SIGNAL / NOISE
Same Meter, Different Seat
Three thousand dollars a month. Call it $36,000 a year, and remember what that price does not include: no payroll tax, no health plan, no PTO, no chair, no manager who needs a Friday one-on-one. That's what a metered AI seat can run now. GitHub Copilot moved everyone to token billing this week and the receipts are ugly: a $19 seat ballooning into a $3,000 bill, one Anthropic user ringing up $150,000 of Claude Code in thirty days. A month ago this stuff felt free. Somebody just turned the meter on.
Ed Zitron published the angry version today, "AI Doesn't Have ROI," and he's half right. The deeper problem was never the return. It's that nobody could measure the cost, because every subscription you've ever touched was subsidized to hide it. But "no ROI" is only the view from one chair. The token seller sees a long-overdue repricing. The over-levered startup sees a funeral. The patient buyer sees a sale. Same event, three seats — and which one you're in is the whole story.
We've seen this movie. Subsidize below cost, let the capital intensity drag you all the way in, then charge what it actually costs. That's the Uber playbook. Two differences. Uber took a decade; AI ran the entire loop — miracle, jobs apocalypse, never mind, basically free, here's your $3,000 bill — in about six months. And Uber never had to kill Lyft. Neither do the labs. They're not fighting each other for AI customers; they're coming for search, SaaS, the consultants, the layers of middle management. Microsoft even trained its new model on McKinsey's own workflows — the consultancy helping build the machine that bills by the outcome it used to bill by the slide.
The meter that priced the machine is the same meter that's about to price you. We pay salaries by default for one reason: we couldn't measure what anyone produced, so we guessed and called it a wage. That excuse just died. Marcelo Lebre — who runs a payroll company, so he'd know — laid it out: once you can stack "forty cents in tokens" next to "four hours of an engineer" a few thousand times, headcount stops being the unit and cost-per-outcome becomes the language. Then the comp math turns cold. If a $3,000 bill does the job, the human doing it had better come in under $36,000, all in.
That's not a layoffs story. It's a pricing story. The meter is just transparency, and transparency — if the labs and Washington let it run — drags any market to its true price and clears it. That's the system working, not the threat. The threat is whoever would rather keep the price hidden: harness lock-in, or a White House handing early access to its friends. Watch who tries to break the meter.
Stop pricing the AI. Start pricing the seat next to it.
At COAI today: the full Signal/Noise — the Rashomon frame, the Uber-compression math, and why the hedge is real but still hacker-only — is live at getcoai.com.
If today's column hit a nerve, that nerve is the work. Outsider Labs is our consulting arm, and the question we run with clients is this: where is your AI spend producing measurable value, and where is it just a meter running with nobody watching. If you can't answer that cleanly, that's the conversation we're built for.
ONE — A NUMBER THAT SUMMARIZES THE DAY
$36,000. That's a $3,000-a-month AI bill annualized — no taxes, no benefits, no PTO, no manager. A month ago that same metered seat cost fifty bucks; this week one Anthropic user ran $150K of Claude Code in thirty days. The subsidy's over and the meter's on. The scary part isn't that AI got expensive. It's that $36K is still a discount on the person it replaces.
THREE — ACTIONS TO TAKE TODAY
Put a meter on your own spend before it puts one on you. Set hard per-seat token caps in every AI tool today, not next quarter. Uber torched its annual token budget in four months; one company spent $500M in a month because nobody set a limit. The default settings are designed to hide the cost from you. Change them this afternoon.
Run the $36K test on one real workflow. Pick a single repeatable task. Price what the AI costs to do it, then price the fully-loaded human cost — salary plus taxes, benefits, and overhead. You're not doing this to fire anyone. You're doing it because your competitor is doing it, and the number tells you where you actually stand.
Price the hedge even if you can't deploy it yet. Open models now run at 5–10% of frontier cost, routing tools claim 25% savings, and unified-memory machines run coding agents for the price of electricity. It's still engineer work, not a Monday plug-in. Get the number anyway — it's your hand the next time a vendor "adjusts" pricing.
FIVE — STORIES TO KEEP YOU INFORMED
Tuesday, June 2
Microsoft builds its own everything so it can stop renting it. (Full analysis above.) At Build, Microsoft shipped seven in-house MAI models, the Scout agent, and an NVIDIA-built stack — what swyx called control "from chip to model to harness." The company that put $18B into OpenAI and Anthropic just unveiled the off-ramp from both.
The meter is coming for payroll, not just tokens. (Full analysis above.) Marcelo Lebre's thread nails the shift: AI is billed per token, humans per hour, and for the first time both sit in the same workflow. Once output is measurable, the salary-by-default era ends. Watch the legal system — labor law is the slowest piece to move.
Trump signs the AI order he walked out on two weeks ago. A 30-day national-security review of frontier models before release, NSA-led, "voluntary." Cato's analyst flagged the obvious risk: the government picks which "trusted partners" get early access, which is a moat handed out by Washington. Lands while Anthropic is mid-fight with the administration.
Roughly half a trillion dollars is trying to get raised at once — and that's just the part we can see. Google ~$80B in stock, SpaceX ~$80B, Anthropic and OpenAI ~$100B each, before the data-center leases. litcapital's "exit liquidity avalanche" hit 800K views. When the richest companies alive borrow to stay at the table, capital, not compute, is the scarce thing.
Anthropic's office vending machine became a chain of AI-run cafés. Andon Labs let an AI agent run a real business with no humans in charge — it hired staff, ran the supply chain, and passed a Swedish labor inspection. The founder's line to sit with: "the future COO won't have many colleagues."
— Harry and Anthony
Sources: Ed Zitron — AI Doesn't Have ROI · @marcelolebre — per-token vs per-hour thread · @garrytan — AI Harness Wars / Factory routing · CNBC — Microsoft MAI models, McKinsey 10x efficiency · Computerworld — Scout on OpenClaw · @litcapital — exit liquidity avalanche · AP/ABC — Trump AI executive order · Fortune — Anthropic's AI-run stores and cafés