Back to Blog
Engineering9 min read

The Abundance Thesis: Why Cheap Code Means More Software Engineers, Not Fewer

2026-08-02

There is a pattern that repeats every time the marginal cost of producing something collapses, and it was formalized as early as 1865, when Jevons observed that Watt's more efficient steam engine - which needed less coal per unit of work - caused England to burn more coal, not less. Efficiency made steam power viable in places it never was before, and total demand expanded faster than the cost fell. The same mechanism played out in food: in 1900, roughly 40% of the American workforce farmed and households spent around 40% of their income eating; today under 2% farm and food takes about a tenth of income. But nobody ate the same meals and pocketed the difference - the food industry exploded into processing, restaurants, delivery, and global cuisine, employing many times more people than farming ever released. Artificial light is even starker: Nordhaus calculated that the price of a lumen-hour fell by a factor of thousands between 1800 and the 1990s, and instead of buying the same candlelight cheaper, we lit streets, stadiums, offices, and screens - light stopped being a product and became an ambient input to everything. The law underneath all three: when cost collapses and demand is elastic, consumption doesn't hold steady at a lower price. It expands until it saturates a much larger space of uses that were never viable before.

The same collapse is now happening to intelligence itself, and it's measurable. GPT-4 launched in March 2023 at $30 per million input tokens and $60 per million output. Today DeepSeek V4 Flash - a model far beyond GPT-4's capability - costs $0.14 and $0.28: a more than 200x price collapse for a given level of intelligence in about three years. And the compression isn't limited to the low end. Anthropic's Opus 5 shipped at $5/$25 per million tokens, literally half the price of Fable 5's $10/$50 within months of it, at comparable capability for most workloads. OpenAI's GPT-5.6 Sol landed at $5/$30, and within a single month of launch OpenAI cut the Luna tier by 80% and Terra by 20%. Frontier capability keeps climbing while frontier prices hold flat or fall - which means cost per unit of intelligence is dropping at every tier simultaneously. Intelligence, the binding input to software production, is on the same curve that coal power, calories, and lumens rode. What follows from that curve depends entirely on one question.

Is demand for software elastic? I think it's the most elastic demand curve in the economy, and the reason is a viability threshold most people don't think about. For seventy years, software only got built when the value of the problem exceeded the cost of a development team. Everything below that line stayed unsolved: the workshop that runs on a whiteboard, the clinic managing appointments through a phone, the mid-size firm gluing four SaaS tools together with an employee's mornings. That long tail isn't a niche - by count of problems, it is most of the economy. Agents don't just make existing software cheaper; they pull millions of previously-unviable problems above the line. And from my own client work, I can add a second dynamic that compounds the first: demand for software is generative. businesses I've built a solution for came back within weeks asking for more things they hadn't imagined before seeing the first one working. Software isn't like light bulbs, where a satisfied need stays satisfied - each deployed system creates new surfaces (integrations, data exhaust, users with requests), and appetite grows precisely by eating. An application doesn't fill a demand; it gives birth to several. The obvious objection is that food and light are physical goods and cognitive work is different - but we already ran that experiment. The spreadsheet was cognitive automation aimed directly at white-collar calculation, and it destroyed the job it targeted: the US lost hundreds of thousands of bookkeeping-clerk positions after 1980, and gained even more accountants, auditors, and analysts - because when a what-if scenario went from three days of recalculation to a keystroke, people started asking what-if constantly. Automation replaces tasks; a job dies only when demand for its output is satiated. Nobody's demand for software is satiated.

So where does the human work migrate? Follow what the models still can't do. I use the best available agents daily, on real production systems, and here is my honest report: with genuinely hard business problems, they still cannot design the solution. They can't sit with a messy, underspecified business need and shape it into a system - defining the edge cases that matter, designing UX that fits how the actual users behave, exercising taste about what not to build, writing evals that capture what "correct" means for this domain, splitting a project into stages that de-risk it. These are not prompt-engineering gaps; they're the parts of the work where the problem itself has to be constructed before it can be solved, and construction requires context, judgment, and accountability the model doesn't have. That is exactly where abundance economics predicts value migrates: when a resource becomes cheap, its complements become precious. The complements of cheap code are the two roles I keep coming back to - eval creation, translating fuzzy business intent into a precise, executable definition of correctness the system can be held against, and comprehension, because code is now written faster than it is understood, and that gap compounds like debt. Someone must be able to hold the system in their head, or maintainability, debugging, and architectural coherence collapse - the human filter I described months ago, applied at industrial scale. I argued in my previous post that implementation is mostly solved and humans are the constraint; this is the same conclusion reached from the economics side.

The strongest counterargument is Leopold Aschenbrenner's drop-in remote worker: an agent capable enough to be onboarded like an employee, using every application on a work computer, doing anything a remote knowledge worker does. Suppose we get exactly that. What happens inside an enterprise? The same thing that happens with human employees - because an enterprise is, at its core, a verification machine. Human workers are onboarded, managed, reviewed, audited, and trusted incrementally; the entire org chart exists because output must be validated before the organization acts on it. Drop-in agents don't remove that layer, they widen its base: someone must onboard them into the company's context, define what they're accountable for, read their output, and decide what gets shipped, signed, or sent to a customer. The bottleneck on reading, validating, and verifying outputs - the one we already live with in human-capital enterprises - persists in agentic ones, except the volume of output to verify is orders of magnitude larger. My honest guess at the distribution: perhaps 80% of enterprises will consume agents the way they consume SaaS today, with little in-house expertise, because for routine work the agents will be trustworthy enough. But the other 20% - companies building at the technical frontier or operating where software touches money, health, and safety - will need the highest tier of talent there has ever been: people who understand the technology deeply and know how to direct these machines. There is exactly one scenario that breaks this equation: a fully self-learning, self-reflecting, continuously adaptive and thinking(i mean it, current models pretend to be thinking) and predicting outcomes of its actions sort of intelligence in a physical body - at which point you haven't built a tool, you've recreated a human-ish being, and the analysis changes completely. Until that limit, every step toward more capable agents adds to the expansion side of the equation, not the replacement side.

I want to be honest about the caveat, because the aggregate argument is true and still cold comfort at the individual level: the farmhand of 1900 did not become a food scientist, and the bookkeeping clerk of 1985 did not automatically become a financial analyst. Expansion of the field is not a guarantee for every person in it - it accrues to those who reposition toward the new scarcities while the old skill still pays the bills. That transition is the actual risk, not obsolescence of the profession. But the first-principles conclusion stands: collapsing production cost plus elastic, generative, threshold-gated demand has produced abundance and industry expansion every single time it has occurred - and this time the collapsing input is intelligence itself, feeding the most elastic demand curve we know of. The engineers of the next decade won't be replaced by the flood of software. They'll be standing in the middle of it, defining what correct means, verifying that it holds, and understanding systems on behalf of a world that suddenly runs on a hundred times more of them.

References

Economic history:

AI pricing:

The counterargument: