The founder supply shock: how AI changes who gets to build

AI is going to make far more founders than it makes unicorns. That may turn out to be the bigger economic story.

In March 2026, at least four million people in the United States used ChatGPT to help plan, start, run, or grow a business, according to a May 2026 OpenAI analysis. Most were not building technology companies. They were running shops, consultancies, restaurants, beauty practices, trade businesses, and small agencies. They were writing estimates, untangling permits, comparing prices, answering customers, and doing all the other jobs that appear the moment you put your own name above the door.

A Shopify survey published in July 2026 found a similar mood across five countries: large majorities of founders said technology and AI have made it easier to start with limited resources or expertise. Anthropic’s June usage research found that conversations about starting a business peak on weekends. People are quietly using Saturday to test the thing they used to only talk about at work on Monday.

AI is widening the founder class itself. An idea that once needed permission, capital, or a missing cofounder can now become a credible test. That is the founder supply shock.

Some are technical. Many are not. Some want venture-scale companies. Many want a profitable business that changes the arithmetic of their household, gives them control of their time, employs a few good people, and eventually pays for a better life than the one their salary was going to buy. That outcome will not ring a bell at Nasdaq. It will still matter enormously to a family, a town, and an economy.

The missing cofounder was often a bundle of capabilities

Starting a company has always required a slightly ridiculous collection of skills. You need to understand a customer, shape an offer, make the thing, price it, sell it, support it, keep records, read a contract, chase an invoice, and decide what to do when the first plan meets the world and loses.

The standard answer was to assemble a team. A technical cofounder built the product. A designer made it usable. An accountant kept the numbers honest. An agency found customers. A lawyer reduced the odds of an avoidable disaster. Each person was valuable, but each also came with a search cost, a salary, a queue, and another handoff.

AI unbundles a first approximation of those capabilities from the people who used to supply them. A founder can move from customer interview notes to a prototype, a landing page, an operating model, and a first sales sequence without waiting to recruit an entire miniature company.

The words first approximation are doing important work. A model is not your lawyer, CFO, engineer of record, or head of safety. It does not care if the product works, the customer renews, or the regulator calls. But early companies are full of jobs where a decent first pass is enough to expose the next useful question. Making those passes cheap changes who can stay in the game long enough to learn.

This is especially powerful for non-technical founders. The restaurant operator who understands kitchen waste, the nurse who sees the same broken handoff every shift, the machinist who knows why one finishing step ruins the week, and the freight broker who lives inside an absurd spreadsheet no longer have to persuade a software team to take the problem seriously before they can test it.

Their lack of technical credentials may matter less than their possession of something the technical world has always struggled to manufacture: close contact with a real, expensive problem.

The new founder advantage is not knowing how to make everything. It is knowing something worth making unusually well.

Technical founders gain a different kind of leverage. They can cross design, research, data, operations, and go-to-market boundaries that used to stop them. They can run several experiments in parallel and spend more of their time on architecture, product judgment, and the hard edges of the system. But they lose an old source of protection too. Being able to build the software is no longer enough to make the company rare.

Cheap attempts change who gets to try

The cost of a failed attempt matters more than the cost of a successful one.

When testing an idea requires quitting a job, raising money, finding a cofounder, and spending nine months building, only a narrow group of people can afford to be wrong. When the same idea can be explored over six weekends for the price of a few subscriptions, the option opens to people with less capital, fewer industry connections, and much less appetite for a theatrical leap into “founder mode.”

That does not remove risk. It lets people buy information in smaller installments.

A new working paper examining more than 160,000 Product Hunt launches found that entrepreneurial entry rose sharply after the release of ChatGPT, driven disproportionately by solo creators. It also found that teams still took a larger share of the highest-ranked launches. The research is limited to one software platform, and a Product Hunt ranking is not a durable business. Still, the shape of the result makes sense: the study suggests AI is increasing the number of shots faster than the number of top-ranked launches.

This distinction will save founders a lot of grief. AI makes a plausible attempt easier. It does not make the market obligated to care.

Expect a much wider middle of business formation: solo firms serving customers around the world, family businesses with serious software inside them, and domain experts selling completed outcomes instead of seats in another tool. Some venture companies will reach meaningful revenue before their first large hiring round. There will also be an enormous pile of polished products nobody needed.

Cheaper experimentation naturally produces more failed experiments. The economic gain comes from reducing the cost of each failure and increasing the range of people who can discover something that works.

Supply explodes. Demand does not.

There is an awkward fact hiding inside all this optimism. If millions more people can produce credible software, services, content, and products, the world does not automatically produce millions more customers with spare attention and money.

Production becomes abundant before demand does.

You can already see the pricing pressure. Upwork’s 2026 workforce index found that freelancers doing AI-related work earned more per hour on average, but lower-complexity generative and creative work grew rapidly while earnings per contract fell. More capability entered the market, and lower-complexity AI work paid less per contract. That is exactly what functioning markets do.

