The cost of
being wrong
collapsed.
We do not predict which B2B software companies will work. We build them cheaply enough to find out inside ninety days.
Venture capital exists in its current form because finding out whether a company worked used to cost millions of dollars and eighteen months. When discovery is that expensive, someone has to guess in advance. The guess becomes the product. Guessing is what venture firms sell, and equity is what founders pay for it.
That expense is gone.
Finding out whether a B2B software company works now costs almost nothing and takes about ninety days. The apparatus built to compensate for expensive discovery is legacy infrastructure maintained for a problem that no longer exists. The deck, the warm intro, the pattern match, and the founder-first bet were reasonable responses to a shortage of evidence. The shortage is over.
We build the companies cheaply enough that finding out replaces guessing.
Venture capital is a technology for making decisions without evidence.
This is a description, not a criticism. For fifty years it was the industry’s historical function, and doing it well was genuinely hard and genuinely valuable.
In 1975, or 1995, or even 2010, the question “will anyone pay for this?” required capital before information. You needed engineers for a year. Hardware, or servers, or a data center. A sales team to find out if the thing could be sold. Eighteen months of runway before the first real signal. Discovery cost seven figures.
When information costs seven figures, you cannot buy information. You have to buy a judgment about information you do not have yet. An industry formed around producing those judgments, and every institution inside it is a tool for guessing better under scarcity:
- The pitch deck. A compressed argument for a future that has not happened, because the present contains no data.
- The warm introduction. Trust used as a filter when evidence is unavailable. If you cannot evaluate the company, evaluate who vouches for the founder.
- Pattern matching. “This looks like the last one that worked.” A heuristic deployed because the specific case cannot be tested.
- “We invest in exceptional founders, not ideas.” The purest statement of information scarcity. When the idea cannot be tested, bet on the person and hope they navigate to something real.
- Staged financing. Buying information in increments because you cannot afford it all at once.
- The power law and portfolio construction. A statistical hedge for an industry that expects to be wrong most of the time. You do not build a portfolio of ninety percent failures if you can tell in advance which ones fail. Portfolio math is the admission that prediction does not work well.
Each of these is a reasonable adaptation to a world where evidence was unaffordable.
That world ended.
Three costs collapsed at roughly the same time.
Building. A competent technical founder with modern tooling ships a working, sellable B2B product in weeks. The nine-month engineering push before the first customer conversation is now a month. This is not a marginal improvement; it is an order-of-magnitude change in the cost of producing a testable artifact.
Infrastructure. Compute, storage, models, and tooling for an early-stage company are effectively effectively free: six figures of credits are available to any properly-formed entity. The fixed cost of running the experiment rounds to zero.
Reaching the buyer. Still hard, and in some ways harder than it was, because every inbox is fuller. It is no longer gated by a hired sales organization. One operator with the right infrastructure can put a product in front of enough real buyers to get a real answer.
Put together, the question of whether an enterprise B2B buyer will pay a meaningful contract value for a piece of software moved from a two-million-dollar, eighteen-month question to a five-thousand-dollar, ninety-day question.
A four-hundred-fold reduction in the cost of finding out.
Institutions built for the old number have not repriced.
When measurement gets cheap, expert prediction becomes a liability.
The transition has happened before, in every field where the cost of testing fell below the cost of guessing. The pattern is consistent, and it has been unkind to the experts each time.
The creative director's intuition about which campaign would land was the product for a century. A/B testing made outcomes measurable at near-zero cost. Intuition did not become a better input. It became a worse one than simply running the test, and now survives as a hypothesis generator rather than a decider.
Medicinal chemists predicted which compounds would bind. High-throughput screening made it cheap to test hundreds of thousands. The prediction did not get better. It got bypassed.
Grandmaster intuition was the pinnacle of pattern recognition until search and evaluation got cheap. Now human intuition is a rough prior that engines routinely overturn.
Scouts were not wrong to trust their eyes. They were solving an information-scarcity problem with the only tool available. When data got cheap, that tool became a source of systematic bias.
In each case the same sequence played out. Expert prediction was revealed to be less accurate than measurement. The expertise migrated from deciding to designing what to test. Institutions built on the old model defended it long past the point of evidence, because their status and economics depended on it.
Venture capital is the last major capital-allocation industry still operating on pre-measurement economics. It allocates on judgment, in advance, at scale, with a widely-accepted failure rate that would be unacceptable in any field where the alternative was cheap.
The industry is optimizing a solved problem.
Enormous effort inside venture goes into making better guesses. Better sourcing, better diligence frameworks, better founder assessment, better pattern libraries, AI-assisted screening.
All of it is investment in a capability whose value is collapsing. The marginal return on being a better guesser drops toward zero the moment the answer is cheaply obtainable. It does not matter how refined a judgment about whether a market exists is, if someone can go find out for five thousand dollars in ninety days. Their answer is not a better guess; it is an answer.
There is a deeper problem. Prediction under scarcity produces specific and well-documented distortions the industry has never escaped:
- Founders are selected on the ability to tell a compelling story about the future. That is a different, largely unrelated skill from building a company that customers pay for.
- Warm-intro filtering encodes network position as a proxy for quality, systematically excluding capable founders outside existing networks.
- Pattern matching reproduces the past, and is worst exactly when a genuinely new thing appears.
- The pitch process rewards conviction over calibration. A founder who says “I don’t know; let’s find out” is penalized relative to one who asserts certainty they cannot possess.
These are not bugs in execution. They are the necessary artifacts of having to decide without evidence. Now that evidence is affordable, they are pure cost.
Run the experiment.
If discovery is cheap, the correct move is obvious and almost nobody is making it. Build the product. Point it at a named buyer. Run real outbound. See whether anyone pays. Do this in ninety days for near-zero capital, and at the end you are not holding a more sophisticated opinion. You are holding a fact.
This inverts the sequence the industry runs on. The standard order is conviction, then capital, then evidence, with the evidence arriving eighteen months later when it is too late to act on. Once discovery is cheap, the correct order is evidence first, capital second.
The reordering is the whole thesis. Everything else is implementation.
It also changes what a negative result means. Under prediction, a company that fails is a mistake, and reputation and capital are destroyed with it. Under experimentation, a company that fails is a completed experiment with a negative answer, delivered for roughly $5,000 in ninety days. The information is real, it was cheap, and it is not a failure of judgment.
An industry that cannot cheaply say “we found out this does not work” will keep expensive dead companies alive to protect the reputation of the people who predicted them. That describes most of the seed market.
Selection doesn't vanish. It moves.
The obvious objection is that you still have to decide what to test. That is correct and important. Judgment does not disappear. It relocates, from predicting outcomes to selecting experiments. Choosing which buyer, which price point, which wedge, and which market has enough surface area to be worth ninety days is a real and demanding skill.
It is also a categorically different bet, with categorically different stakes.
| Dimension | Predicting outcomes | Selecting experiments |
|---|---|---|
| Cost of being wrong (cash) | $2M and 18 months | ~$5K and 90 days |
| Reversibility | Effectively none | Total |
| Feedback | Delayed years, ambiguous | Ninety days, unambiguous |
| Attempts affordable | Few | Many |
| What compounds | Reputation for being right | Knowledge of what is true |
A wrong prediction is a catastrophe you must defend, because eighteen months and a large check are already spent. A wrong experiment is an ordinary week: a few thousand dollars and ninety days. When the cost of error collapses, the optimal strategy shifts from being right in advance to finding out quickly and often.
The competence being built here is not better taste in companies. It is a better apparatus for producing answers.
An experimental apparatus for a question that used to be unaffordable to ask.
We are not an accelerator, an agency, or a seed fund writing early checks. We are an operating apparatus for producing evidence.
Every structural decision follows from the thesis. None of it is arbitrary.
- Ninety days
- The duration of an experiment, chosen so that a real buyer has time to say yes or no and the answer stays cheap.
- Near-zero build cost
- A discipline, not frugality. Experiments have to be cheap enough to run in volume, or the model reverts to prediction. Expensive experiments force the operator to guess which ones to run.
- Cohorts, not one-offs
- One experiment tells you about one company. A cohort pointed at the same buyer tells you about a market.
