The $750k Developer
In the past two months, three independent bodies of research landed within weeks of each other. None of them were coordinated. All of them arrived at the same place.
Goldman Sachs Chief Economist Jan Hatzius told the Atlantic Council that AI investment contributed “basically zero” to U.S. GDP growth in 2025. Not a rounding error. Not measurement lag. Zero.
The National Bureau of Economic Research surveyed nearly 6,000 CEOs and CFOs across the U.S., UK, Germany, and Australia. 80 to 90 percent reported no measurable impact on productivity or employment, despite 70 percent of them actively using AI. These are not skeptics. These are organizations that already bought in, already deployed, and still cannot find the return.
MIT studied $30 to 40 billion in enterprise GenAI investment and found 95 percent of organizations got no return on it.
Apollo’s chief economist summarized it plainly: “AI is everywhere except in the incoming macroeconomic data.”
$250 billion spent. Nothing moved.
This is the part where most people reach for an excuse. The technology is still maturing. The ROI takes time. We’re measuring wrong. But we’ve been saying that for three years and the gap between the narrative and the data keeps widening, not closing.
So where did the money go?
Start with a number that didn’t make it into any of those studies.
A developer running a frontier model through a real agentic coding session reading a large codebase, generating implementations, and iterating on failures burns roughly $3,000 a day in tokens. That’s arithmetic from Anthropic’s published pricing: $5 per million input tokens, $25 per million output tokens. One developer. One day. Verify it yourself.
Multiply by 250 working days: $750,000 a year. Per developer. In tokens. Before salary. Before benefits. Before infrastructure.
Now the conference circuit answer: run more agents. “We’re running 10 agents, 24 hours a day.” I’ve heard this at every event this year. Here’s what that actually costs:
1 developer during business hours would be $3,000 a day or $750,000 annually
10 agents working 24/7 would be $90,000 a day or $32,850,000 annually
$32.8 million a year. Before a single line of rework. Before the AWS bill that nobody has looked at closely yet.
But there’s a cost the spreadsheet doesn’t capture at all.
A developer managing 10 agents isn’t 10x more productive. They’re a single human context-switching across 10 parallel workstreams — 10 PRs, 10 CI pipelines, 10 review queues, all demanding attention at once. Research on cognitive load puts the cost of a single context switch at 23 minutes of recovery time. Ten agents means that recovery time is theoretical. The developer never gets it back.
The math that makes agents look cheap assumes the human in the loop is infinitely scalable. They’re not.
This week we’re going to show you exactly what that costs — in code quality, in organizational dysfunction, in budget categories that don’t exist, and in a historical pattern the industry has been through before.
Tomorrow: what those 10 agents are actually producing.
Sources
Goldman Sachs Chief Economist Jan Hatzius, Atlantic Council, via Gizmodo (February 2026)
NBER Executive Survey (February 2026), via Fortune
MIT GenAI Divide: State of AI in Business 2025, via The Outpost
Gloria Mark, UC Irvine, cost of interrupted work
Anthropic API pricing: $5/MTok input, $25/MTok output (Claude Opus 4.6)
Be Atomic covers the infrastructure, primitives, and ideas behind agent-native software development, written by the team building it. atomic.dev