The Number Under Every Business Plan Fell by Two Orders of Magnitude and Nobody Repriced
cr4fts.com/n/wypqk
Every business plan contains a number that nobody writes down and everybody assumes. It is the cost of building the thing. Not the price of the product, not the cost of goods, but the cost of getting from nothing to a version a customer can use. Market sizing rests on it. Funding strategy rests on it. Every judgement about which opportunities are worth pursuing rests on it entirely.
That number has fallen by roughly two orders of magnitude in fifteen years, and almost nobody has updated the conclusions that were derived from it. The result is an industry making decisions with arithmetic from 2010, systematically declining to build things that are now obviously worth building, and systematically overpaying for things that are now cheap.
This is the least discussed and most consequential fact in technology, and it is treated as a footnote about infrastructure pricing when it is actually a repricing of which businesses are permitted to exist.
The pattern is old and it repeats across inputs that have nothing to do with each other.
Storage cost hundreds of thousands of dollars per gigabyte in the early 1980s and is now a rounding error on any invoice. Bandwidth followed the same shape. Compute tracked Moore's law for four decades, and when the per-transistor curve began flattening, cloud provisioning changed the economics anyway by converting a capital purchase into an hourly rental. That last shift mattered far less for the price of a server than for the minimum size of an organisation permitted to have one.
The fourth curve is the sharpest and the best measured. Stanford's 2025 AI Index tracked the cost of querying a model at a fixed capability level, which is the only methodologically honest way to measure this. For performance equivalent to GPT-3.5 on the MMLU benchmark, the price fell from twenty dollars per million tokens in November 2022 to seven cents per million tokens by October 2024. Beneath it, hardware costs are declining around thirty percent annually and energy efficiency improving around forty percent.
Model pricing is partly a competitive weapon rather than a pure cost readout, and a portion of that decline is therefore reversible. But four independent inputs have moved in the same direction for four decades, and every founder who has assumed the current price was the floor has been wrong every single year.
Here is where most analysis of this goes wrong, and the error leads directly to bad hiring.
Infrastructure was never the problem. For a typical early-stage software company, hosting, tooling and third-party services have represented a small fraction of burn for at least a decade, frequently under ten percent. Halving that changes almost nothing. A founder who reads that compute got cheaper and concludes their costs have collapsed has misidentified which cost.
The cost of building software has always been people, and more precisely the coordination of people. A team of thirty exists because the work was divided thirty ways, and it was divided thirty ways because no smaller number of people could hold all of it. Specialisation was the response to complexity. Specialisation requires coordination. Coordination is where the money goes.
What has actually changed is the amount of specialist work a generalist can now do adequately. Not excellently. Adequately, which is the correct standard for a first version and the wrong standard for a mature product. One competent engineer can now produce a passable interface, a reasonable schema, a working deployment pipeline and a functional analytics layer. Each of those was somebody's full-time job in 2015.
This is why the fall in build cost exceeds the fall in any single input. Removing a person from a team does not remove one unit of cost. It removes a salary plus the coordination overhead that person's presence created, and as Fred Brooks established in The Mythical Man-Month in 1975, the second term grows with the square of team size rather than linearly. Five people maintain ten communication channels. Two people maintain one.
The conclusions do not shift gradually as this number falls. They invert at thresholds, which is why so many experienced operators are carrying beliefs that were correct when they formed them and are now precisely wrong.
Consider a workflow tool for a specific trade with perhaps forty thousand potential customers in one country, each able to justify thirty dollars a month. That is a ceiling in the low tens of millions annually if you capture nearly all of it, which nobody does. Assume a realistic share and the business does a few million a year at maturity.
Against a build cost of two million dollars and eighteen months, that business is uninvestable and effectively unbuildable. The payback period is too long, the capital must be committed before any evidence exists, and the risk-adjusted return does not justify the effort for anyone with alternatives. So nobody builds it, and the market gets recorded in everyone's notes as too small.
Against a build cost of fifty thousand dollars and four months, the identical business is straightforwardly excellent. It is not a venture-scale outcome by the standards of a fund that requires one company to return everything, which is exactly why it remains unbuilt while capital chases larger stories. It is a very good outcome for a founder who owns most of it, and it compounds into adjacent trades the moment the first one works.
Nothing about the market changed. The number underneath it changed, and the conclusion reversed. Multiply that across every trade in every country that somebody once modelled and abandoned, and you have a fair description of where the next decade of company formation actually happens. It is not where the attention is.