For founders, this means a working product has become weaker evidence. So has a handsome demo, a week of impressive velocity, and a pitch built around how hard the code was to write. Customers cannot see your effort, and they have no reason to price it.

The scarce assets move elsewhere: access to a painful problem, a trusted name in a narrow market, proprietary feedback, unusual data rights, regulatory permission, physical distribution, a community with a reason to stay, and the ability to keep learning after the first answer fails.

This is why the current flood of generic AI wrappers feels both exciting and strangely dead. Many are competent. Very few contain a reason to be chosen. The tool made production easier, so the founder spent the saved effort producing more. The customer needed the founder to spend it understanding them.

The best opportunities will often look smaller at first. Shopify reported in May 2026 that long-tail categories outside its top 100 generated nearly 55 percent of sales, and that most orders it attributed to AI discovery went to long-tail products. Shopify has an obvious interest in this story, and its internal attribution is not the whole internet. But the direction is useful: when search and recommendation can understand intent more precisely, an exact solution to an obscure problem can find a global market without pretending to be for everyone.

Small niches are becoming bigger businesses. Generic markets are becoming louder rooms.

Your company is the learning loop

The models will improve. Your advantage cannot depend on them staying bad.

Sequoia’s Pat Grady advised founders in May 2026 to build moats backward from the customer and exploit the gap between what frontier models can do and what businesses have actually deployed. That is a more useful frame than asking which model is ahead this month. The opportunity is in the distance between available intelligence and a completed job inside a real workflow.

But even workflow integration is not enough by itself. Over time, competitors can buy the same models, study the same interface, and copy the obvious path. What they cannot instantly copy is the record of how your company learns: which outcomes count as good, which exceptions matter, which corrections changed the process, and what you now know about the customer that was invisible at the start.

Satya Nadella made this point sharply in a June 2026 essay on the future of the firm. He argued that companies need to retain control of the proprietary context and corrections that make rented models useful, and own the learning loop between their people and AI systems. Microsoft is hardly a neutral observer in the contest over enterprise AI, but the warning is sound.

Suppose a customer corrects a generated quote because one type of site always needs a second inspection. That correction should become a test the system must pass next time. If an output fails, the reason becomes a guardrail. The next job improves because the lesson stayed inside the company instead of vanishing into a chat window. The model is rented intelligence. The accumulated judgment around it is the asset.

This changes the founder’s daily work. Build short customer loops, turn corrections into quality checks, name the person responsible for consequential decisions, and keep a record of why the company changed its mind.

Software is only the cleanest first case

Software moves first because mistakes are cheap, distribution is instant, and the raw material is symbols. It would be a mistake to assume the founder boom stops at the edge of the screen.

Robotics is starting to absorb the same pattern. In a June industry briefing, the International Federation of Robotics argued that generative and agentic AI could reduce specialist programming and eventually let non-experts deploy robotic systems. That could let smaller manufacturers automate tasks without hiring a full robotics team. It does not make robots reliable, safe, or compliant. The same briefing says humanoids remain mostly trials and prototypes. The stubborn parts become easier to see once programming gets simpler.

Look at PickNik, a 35-person robotics software company whose MoveIt Pro system has been used in NASA tests to open spacecraft hatches and move cargo. NASA’s June account also describes the same software being used by BMW on assembly lines, by a construction company programming large robotic arms, and by a cleaning-robot business. One small team built a capability that travels across space, factories, housing, and services. Government research support helped make the commercial product possible.

Another useful example is Rivelin Robotics. According to a June UK Defence Innovation case study, the company grew from one founder to 21 people by attacking a narrow, unpleasant manufacturing bottleneck: finishing metal parts. Its robotic microfactories now serve customers across four countries. UKDI says one US customer cut completed-component cost per kilogram by 90 percent against hand-finishing. That is one deployment, not an industry benchmark, but it shows the appeal of making one stubborn physical bottleneck programmable.

That is the physical opportunity map in miniature. AI lowers the cost of perception, programming, simulation, and design. New commercial models lower the upfront cost of adoption. Digital twins shorten the loop between a change and its consequence. The remaining constraints—uptime, materials, certification, integration, maintenance, and production yield—become the valuable work.

The World Economic Forum’s June manufacturing review describes selected showcase factories cutting lead times, defects, energy use, and inventory with AI, automation, digital twins, and redesigned operations. The founder inference is that newly measurable workflows expose smaller, specific problems a focused supplier can attack.

Atoms are not becoming free. They are becoming more legible.

Space and defence are opening from the edges

The same transition is visible in markets once reserved for governments and enormous contractors.

In late June, NASA and the US Small Business Administration created a program intended to steer growth capital toward small manufacturers and suppliers working on technologies needed for lunar and Mars missions. NASA also plans to select up to three winners in September for its robotically manipulated payload challenge, award as much as $500,000 each, and offer a no-cost in-orbit demonstration currently slated for early 2028.

The money matters. The test access may matter more. A founder can now imagine proving a component in orbit without first financing an orbital mission. That moves a company across the line from clever laboratory work to something a customer or investor can underwrite.

