Anthropic's landlord is Elon Musk

Anthropic's Landlord Is Elon Musk

SpaceX filed its IPO prospectus last week. The most interesting line item wasn't satellites or Starships. It was Anthropic. The S-1 disclosed that Anthropic is paying SpaceX $1.25 billion every month through May 2029, close to $45 billion total once you account for discounted rates during the early ramp. In exchange, Anthropic gets access to more than 220,000 NVIDIA GPUs and 300+ megawatts of capacity at Colossus, the data center campus SpaceX is leasing alongside xAI. Either side can walk with 90 days notice.

Now hold that in your head, and add this: Anthropic is also reportedly weighing a $40-50 billion funding round at a $900 billion valuation, with a potential October IPO on the calendar. OpenAI filed its own confidential S-1 a few days later, targeting September at $852 billion to a trillion. And in the same seven-day window, OpenAI co-founder Andrej Karpathy joined Anthropic's pre-training team (the people who train the model from scratch), specifically to help launch a sub-team that uses Claude to accelerate the next round of Claude's pre-training. Two of the three biggest AI companies are about to be public, combined valuations approaching $2 trillion, and one of them is wiring $15 billion a year to a data center colocated with a direct competitor while collecting the talent that builds the talent.

Our read: at this point in the cycle, the durable moat in frontier AI looks more like a power-and-real-estate business than a software one. The companies that compound for the next decade are the ones that own the megawatts and the chips, or have a generational lease on them. Anthropic's revenue is past a $30 billion run-rate and reportedly on track to land its first profitable quarter, coming in roughly $559 million ahead of what its own board projected, which is incredible, and which only works because they have enough capacity to serve the demand. The economics of being a tenant are fine right up until the landlord raises the rent.

Here's who should be uncomfortable. The smaller AI company whose whole pitch deck still leans on "we have proprietary GPU access." That sentence used to mean something. After last week, it means you signed a 90-day lease with a counterparty who just told the public market it has more capacity to sell. And the enterprise CIO whose AI risk register is mostly about which model to standardize on: that's the wrong worry. Whoever you pick is also renting, and their landlord can change the deal.

The other half of this story is the bottom of the market. Cheap open models keep getting close enough to frontier reasoning for most production workloads, which means the pricing power the IPO math depends on is already eroding from below. Anthropic and OpenAI pay $15 billion a year for the chips that justify the high-end pricing, while a startup down the street can run a near-frontier model for pennies per call. The durable AI businesses of this decade are going to be the ones that own the workflow the agents actually live inside, not the ones training the next round of foundation models.

What an Agent Looks Like When It Stops Chatting

Eric Zakariasson at Cursor posted last week that the most-used skill inside Cursor right now is something called /thermo-nuclear-code-quality-review. What it does: deletes complexity instead of moving it, blocks files over 1,000 lines, and rejects pull requests that technically work but leave the codebase messier than they found it. 451,000 views in three days. The replies are mostly other engineers asking how to copy it.

Why this matters even if you don't write code: this is what an agent looks like when it stops chatting and starts working a shift. Three things shipped in the last six weeks that together change what "using an agent" means: a way to set a destination, a way to check work along the way, and a way to remember what the codebase actually looks like. None of them is its own product. They're the layers underneath one.

• /Goal: Codex CLI shipped it first in late April; Claude Code followed May 12. You set a completion condition, the main model works, and a separate validator model checks "is the goal met?" after every turn. If not, it keeps going. Hours, sometimes overnight. Stops when the work is actually done.
• Skills: Eric's /thermo-nuclear-code-quality-review is the most-shared example, but every coding team is now building their own. The skill is the agent's standing orders. "Reject anything that ships with TODOs." "Block PRs over a thousand lines." "Don't move complexity, delete it."
• Knowledge graphs: Safi Shamsi shipped Graphify in April. Reportedly 30,000 GitHub stars and 250,000 PyPI downloads in roughly a month. Turns your codebase, docs, PDFs, and diagrams into a structured map the agent can ask questions of, instead of digging through files one at a time. Free, MIT-licensed.