The first is that you need institutional capital before evidence. When a first version cost two million dollars, you had to sell a story rather than a result, which meant the people who could raise were the people who could tell stories to investors. When it costs a fraction of that, the sequence inverts entirely: build the thing, discover whether anyone wants it, raise on the answer. Capital does not become unnecessary. It becomes optional at the precise stage where it was previously mandatory, and that changes who is permitted to try.
The second is that small markets are not worth serving. Minimum viable market size is a direct function of build cost, and the arithmetic is unforgiving in both directions. Two orders of magnitude off the build cost moves the threshold by two orders of magnitude, and a very long tail of industries that everyone agreed was too small was priced against a number that no longer exists.
The third is that the product is the moat. If you can build it cheaply, so can the next person, and this objection is correct. It is also an argument against treating the artefact as defensible, not an argument against building. What does not commoditise is accumulated understanding of a specific market: which customers actually pay, what the workflow does when it breaks, which regulator matters, who must be persuaded internally before anything is purchased. That knowledge accrues at the speed of time spent inside a problem and is entirely unaffected by the cost of compute.
A founder who internalises the first half of this argument without the second will build something adequate that nobody buys, and this failure mode is now extremely common.
Distribution did not get cheaper. Getting a person to notice, trust and pay for something has become measurably harder, because everyone else's build cost fell simultaneously and the resulting supply of adequate products is enormous. The scarce resource moved from the ability to build to the ability to be found and believed, and almost nothing in the current discourse acknowledges that the constraint relocated rather than disappeared.
Trust did not get cheaper. In categories where being wrong is expensive, health, money, anything regulated, buyers require evidence that accrues only at the speed of time passing. No cost curve compresses a track record, and no amount of product quality substitutes for one.
Operations did not get cheaper proportionally. Software touching the physical world still requires vehicles, warehouses and people who show up. The software got cheap; the world it operates in did not.
And the last mile of quality did not get cheaper at all. Reaching eighty percent is what collapsed in price. Getting from eighty to ninety-nine costs approximately what it always did, and for any serious product that is where most of the work lives. A great deal of current disappointment with these tools is people mistaking the first number for the whole job.
The saving is real at the first-version stage and shrinks as a product matures, because maturity consists almost entirely of the last-mile work that did not get cheaper. This has a specific operational consequence that most teams get wrong in one of two directions.
Teams that staff for the mature product from the beginning lose the advantage before they have used it. They hire the specialists they will eventually need, absorb the coordination cost immediately, and move at the speed of a company three times their revenue. Teams that never staff up hit a hard ceiling at roughly the point where the remaining work is all edge cases, and then spend two years failing to cross it.
The shape that works is a very small team through the first version and the first paying customers, then deliberate specialisation only where the last mile actually demands it. That is usually narrower than expected and almost never in the functions people hire for first.
If the cost of building has fallen by two orders of magnitude, then every judgement derived from it needs redoing, and almost nobody has done the work. The consequences show up in three places, and all three are currently mispriced.
Venture returns are the first. A fund model requiring each investment to be capable of returning the whole fund forces every company toward the largest possible market, which was rational when building was expensive because the expense had to be justified by an enormous prize. That constraint is now imposed by the fund structure rather than by the economics of the business, which means an entire category of companies is being declined for reasons that stopped being true. The businesses are good. They are simply the wrong shape for the instrument being offered.
Acquisition prices are the second. Firms continue to pay for software assets on multiples that assume the asset would be expensive to rebuild. For a meaningful share of what changes hands, rebuilding is now cheaper than buying, and the acquirer is paying for a customer base and a brand while telling itself it is paying for technology. That is not always a bad trade, but it is a different trade from the one being described in the announcement.
Internal build-versus-buy is the third and the most immediate. Enterprises still default to buying because internal development has a reputation for cost and delay earned across two decades of accurate observation. That reputation is now out of date for a specific and growing class of problem, and the vendors who price on the old assumption are the ones most exposed.
All three are the same error in different clothing: a decision rule calibrated against a number that has since moved by a factor of a hundred, applied with confidence because it used to be right.
The discipline this demands is about which costs you plan around. A cost sitting on a curve that has halved repeatedly should never be planned at today's level, because planning at today's level is planning to be wrong in a direction you can already see.
This cuts both ways and the second direction is the one that kills people. A business whose advantage is that something is expensive should assume the expense is temporary, and should know what it becomes when the expense goes away. A business that only works once an input is far cheaper than today should assume it will get there, and should find a version that is useful at current prices in the meantime, even if that version is much smaller than the ambition.
The cost of building fell by two orders of magnitude and the received wisdom did not move. That gap is the opportunity, and it will close as the arithmetic propagates. Anything priced on today's cost of building is already priced wrong, and the people who reprice first will take the markets everyone else has already written off.