They're now running higher revenue than OpenAI, and the latest projections have them hitting profitability first. I'm not saying anyone's "winning" here, both are still burning billions a year. But the underdog framing for Anthropic is getting harder to defend. The thing I keep coming back to is why a company growing this fast still needs another $65 billion. It comes down to compute. Dario, their CEO, said it himself a few weeks back. They planned for 10x growth in 2026. They got 80x. They genuinely cannot build datacenters fast enough. Earlier this month they leased the entire Colossus 1 datacenter in Memphis from SpaceX. 300 megawatts, over 220,000 Nvidia GPUs, $1.25 billion a month. They didn't pick SpaceX over Amazon or Google for any clever reason. Colossus was the only compute available right now. The rest doesn't come online until 2027. This round also brought chip makers Samsung, SK Hynix and Micron on as investors. Those three make the high-bandwidth memory that sits on every Nvidia GPU. Getting them invested locks in supply at the single most constrained part of the chip stack.

You run /code-review --fix and it applies the changes straight to your working tree. No copy-paste, no "accept this suggestion" loop. It just does it. This saves real time. It's also where a lot of people are about to get burned. Because the job has quietly changed. A year ago you were writing the code. Now you're reviewing it. And the better these tools get, the more tempting it is to stop reviewing and just accept whatever lands. That's the actual risk. Not that the AI is wrong. It's that you stop checking whether it's wrong. There's a line I keep coming back to: you can outsource your thinking, but you can never outsource your understanding. When your name is on the commit, you take the credit when it ships clean. You also own it when it breaks at 2am. So here's the filter I use with any auto-fix, code or content: If it's low stakes and reversible, let it run. Formatting, a rename, a test stub. Who cares. If it touches logic, money, or anything a user sees, it drafts and I review. Every time. The auto-fix gets me to a reviewable diff faster. It doesn't get me out of reading the diff. That's the whole game now. Reading better, not shipping blind. Try it this week. Next time you accept an AI change without reading it, stop and ask whether you actually understand what changed. If you can't explain it, you don't own it yet. What's one thing you let AI run fully hands-off, and one thing you'll never stop reviewing? Curious where people draw the line.

In this video, I show how I run Claude Code from my phone all day through a Telegram bot wired straight into my Claude Code project. Voice notes, transcription, replies in the chat, the bot running 24/7 on my Mac. No VPS, no Mac mini, no monthly fee. Step 1, Get the skill: Drop the Claude Code Telegram Bot skill into your Claude project. Tell Claude "proceed with the Telegram setup" and it walks you through the whole thing. Step 2, Create the Telegram bot: Message BotFather in Telegram, hit new bot, name it, copy the access token. Paste that into Claude. Step 3, Group and topics: Create a Telegram group, add your bot, promote it to admin, enable topics under group settings. The topics step needs a message in the group before Claude can confirm it. Step 4, API keys: You don't need a new Anthropic API key. Tell the setup to use your existing Claude Code login. Only the OpenAI key is required, and only if you want voice notes ($5 deposit, lasts months). Step 5, Make it run 24/7: Tell Claude to set it up to run all the time. It installs a launchd service that keeps the bot online with the lid closed, auto-restarts on crashes, auto-starts on every login. --- Send this prompt to Claude Code: **copy below this line** I'm giving you a skill called Claude Code Telegram Bot. Get the files with: git clone https://github.com/automatorsplus/claude-code-telegram-bot Follow the README to set it up. Recommend how it would best apply to our setup in plain language, and ask me a few questions to clarify my intent before you install it. --- If you want to drive Claude Code from anywhere, this is the setup.

Most Claude Code skills on YouTube are built for a YouTube clip. These are the five I actually keep installed when work is going out to a paying client. Here they are in install order so they stack: 1. Skill Creator → /plugin install skill-creator@claude-plugins-official The factory. Turns natural language into reusable skills. 2. Superpowers → /plugin install superpowers@claude-plugins-official Stops Claude sprinting. Plans, tests, verifies before claiming done. 3. Claude Mem → /plugin marketplace add thedotmack/claude-mem Carries memory across sessions automatically. 4. GSD → npx get-shit-done-cc --claude --global Beats context rot by splitting work across sub-agents. Heads up: uses more tokens. 5. Frontend Design → /plugin install frontend-design@claude-plugins-official Fixes the AI-slop look on UI work. Strongest before-after of the bunch. Install them in this order. Each one fills a hole the next one doesn't try to cover. If you've got a sixth one I should be running, drop it below — always testing.

An easy headline is a big AI guy joins a big AI lab. The actual story is why Anthropic, why now. Look at what he's been publicly building the last few months. His LLM wiki. Auto research loops. Pushing the term context engineering over prompt engineering... Then look at what Claude Code has been shipping. Skills. Sub-agents. Hooks. /goal. Project memory. Same direction. The thing sitting underneath all of it: the model isn't the product anymore. The wrapper around it is. Your data, your context, your workflows. Same model, totally different outcome depending on what's around it. When people say "data is the moat" they usually picture some giant enterprise database. For normal builders, it's way smaller and way more practical. Your meeting notes. Your internal SOPs. Your customer calls. Your style guides. Your naming conventions. The stuff that makes your work actually yours. If Claude can read all of that and use it the right way, the model gets smarter and more useful to you specifically every single week. That's the real lock-in. Not because you can't switch models, you can. But because the longer you build context inside one operating system, the less you want to leave it. So worth thinking about how you're feeding your AI before worrying about which model you're using.