Stack them and the agent isn't waiting on you between turns. It's working toward a stated outcome, honoring rules you set, and looking at a structured map of your codebase instead of a wall of text.

1. Pick a goal worth a day, not a turn: A goal is "refactor the auth module so it stops leaking session state" or, if you're not on the engineering team, "audit our top 50 customer emails for compliance issues and draft fixes." A task is "rename this variable." The /goal command is built for the first kind.
2. Set /goal in Codex or Claude Code: Type /goal followed by your completion condition. Both CLIs ship it now; the syntax is nearly identical. The validator runs after every step, so you don't need to babysit.
3. Wire in a quality skill: Steal Eric's idea or write your own. Something that vetoes "works, but messier than before" code. Two paragraphs of plain-English rules is enough to start.
4. Drop in a knowledge graph: Install Graphify and let it index your repo overnight. The agent stops thrashing through grep and starts answering questions like "what calls this function, and what breaks if I change it."
5. Run it overnight. Audit in the morning. Don't watch: The point of the supervisor is that you don't have to. Look at the diff over coffee, read the goal validator's summary, decide what to merge.

The lucky part: Codex has been making the most aggressive push at Claude Code we've seen since either tool existed, and on the hard, long-running engineering work, the operators we trust most are currently putting Codex slightly ahead. Both labs are subsidizing the credits hard to fight for you. No need to pick a side this week: use whichever CLI your team prefers, learn the pattern, switch later if the lead changes. The skills and the knowledge graph travel with you.

Try it: Codex /goal ships on Codex CLI 0.128.0 and later. Claude Code /goal ships on 2.1.139 and later. Graphify is free and open source. Try the whole stack on one refactor nobody on your team wants to touch.

Also worth your time

Leaked Zuckerberg audio drops on the same day as 8,000 Meta layoffs

In a leaked all-hands recording, Mark Zuckerberg explained that Meta has been tracking employees' Gmail, GChat, VS Code sessions, mouse movements, and periodic screenshots to train internal AI through a program called the Model Capability Initiative. The audio surfaced publicly the same day roughly 8,000 employees got layoff emails. More than 1,000 staff have signed a petition to halt the program. Even if every word of the policy is legal, this is the case study every employment lawyer will be citing for the next five years.

SAP shipped the Autonomous Enterprise at Sapphire

50+ Joule Assistants running on top of 200+ specialized agents across finance, procurement, supply chain, HR, and CX, with Claude doing the reasoning. The pitch is that your ERP starts running itself. Whether or not the demo holds up in production, every CFO in the mid-market is about to walk into a sales meeting that opens with this slide.

Perplexity open-sourced the first real security tool for AI coding workflows

Bumblebee runs quietly on a developer's laptop and scans for sneaky packages, suspicious browser extensions, and AI-tool configs that could be leaking access. Covers Claude Code, Codex, Cursor. An independent researcher audited the code and confirmed it's clean. For two years AI coding tools shipped with zero security around them; Perplexity just shipped one. Free.

Every US frontier AI lab is now under voluntary federal pre-deployment review

Google DeepMind, Microsoft, and xAI signed agreements with the Commerce Department's CAISI for pre-release model evaluations, joining OpenAI and Anthropic. That's all five major American labs. CAISI has now completed 40+ evaluations, including on models that haven't shipped publicly.

Pennsylvania sued an AI chatbot for practicing medicine without a license

A state investigator chatted with a Character.AI persona called Emilie, who claimed to be a Pennsylvania-licensed psychiatrist and produced a fabricated license number on the spot. Governor Shapiro filed for an injunction, the first of its kind from a US governor. The "AI made it up and the user got hurt" lawsuit category just opened officially.

See you next Tuesday. Sven