1. The Wrong Debate
The debate about AI and programming jobs has been running for two years, and it's stuck.
One camp says AI will replace programmers. Dario Amodei predicts 50% of white-collar jobs gone within five years. The New York Times runs headlines about coding students seeking work at Chipotle. Every month another CEO announces a hiring freeze.
The other camp says AI will augment programmers. Jensen Huang predicts the number of coders will grow from 30 million to a billion. Every carpenter will become a coder. AI is a power tool, not a replacement.
Both sides share a hidden assumption: the unit of analysis is the existing tech industry. The replacement camp asks whether Google will need fewer engineers. The augmentation camp asks whether Google's engineers will become more productive.
Neither asks what happens at the millions of companies outside the software industry that currently employ zero developers.
In the United States alone, to put some numbers on it: 643,000 retail businesses employ nine thousand software developers between them. That's one developer per 71 companies. Construction has one per 190. Real estate has one per 75.
These are distribution firms, retail chains, contractors, property managers. They have never employed programmers — not because they don't need software, but because software was too expensive to build themselves. When they needed something bespoke, they didn't hire developers. They outsourced the work to a local software agency.
2. The Collapse
The typical local software agency is not a cutting-edge company. It's a fifteen-person shop building business applications with the same tools it used ten years ago. Its competitive advantage has never been brilliance. It has been that its clients don't know how to code, and it does.
The agency sells a relative advantage — not excellence, but access to a scarce skill. It wasn't charging $200,000 because its software was world-class. It was charging $200,000 because the alternative — hiring an in-house team — would have cost half a million a year, and the client wasn't in the software business.
Around December 2025, that skill stopped being scarce. AI coding tools crossed a threshold — not the threshold that makes programmers obsolete, but the threshold that makes one competent developer, armed with AI, more productive than a mediocre agency of five.
The agency's economic niche — built on the fact that coding was hard for everyone else, not on the fact that it was done exceptionally well — collapsed. And most agencies are mediocre.
But AI didn't just kill the agency's business model. It reset the playing field entirely. The software agency with fifteen years of process — estimation templates, spec documents, QA checklists — now has more to unlearn than the distribution company that's never written a line of code.
The agency opens Claude Code and fights the urge to do it the old way: write the spec first, estimate the timeline, assign the tickets. The distribution manager just describes their warehouse and asks the AI what it can do. The incumbent's institutional experience, built up over years of doing things a certain way, becomes an anchor.
In a reset of competitive advantage, the domain expert — the person who actually knows what the business needs — starts ahead.
3. Free Cognition
But what gives the domain expert the confidence to start at all?
The engine underneath everything here is a change in the economics of cognition itself. In January 2025, the Chinese AI lab DeepSeek demonstrated that powerful reasoning could be delivered at fractions of a cent per token — a fraction of what competitors were charging.
The immediate reaction was to invoke Jevons Paradox, the 19th-century observation that more efficient coal engines led to more coal consumption, not less. Lower AI costs, the argument went, would simply mean more AI usage.
DeepSeek was deeper than Jevons. When the cost of reasoning approaches zero — when tokens become, in practical terms, too cheap to meter — the economics of cognitive work don't just shift. They break and re-form.
The question for a business flips from "can we afford to build this?" to "is there any reason not to try?" The cost of trying — of describing your warehouse to an AI and seeing what it produces — rounds to zero.
And so the domain expert tries. Invoicing automation. A customer report card. A logistics dashboard. They discover what's possible. And they discover what they can't do themselves.
That's when they hire someone who can.
4. The Migration
The person they hire is likely a senior developer from exactly the kind of agency whose business model just collapsed.
This developer already knows CI/CD, architecture, testing, and deployment — the institutional knowledge that used to live inside the software house. AI now multiplies their output to what previously required a team of five.
A non-software company hires one or two of these people, and the organizational scaffolding walks in the door with them. Code review, deployment pipelines, security practices — the hiring company doesn't need to build any of it. The developer is the carrier of best practices, not the organization they came from.
The developer doesn't need the software house's organization. The software house needed the developer. The client gets better software, built faster, for less money — built by someone who now understands the business because they're inside it.
And the developer? They go from being one cog in a fifteen-person agency to being the person who owns the entire software capability of a company. That's not a demotion. It's a promotion.
5. The Evidence
Jensen Huang has been saying something adjacent to this for years. He likes to tell a story about radiologists.
Years ago, AI researchers predicted that radiologists would be the first profession eliminated by AI. Computer vision achieved superhuman performance on scan reading by 2019. Radiologists were supposed to be obsolete.
