Coding is the first place AI has found product-market fit.

For the U.S. workforce: computer and mathematical jobs are just 3.4% of workers.

But in Anthropic’s February 2025 data, those jobs account for 37.2% of Claude conversations—almost 40%.

That’s about 11× their share of the workforce.

That was all before winter 2025, Opus 4.6, and long-running agents!
Before the job of every coder changed drastically.

Monthly spend per employee has 10x’ed in the last 18 months. Coders now write only a small percentage of their code.

A 10-person human team would’ve taken 18 months and $3.6M.

DHH’s one-person team took 3 months and $69,000, which is 50x cheaper.

We can then break down the software factory pipeline into its parts, like they did in this study of 484 developers.

There’s planning, design, coding, QA, and other. This chart shows two categories of companies: modest AI and AI-pilled.

The modest AI company will save 50% of its coding time, but won’t save time in the other parts. This only causes a 1.2x productivity improvement.

Meanwhile, the AI-pilled company will completely automate design, coding, and security. They’ll spend a bit of time on planning and other.

This causes a 10x productivity improvement, not just 1.2x. It’s essentially the DHH Omarchy example above: 1 person + AI instead of 10 people.

And here was the impact on farmers. We became tractor-pilled.

And became horse un-pilled.

First, you can see the historical trend from 1980–2025. Before agents! We’ve gone from spending $300B/yr to $3.5T/yr.

Although we’ve 10x’ed spend, we’ve 70x’ed lines of code (LOC) from 7B a year to 500B a year.

In other words, we’ve gotten 7x more productive. 10x spend, 70x output.

(And yes, I know, LOC is a terrible metric.)

The right part of this chart shows the 2040 scenario with AI agents.

Given the software pipeline above, we can imagine productivity to increase anywhere from 1.2x to 10x. This chart assumes a 2x increase.

To model correctly, we need to assume a number for demand elasticity.

If coding becomes 50% cheaper, will we still want to spend the same amount on coding? In other words, how much code does the world want?

This model assumes an ε of 1.

In this scenario, we’d still spend $3.5T on code, but we’d spend 50% on agents, and produce a total of 1T LOC a year (2x).

I think this model is too conservative. I can imagine us spending $10T on code, producing 10T LOC a year, and actually increasing the number of coders.

We’ve been on a massive upward trajectory for 40 years already. I don’t see it flattening out.

This is the trillion-dollar AI agent opportunity.

(Compared to the overall SaaS market.)

The other way to think about the market for code is by charting it like other Jevons-y experience curves.

We can see the solar market on this chart: as production rises, costs fall. Historically, each doubling of installed solar capacity decreased the price per watt by about 20%.

We can make a similar curve for cars.

Or code.

Over the course of a decade or two, we've increased production 100x and decreased cost 10x.

You can also compare AI to the cost curves of other tech.

The price of AI thought is becoming cheaper, faster.

The speed of the AI decline is most a function of R&D investment.

And here’s AI inference.

One weird thing is that we haven’t seen AI in the productivity statistics yet.

As the Stripe Econ team shows here, there’s no correlation between productivity growth and AI adoption.

Information and retail trade have the same productivity growth, but a 3x difference in AI adoption.

This is the Solow paradox: “you can see the computer age everywhere except in the productivity statistics.”

Since WWII, overall TFP has been roughly constant, if slightly decreasing.

We’ve gotten TFP from electricity, cars, and computers.

But nothing for AI yet. We’ll see how much of AI productivity is captured in GDP.

Some goods, like electricity and search, became available at extremely low prices, creating enormous consumer surplus.

Other goods, like steam power and automation, were mass-produced at higher prices, creating more producer surplus. Those market sales show up in GDP.

Overall, I think AI is more like search than steam power. It won't show up in GDP.

Coding is one piece of white-collar labor. What about the rest?

Other big white-collar jobs are managers, finance, and lawyers.

Here are AI revenues across knowledge work. Coding is at $29B, while others are at $1B.

But perhaps the most important white-collar market is just admin work.

AI personal assistants like Instinct and Muse will be the second big consumer app after ChatGPT in 2022.

The internet made marketplaces. AIs will soon be their main users.

Instinct has over $1B in transaction volume.

That’s 100x less than eBay. But it’s only been 100 days.

Agents still make 100,000x fewer transactions than people. But that’ll flip soon enough.

Interestingly, acting as a personal assistant requires human social intelligence that’s a bit alien to AI. In the end though, social interactions are just another learnable vector in trajectory space.

The blue-collar labor market in America is roughly half the size of the white-collar market.

The top blue-collar jobs by wage pool are care aides, retail salespeople, freight movers, fast-food workers, truck drivers, stockers, and cashiers.

Let's look at driving.

Trucking and driverless cars are one of the holy grails in robotics. We’ve made surprising progress on this. Waymo now serves 20% of rides in the SF Bay Area, up from 0% only three years ago.

How?

Over the last decade, companies like Waymo got millions of road miles and a billion simulation miles before even hitting the market.

Part of it is just more data. GEN-1.5 was trained on ~650k hours of human data. (That’s a lot! 650 people each working for 100 days.)

It’s roughly 10x more than human children and similar to Waymo’s driving hours.

Interestingly, Gen-1.5 didn’t use any simulation hours, while Waymo spent 600M hours driving in simulation before deployment.

Real-world success rates...

...are increasingly correlated with simulation success rates.