- Four outcomes at day 90
- GO, GROW, PARK, KILL. The industry's binary of unicorn or dead exists because prediction can only be vindicated or embarrassed. Empirical work produces a distribution, and honest instruments report all of it.
- Cheap, clean kills
- A kill is a completed experiment with a negative result, delivered in ninety days for roughly $5,000. The willingness to run it is what separates measurement from theater.
- $100K ARR
- A measurement, not a goal. The threshold at which “someone will pay for this” stops being an opinion and becomes a fact with a slope attached.
- The fund comes after the studio
- We do not raise capital in order to make better predictions. We build companies until the evidence exists, then invest in the ones that have already demonstrated they work. Other seed funds buy guesses at a discount; we pay a premium for proof, because proof at a premium tends to beat guesses at a discount when most guesses are wrong.
Where this thesis does not apply.
A thesis without boundaries is marketing. This one has a specific scope.
It applies to B2B software where a buyer can plausibly say yes within ninety days at a meaningful contract value. That is most of B2B SaaS, and it is what we build.
It does not apply to deep technology with multi-year research horizons, biotech gated on trials, capital-intensive infrastructure, or marketplaces with severe cold-start dynamics where nothing meaningful is observable early. In those categories, discovery is still expensive, prediction is still the only available tool, and traditional venture capital remains the right instrument. We do not build there, and we make no claim about it.
A narrower caveat, plainly: ninety days of revenue proves that someone will pay. It does not prove durability, retention at scale, or a billion-dollar outcome. Early revenue is evidence, not proof of everything. The claim is only that evidence beats a deck, and that it is now cheap enough to obtain that continuing to prefer the deck is indefensible.
What would prove us wrong.
We hold this as a bet. Here is what would break it.
If founder quality is so decisive that no volume of cheap experiments substitutes for backing the right person, then we are an expensive random-number generator and the pattern-matchers were correct. It is the strongest objection and we take it seriously. Our position is that founder quality matters enormously, and that ninety days of observed behavior against real buyers measures it better than a pitch meeting does.
If companies that reach $100K ARR in ninety days fail at the same rate as companies that do not, the measurement is worthless and the apparatus is theater. That is directly falsifiable, and we will know it from our own data before anyone else does.
If every founder can run this experiment alone, the arbitrage compresses and the studio's edge reduces to operating capacity. We think that takes longer than it appears, because running a real distribution experiment is still operationally hard even when it is financially cheap. The direction of travel is against us, and we know it.
Seed funds are not unintelligent. If they restructure around funding cheap experiments instead of buying predictions, they have more capital and better networks than we do. Our bet is that institutions rarely dismantle the thing their status is built on. That is a bet about human incentives rather than technology, and it could be wrong.
We would rather state the bet and be wrong in public than dress a business model up as a law of nature.
Built by people who have run the prediction machine and produced the evidence.
A thesis about the obsolescence of prediction is only credible from people who have seen the prediction machine from the inside and built the alternative with their hands.
We have run the prediction machine. Several years sourcing early-stage software into venture pipelines, evaluating dozens of companies on decks, teams, and narrative before any of them had produced evidence worth the name. We watched capable people make confident judgments about futures that had not been tested, on the assumption that testing was expensive. Testing is not expensive anymore.
We have been the evidence. One company built and sold as a founder; two others sold as a senior operator in VP and CMO roles, all in regulated healthcare, which is the most punishing environment there is for producing commercial proof: long cycles, gatekept buyers, compliance at every step. When acquiring a single customer takes a year, you learn precisely what real evidence looks like and how it differs from the vanity metrics that decorate a deck.
We can build the apparatus. The thesis is only actionable if the operator can run cheap experiments in practice: build fast, reach buyers, and read the results honestly enough to kill without ego. That is an operating capability rather than an investing one, and it is the one we have.
The combination is unusual: enough time inside venture to know what it actually optimizes for, and enough time operating to know what a real answer looks like.
If the cost of finding out has collapsed, then guessing is no longer a service. It is overhead.
The next decade of company formation belongs to whoever builds the cheapest, fastest apparatus for producing answers, and to the capital that buys those answers instead of opinions about them.
We are building the apparatus. The fund follows the evidence.