For most of my career, capital was one of the scarcest inputs in building a company. Software took armies of engineers, months of runway, and a plan credible enough to convince someone to fund it before a single customer ever saw the product. That friction was painful, but it imposed discipline. Founders had to choose.
AI is removing much of that constraint. A single developer with a coding assistant can now build in a weekend what a seed-funded team of six might have spent a quarter building in 2024. As the cost of turning an idea into working software collapses, building stops being the hard part. The scarce resource becomes judgment: knowing what is worth building in the first place. The founders who win this era will prove what customers actually pay for before they scale.
Find the problem before you scale it
One of my favorite examples started as a cluttered little app called Burbn. In 2010, Kevin Systrom and Mike Krieger had built a location check-in product packed with features: plans, points, photos. Users mostly ignored them. But when Systrom and Krieger looked at what people actually did, one behavior stood out: they posted photos. So they did something that becomes harder as companies get larger and better funded. They threw almost everything away. They rebuilt the product around that single behavior, renamed it Instagram, and signed up 25,000 people on launch day.
Slack has a similar origin story. Stewart Butterfield's team spent years building an online game called Glitch that never found a large enough audience. But the team loved the internal communication tool they had built to work on it. They shut down the game and followed the signal. Neither company found its opportunity by perfectly predicting the future. They found it through experimentation, observation, and the humility to admit which experiment had actually worked.
That is the approach I keep coming back to when I talk with founders today. Start with the narrowest painful problem you can find, ideally work customers hate doing or that consumes an unreasonable amount of their week, and solve it completely. Run small experiments. Kill the ones that fail. Follow the ones customers pull out of your hands. Don't build a sprawling system before you know which piece anyone actually wants. This matters even more in AI because nobody can see that far ahead. When the technology itself changes every few months, long-range product certainty is mostly fiction. The advantage goes to teams that can learn and adapt faster.
When an experiment fails, the explanation I hear most often is that customers just don't get it. My answer to founders is always the same: customers are voting with their wallets, and you are voting with your instinct. When the two disagree, I side with the wallets every time. If a product needs customers to "get it" before they will pay, it usually has not found its problem yet.
Treat the journey as a series of experiments
See your startup as a series of experiments, not a single bet. Most of them will fail, which is why resilience and grit matter as much as the idea. It is also why experience on a prior startup counts for so much: you already know the drill, how long things really take, and what it feels like when an experiment dies. No wonder investors write checks most readily to founders who have already been on the journey.
What investors generally will not do is pay for the search. Venture capital funds a company that has found something, not the experimentation it takes to find it. That part is yours to figure out. Switching to a completely different experiment, the pivot, is not a failure. Burbn and Glitch were both abandoned experiments before they became Instagram and Slack. And AI has made pivoting cheaper than it has ever been, so you can keep running new experiments until you've found your niche.
Small teams, outsized outcomes
And when a small team does lock onto the right problem, the results can be extraordinary. Instagram had roughly 13 employees when Facebook agreed to buy it for $1 billion in 2012. WhatsApp had around 55 when Facebook bought it for $19 billion two years later. Mailchimp never took venture capital, grew for two decades on customer revenue, and sold to Intuit for $12 billion in 2021. These companies did not win by having the largest payroll. They found something people cared about intensely, stayed focused on it, and resisted diluting that focus.
AI should make this kind of company more common. The amount of labor required to build software, operate a business, and serve a customer is falling. A company that once needed 100 people may need 20. One that needed 20 may need five. But that does not mean scale is obsolete.
When scale is the moat
Some of the greatest companies of my lifetime were built by going big deliberately. Amazon started with books, then spent years pouring money into warehouses, logistics, and eventually AWS. Uber raised tens of billions of dollars and subsidized rides city by city to create networks dense enough that a car would reliably appear within minutes. It lost money for more than a decade while it scaled. Facebook raced from campus to campus because a social network has little value unless the people you know are already on it.
In each case, scale was not vanity. Scale was the moat.
Network effects, logistics density, capital-intensive infrastructure, and winner-take-most markets can reward whoever gets big first. When the underlying economics have that shape, capital is fuel.
Fund math is not company math
The mistake is assuming every business has that shape, and the venture industry increasingly has incentives to do exactly that. As funds get larger, they have to write larger checks and pursue larger outcomes for the economics of the fund to work. A billion-dollar fund cannot generate a great return through a portfolio of companies selling for a few hundred million dollars each. It needs outliers. That logic makes perfect sense for the fund. It does not necessarily make sense for the company.
