Key Takeaways

  • Software engineers justify up to $100,000 in annual token budgets against salaries of $200,000 to $400,000, creating the primary recurring revenue driver for labs like OpenAI and Anthropic.
  • Non-technical enterprise departments like marketing, legal, and finance remain stuck at seat caps around $200 per month per user because their operational output does not scale with token volume.
  • Scientific AI and automated chip design cannot expand software margins quickly because physical constraints like wet lab synthesis, FDA trials, and hardware fabrication dictate deployment speeds.
  • Frontier AI model providers risk an ARR growth wall once developer adoption reaches saturation and enterprise buyers refuse to pay six-figure token bills for back-office seats.

The High-Yield Developer and the $200 Seat Cap

Wall Street is running the math on AI infrastructure capital expenditures, and the numbers are getting uncomfortable. Frontier AI labs face a hard question: once every software engineer has an AI coding assistant, where does the next wave of annual recurring revenue come from?

Right now, software engineering carries the entire monetization thesis for foundation model companies. John Coogan pointed out the direct math behind this imbalance: “an engineer can easily justify $100,000 token budget a year if they're making 200 300 $400,000 in their comp package.” (play from 13:00) If an AI tool increases an engineer's output by thirty percent, spending a six-figure sum on inference compute creates instant return on investment. The feedback loop is pure bits. You write code, test it in an isolated environment, deploy it to a server, and measure performance in seconds. The software writes software, which consumes more tokens, which feeds lab revenue.

Outside of technical teams, that math collapses. Jordi Hays noted the reality facing commercial teams: “there are a bunch of businesses in the world small and large where they want to be using more AI but they can't really find a way to spend more than a couple hundred bucks a month per person let's say.” (play from 12:04)

Founders and executives cannot justify giving non-technical operators a massive API balance. As Coogan explained: "Right now everyone who has a marketer on their team, everyone who has a an in-house counsel, everyone who has a CFO, they would all love to be able to go to them and say, 'Yeah, your budget for AI is $100,000 a year. Go spend it and let's make sure the ROI is there.' But we're generally confident the AI that the ROI is there right now. People are like, you can have 200 bucks a month."

A lawyer or a marketer can only read, verify, and publish a finite amount of text per day. Their token consumption hits an organic human ceiling. That difference turns enterprise AI into two distinct businesses: a high-margin compute engine for developers, and a cheap SaaS subscription for everyone else.

The Real World Slows Down Pure Compute

If back-office white-collar workers cannot absorb six-figure token allowances, what can? The standard bull case points to deep technology: automated chip design, materials science, and computational drug discovery. The thesis assumes that labs will replace marketing software revenue with breakthroughs in hard sciences.

That pivot runs into a wall of physical reality. Software loops run in milliseconds, but atoms move on their own schedules. Hays highlighted this problem: “My concern with like how that revenue will ramp over time is that like you just can't even with AI, you're not gonna pick I think at this point anything where it still has to tie into the real world.” (play from 15:15)

An AI model can propose ten thousand novel molecular structures before lunch. Testing those compounds still requires a physical wet lab, chemical reagents, and human lab technicians. Even when those stages clear, human clinical trials take years under FDA supervision.

Automated semiconductor layout hits the same barrier. A model can design an optimized microchip architecture overnight. Getting test silicon back from a fab like TSMC takes quarters, costs millions in physical mask sets, and requires cleanroom line time that software cannot compress.

Because physical validation creates months of waiting time between token generation cycles, labs cannot scale scientific API consumption at the velocity public markets demand. The revenue curve stretches out, exactly when debt and capex bills come due.

What to Do With This

Audit your company's AI spending this week across every department. Separate your software engineering token and seat costs from your sales, marketing, and legal tools. If you find non-technical teams sitting on unused $200 per month seats, downgrade them to standard flat tiers and redirect that software budget directly into engineer-led automation workflows that generate measured line items.