Instead, their numbers grew. We now have a shortage. Why? Because, as Jensen puts it: "The purpose of a radiologist and the task of a radiologist reading scans are related, not the same." AI automated the task — reading scans — and amplified the purpose — diagnosing disease. More scans read meant more diagnoses, more patients treated, more radiologists needed.
Jensen applies the same logic to programmers. "The number of software engineers at NVIDIA is going to grow, not decline." He predicts the world will go from 30 million coders to a billion. "Every carpenter in the future will be a coder, except a carpenter with AI is also an architect."
He is right about the expansion. AI will make software creation so accessible that millions — possibly billions — of people will do some form of it. But Jensen never asks where they will work. He never asks whether the companies that employ those billion coders will look anything like the companies that employ today's 30 million.
That's the question the debate is missing. And the math demands an answer.
So here is the math. In the United States, 1.69 million software developers are currently employed across all industries. The computer systems design sector — essentially, software agencies and IT consultancies — employs over 500,000 of them, roughly 30%. Retail employs 9,000.
If just 5% of America's 643,000 retail businesses hired a single developer, that's 32,000 new positions — more than tripling the sector's current developer headcount. Repeat that across construction, real estate, wholesale, and healthcare. The total number of non-software businesses in the United States runs into the millions.
The untapped pool of potential developer jobs isn't measured in thousands. It's measured in companies that currently employ zero developers deciding to hire one.
The US Bureau of Labor Statistics projects 267,000 additional software developer jobs by 2034. That projection implicitly assumes that the industry distribution of developers stays roughly the same — that thirty cents of every developer dollar still goes to a solution provider or IT consultancy.
It hasn't priced in the structural break that began in December 2025. If my thesis is right, the BLS number is not a forecast. It's a floor — and a low one.
So what does this look like inside the companies that actually hire these developers? For a non-software company, software touches everything. Process invoices faster. Generate business intelligence sharper and sooner. Write business proposals in half the time. Give customers better report cards and advise them with data instead of instinct. Process credit notes, rebates, and logistics better, faster, cheaper. Build better business plans, clearer trading programs for sales teams, better systems for managing people.
Every one of these surfaces can be improved by software. In the old world, improving any one of them meant either buying expensive off-the-shelf software or hiring an agency at $200,000 a pop. In a free-cognition world, where the cost of trying rounds to zero, the only constraint on how many surfaces get improved is how fast you can ship — which is exactly what the AI-amplified senior developer makes possible.
The competitive advantage isn't a single killer application. It's compound operational advantage across every surface at once. Margins increase because expenses go down everywhere simultaneously — on invoicing, on logistics, on rebate processing, on reporting, on planning.
The competitor doesn't see what hit them. They see the results — better pricing, faster response times, sharper proposals — but not the software that produced them. They don't know they're being out-engineered. They just know they're losing.
This is the force that will drive demand for developers far beyond anything the official projections anticipate. Every company that watches a competitor pull ahead on a dozen operational fronts at once will eventually figure out why. And the answer will be: hire developers.
That demand will eat through the existing supply of developers — first the AI-native ones who don't need convincing, then the ones who can adapt — far faster than the pipeline can replenish them. We are not looking at a gradual rebalancing of the labor market. We are looking at a shortage.
6. The Honest Parts
If this is right, you would expect to see the outsourcing market contract as companies pull work in-house. Instead, the global IT outsourcing market is projected to grow from $564 billion in 2025 to $977 billion by 2031 — a compound annual growth rate of nearly 10%.
This looks like a contradiction, but it isn't. It's exactly what Jevons Paradox predicts. The total demand for software is exploding so fast — because AI makes software cheaper to produce and more valuable to have — that every channel expands simultaneously.
The outsourcing market can grow in absolute terms while losing share to in-house development. The pie gets bigger for everyone, even as slices shift. My thesis does not predict the death of the outsourcing industry. It predicts that the most interesting growth in developer employment over the next decade will happen somewhere else entirely.
One more assumption needs to be named. This thesis presumes that we do not achieve AGI — artificial general intelligence — in the foreseeable future. Not the weak-form AGI that Jensen Huang already claims we've reached, where AI can autonomously generate economic value in narrow domains. I mean strong-form AGI: a system that can do everything a human can do, at which point the concept of "work" dissolves and the question "where will developers work?" becomes meaningless.
My thesis doesn't need AGI. It only needs AI to be good enough to multiply a developer's output several times over — and that threshold has already been crossed. The Dario Amodeis and Sam Altmans of the world need AGI for their predictions of mass job elimination. I don't.
There is also a structural reason to believe this redistribution is not just possible but inevitable. The economist Carlota Perez has studied every major technological revolution since the Industrial Revolution and found a recurring pattern.