Many businesses are naturally excellent niches: focused products serving specific customers extremely well, with strong margins and durable economics at modest scale. Put a megafund's growth expectations on one of those businesses and you can destroy what made it attractive in the first place. The company hires ahead of demand, enters adjacent markets before owning its core one, and trades a durable niche for a fragile attempt at dominance. (There are funding models built for exactly these companies, and they deserve more attention than they get.)
Scaling before you have earned it
The wreckage from that mistake is instructive. Webvan went public in 1999 promising rapid grocery delivery, then committed enormous sums to automated warehouses and national expansion before proving that the economics worked in a single market. It was bankrupt by 2001. Juicero raised around $120 million to build an internet-connected juicer that cost hundreds of dollars, only for reporters to demonstrate that its proprietary juice packs could be squeezed effectively by hand. Then there was Quibi: two of the most accomplished executives in media, roughly $1.75 billion in funding, a Super Bowl ad, and a roster of Hollywood stars. It shut down about six months after launch, never having answered the most basic product question: why would consumers pay for short-form video on their phones when enormous quantities of it were already available elsewhere for free?
These companies had different problems, but they shared a failure of sequencing. They scaled before they had earned the right to scale. Capital allowed them to execute assumptions that should still have been experiments.
Good reasons and bad reasons to raise
That leads to a rule that should always have been obvious: raise money when you know what the money is for. A good reason to raise is that you have an engine that works and needs fuel: a distribution channel that converts profitably, a product customers want faster than you can deliver it, infrastructure that genuinely requires upfront investment, or a market in which scale itself creates defensibility. A bad reason is that a fund is eager to invest, a competitor just announced a round, or a larger number feels like validation.
Money raised before you have found the right problem is expensive in ways that are easy to miss. It dilutes ownership. It creates a valuation you now have to grow into. More subtly, it changes behavior. A large bank balance creates pressure to hire, launch, expand, and "put the money to work" when what the company may actually need is six more months of inexpensive learning. AI strengthens the case for restraint. If a tiny team can build and test a product cheaply, much of the early learning that once required venture capital no longer does. Raise to accelerate discovery only when discovery itself is expensive, and expect to fund most of that search yourself. Otherwise, raise to accelerate something that has already been discovered.
What still counts as a moat
AI changes the nature of defensibility, too. When software becomes cheap to build, it also becomes cheap to imitate. A large engineering team and a long feature list are weaker moats when a competitor can reproduce much of your interface in weeks. What remains difficult to copy is everything accumulated around the software: proprietary data, deep understanding of a particular customer or industry, integration into critical workflows, distribution, trust, and becoming a system of record that customers are reluctant to remove. Advantages like these come from depth, not breadth.
Generic AI products face another problem: incumbents already own the customer. Every established software product can add AI to workflows their customers already use. For many buyers, "good enough and already integrated" will beat "slightly better but new." That leaves startups with an opening, but often a narrower one: understand a specific customer or workflow so deeply that a horizontal incumbent cannot justify matching you.
The economics of AI make choosing correctly even more important. Traditional SaaS could add users at relatively low marginal software cost. AI products often incur meaningful inference costs every time customers use them. Scale the wrong product and you do not merely waste salaries and marketing dollars. You can also accumulate a large usage bill proving that people do not value what you built.
Earn the next problem
None of this means companies should stay small forever. A niche is a beachhead, not a ceiling. Amazon started with books. Great companies keep expanding, but the best expansions are earned. Own one problem. Let that ownership generate data, trust, distribution, and insight. Then follow those advantages into the next problem. Capital should follow demand rather than attempt to manufacture it.
The winners of the AI era will treat capital as scarce even when it is abundant, and measure what each dollar of burn buys. They will run more experiments, keep teams smaller for longer, kill bad ideas faster, and raise serious money only when they can explain precisely what proven engine that money will accelerate. They will pursue scale when scale creates an advantage, not because scale itself looks like success.
For investors, that means accepting an uncomfortable truth: not every great company is a venture company. For founders, the lesson is simpler. AI can build almost anything you ask it to build. The hard part is deciding what to ask AI for.
Prove it first. Stay small until customers start pulling the product out of your hands, then, and only then, pour on the fuel.If you are weighing whether it is time to raise, or what the money would actually accelerate, book a 30-minute consultation. We will look at your numbers together and tell you plainly whether you have an engine ready for fuel.
