Another Wednesday, another selection of AI news and resources to help you become more AI native. This week:
Anthropic blows past OpenAI on valuation and revenue, and Elon Musk is its landlord
Microsoft puts computer-using agents into general availability
Meta trained internal AI on staff data, then laid 8,000 of them off
The Codex /goal stack that turns agents into overnight shift workers
Next Gen
Frontier AI Is a Real Estate Business Now.
Whats happening: SpaceX's IPO filing this month disclosed Anthropic is paying it $1.25 billion every month through 2029, close to $45 billion total, in exchange for 220,000 GPUs and 300+ megawatts at the Colossus campus it co-tenants with xAI. Goldman Sachs lifted its 2026 AI infrastructure forecast to $800 billion with capex growth at 7.8%. Anthropic and OpenAI have both filed S-1s pointing at IPOs that combine to roughly $2 trillion in market value.
Why it matters: The durable moat in frontier AI is starting to look more like a power-and-real-estate business than a software one. Even the labs are tenants, and the landlord can change the rent with 90 days notice. Which means any company whose differentiation rests on "we use Claude" or "we use GPT" is two layers down from the actual scarce resource. The competitive question in 2026 stopped being which model you picked. It's which workflow you own.
In the wild: a16z published an essay this week called Avoiding Death on the Yellow Brick Road making exactly this case from the application layer. Their portfolio company 11x has quadrupled positive email reply rates over a few months while running on a stack the labs could in principle clone, because workflow knowledge, domain data, and per-customer guardrails compound in production in a way model upgrades alone cannot. The model is fungible. The system of work is not.
Looking ahead: Write down which of your AI features would still matter if every customer could swap your model with a config change tomorrow. Anything not on that list is rented from a vendor that is itself renting from Elon. The move this week, whether you build agents or buy them, is to invest in the data, governance, and workflow surface that compounds inside your customer's account, and treat the underlying model as the most replaceable thing in your stack.
Need help finding where your company is bleeding money?
I can use AI to save your company money and build a plan to get it back. If I can’t find any meaningful savings, you don’t pay.
AI First
The Codex /goal Stack
Codex and Claude Code both shipped /goal, the command that turns an agent from a chat partner into something that works an overnight shift.
It's worth knowing about this week because the three pieces that make /goal actually useful, goal mode, skills, and a knowledge graph, all hit production in the last six weeks and stack cleanly on top of each other.
Pick a goal worth a day, not a turn. "Refactor the auth module so it stops leaking session state," not "rename this variable."
Set /goal in Codex CLI 0.128.0 or Claude Code 2.1.139 with your completion condition. A validator model checks the work after every step.
Wire in a quality skill. Two paragraphs of plain-English rules is enough to veto messy output before you ever see it.
Drop Graphify into your repo. The agent reads a structured map instead of grepping a wall of files. Free, MIT-licensed.
Run it overnight. Read the diff and the validator's summary over coffee.
Pro tip: Borrow Eric Zakariasson's /thermo-nuclear-code-quality-review skill from Cursor. It deletes complexity instead of moving it, blocks files over 1,000 lines, and rejects PRs that technically pass but leave the codebase worse than they found it.
This is the first credible answer to "what does it actually mean to delegate to an agent?" Aimed at any team that ships code, but the pattern (goal, rules, structured context, audit) generalizes to any long-running agent work.
AI News
Anthropic became the world's most valuable AI startup, and its landlord is Elon Musk.
SpaceX's IPO prospectus disclosed Anthropic is paying SpaceX $1.25 billion every month through May 2029, roughly $45 billion total, for 220,000 NVIDIA GPUs and 300+ megawatts at Colossus, the data center campus it shares with xAI. Separately, Anthropic is closing a round above $30 billion at a pre-money valuation north of $900 billion, lapping OpenAI's $852 billion for the first time. Revenue grew 5x in five months to roughly a $45 billion annualized run-rate while OpenAI lost at least $7 billion last quarter. Andrej Karpathy quietly joined Anthropic's pre-training team in the same week.
LINK
Microsoft pushed computer-using agents to general availability.
The biggest Copilot Studio update of the year took computer-using agents to GA across every commercial Power Platform region. The agents drive software the way a person does, reading the screen, clicking buttons, and typing into fields, so they can finally automate legacy internal tools that never got an API. Microsoft bundled in sub-500ms voice latency, a redesigned workflow canvas with conditional branching, and full session logging to Microsoft Purview and Dataverse. This is the moment "computer use" stops being a research demo and becomes a production feature any team can ship.
LINK
Meta trained internal AI on employee data, then laid off 8,000 people.
A leaked Zuckerberg all-hands revealed Meta has been tracking employee Gmail, GChat, VS Code sessions, mouse movements, and periodic screenshots through a program called the Model Capability Initiative. The recording surfaced 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 line of the policy is legal, this is the case study every employment lawyer will cite for the next five years.
LINK
Reply to this email if you have any questions or further content you want covered.
Talk soon,
Cam
P.S: Try out the money bleed calculators here.