Each begins with an installation period — infrastructure is built, speculation runs wild, a handful of companies capture the value. Then comes a deployment period — the technology diffuses through the broader economy, transforming industries far from its origin.
Her key insight: "The companies that lead during installation are rarely the same companies that lead during deployment." The railroads weren't dominated by the companies that laid the track. The internet wasn't dominated by the companies that ran the fiber.
AI is in its installation period. Big tech owns the infrastructure and the headlines. My thesis is about what the deployment period looks like for the labor market — and in a deployment period, the incumbents never keep all the gains.
But there is a real danger here, and it would be dishonest not to name it. Jim Rutt, the complexity researcher and former chairman of the Santa Fe Institute, describes AI's progress as a scaffolding plank that keeps rising from below.
"Today's coding assistants handle boilerplate. Tomorrow's may handle architecture. Next year's may translate business requirements directly into system specifications. The window of 'uniquely human' judgment is not a ledge. It is a scaffolding plank that AI keeps raising from below."
If Rutt is right, the developer's role keeps narrowing, and the person whose superpower is writing clean, careful code may not be the same person whose superpower is specifying constraints for a code-generating AI and auditing its output. Different people. Different skills.
The redistribution creates winners and losers within the developer profession itself. Rutt notes that "the paradox does not promise comfort. It describes a mechanism." A mechanism does not care who it benefits.
7. Who This Is For
If you are a CS student being told your degree is a dead end — it isn't. The demand for developers is about to expand, not contract. But the jobs may not be where your professors assumed they would be.
You may not work at a tech company. You may work at a distribution firm, or a hospital, or a construction company — and that will not be a compromise. It will be the point.
The skills you need are not just clean code and algorithms. You need to know how to talk to non-technical people. How to translate a warehouse manager's frustration into a specification. How to own the outcome of software that touches every surface of a business. Those skills are not taught in most CS programs today. They need to be.
If you are an educator designing the curriculum that produces the next generation of developers — pay attention to what the job market is about to demand. The developer of the near future does not need to be optimized for passing a FAANG whiteboard interview. They need to be equipped for a world where their employer is a retailer, a manufacturer, a logistics firm, a hospital network.
That means teaching domain fluency: how to listen to a non-technical stakeholder, how to extract requirements from someone who doesn't know what a requirement is, how to design software that fits into a business process rather than expecting the business to adapt to the software.
It means teaching AI-native workflows — not just using Copilot as autocomplete, but directing agentic coding systems, writing specifications, verifying AI-generated output. The curriculum that prepares developers for Google is not the curriculum that prepares developers for the other 643,000 businesses in the United States that have never employed a programmer. Someone needs to build that curriculum. It might as well be you.
If you are a developer currently working at a solution provider — your career path is forking. One path leads to being hired by a domain company, where you become the owner of its entire software capability, amplified by AI into a one-person army. That is not a step down from agency life. It is the most leveraged a developer has ever been.
The other path leads to staying at an agency that survives by being genuinely excellent at something the commoditization of coding cannot touch — specialized security, regulated industries, legacy systems with decades of embedded logic. The mediocre middle is disappearing. Pick a side.
If you are the CEO of a non-software company — the economic calculus has changed. Hiring your first developer is no longer a luxury. It is the difference between compounding operational advantage across every surface of your business and watching a competitor do it first.
The developer you hire is not a cost center. They are your competitive advantage — but only if you give them the right role. Do not hire a developer and then bury them in tickets written by people who don't understand software. Hire a developer who can talk to your warehouse manager, understand the problem, and own the outcome.
That developer exists. They are about to become available in large numbers. And if you don't hire them, your competitor will.
This is a prediction. It cannot be proved with past data because the structural break it describes — the emergence of powerful agentic AI platforms around December 2025 — has not yet worked its way into the employment statistics. Past data is rearview-mirror data.
The testable version of this thesis is simple: over the next three to five years, the share of software developers employed outside the tech sector should rise measurably. If retail goes from employing 0.5% of developers to 2%, the thesis is vindicated. If it doesn't, I'm wrong.
But the math, the mechanism, and the history of technological revolutions all point in one direction. AI will not replace programmers. It will not even reduce the total number of programmers. It will redistribute them — from solution providers to domain companies, from the software industry to everywhere else.
And when millions of companies that have never employed a developer decide to hire their first, there will not be nearly enough to go around. The programmer shortage the experts are predicting is not an overestimate. It is a dramatic underestimate, because the experts are looking at the wrong part of the economy.
Look where there are zero developers today. That's where they'll be tomorrow.
Top comments (0)