The interesting space companies will not all build rockets. They will build inspection, power, propulsion, navigation, servicing, materials, communications, and the unglamorous supply-chain pieces that make repeated missions possible. NASA made that market more concrete on 30 June with nearly $600 million in awards for four commercial lunar deliveries in 2028, while signalling more work in power, avionics, communications, and navigation. The winners are established providers, and lunar delivery remains brutally difficult. But repeated procurement lets value spread into the ecosystem around the vehicle.

Defence is moving in a similar direction, driven by much darker incentives. On 14 July, the UK launched a new autonomy and robotics competition with more than £1 million across two stages for early ideas spanning land, sea, and air. In June, the European Innovation Council opened direct funding to defence and dual-use startups, including a €100 million scale-up call.

The UK made the new-supplier logic explicit on 13 July when it awarded £3.16 million to three small companies developing cheaper counter-drone interceptors. Those are development contracts, not proven production systems. They still show procurement beginning to favor cheaper mass, faster iteration, autonomous systems, and new suppliers. That creates openings in simulation, sensing, resilient communications, distributed manufacturing, and logistics.

It also carries consequences that a normal software startup can avoid thinking about. A product that influences the use of force is not an efficiency tool with a khaki interface. Founders, employees, and investors have to decide which systems they are willing to build, who controls them, how errors are caught, and what accountability survives deployment.

The ability to enter a market is not the same as permission to stop thinking.

Labour does not disappear neatly into the new companies

The optimistic version of this story says displaced workers become founders and AI does the dull parts. Real economies are not that tidy.

The European Parliament’s July briefing on AI and jobs describes the likely outcome as deep job transformation rather than immediate mass destruction, while noting early pressure on white-collar entry-level work. An ILO study across Southeast Asia published in July 2026 similarly found broad exposure but no evidence yet of large-scale job loss. That is reassuring for the present, not a guarantee about the next model cycle.

There is a nasty sequencing problem. Junior work is often exactly the routine, supervised work that models can absorb first. It is also how people build the context and judgment required for senior work later. If companies remove the bottom rung without redesigning how people learn, they may save money now and discover a shortage of experienced operators in five years.

Entrepreneurship can absorb some of this energy. A person who can no longer rely on one employer may be able to sell a specialized service, run a small product company, or turn professional knowledge into an outcome-based business. More independent work is already visible: Upwork reports that more than one in three skilled US knowledge workers now freelance, though its marketplace position gives it every reason to emphasize the trend.

But founding a company is not a social safety net. It transfers risk to the individual. The person still needs time, confidence, health care, a financial buffer, access to customers, and enough room to fail without taking the family down with the experiment.

If we want the founder boom to be broadly useful, access to models is the easy part. People need paid training on real customer work, benefits that follow them between employment and self-employment, small checks for bounded experiments, and at least one large buyer willing to take a chance on an unknown supplier. Otherwise the people with savings and strong networks will compound the new leverage while everyone else gets a better chatbot and a shakier job.

Investors have a prototype problem

Venture investors are about to see more companies than ever, each arriving with a better demo and a smaller team. Their old filters will fail in predictable ways.

Technical difficulty used to act as a rough screen. If a team had built something impressive, that fact carried information about talent, persistence, and the scarcity of the result. Now a small team can produce impressive work quickly, including a great deal of the pitch itself. The demo still shows what exists. It reveals much less about why this team should win.

The capital market is already splitting. Startup Genome’s 2026 ecosystem report says funding for AI-native startups grew 218 percent from 2021 to 2025 and their ecosystem value grew 969 percent, while North America captured 73 percent of early-stage and 86 percent of late-stage AI-native funding. PitchBook’s global data shows the same barbell: AI accounted for 76.7 percent of venture dollars but 35.9 percent of deals in 2026. A few model and infrastructure rounds make capital look abundant while the ordinary founder still hears that the market is disciplined.

Investors need to underwrite three things. First, evidence of painful demand that persists when the founder is not standing beside the customer. Second, a fast learning loop whose feedback and data belong to the company. Third, dependency economics that survive a model-price change, a platform policy shift, or three competitors reproducing the interface by Friday.

For physical companies, the questions become even more concrete. What has been tested outside the lab? Which certification or procurement gate is next? Can the team manufacture the tenth unit with the same quality as the first? What exactly will new capital retire: technical risk, production risk, market risk, or merely time?

The investor’s value has to move too. Founders with cheap execution need less advice about hiring a standard software team. They need access to customers, test facilities, manufacturing partners, regulators, specialist operators, and follow-on capital suited to the actual constraint. In an age of abundant introductions generated by software, a trusted introduction backed by judgment becomes more valuable.

The map gets weird, not flat

AI gives a founder in Nairobi, Newcastle, Naples, or Nagpur access to capabilities that once sat inside a well-funded team in San Francisco. Translation improves distribution. Cloud infrastructure turns tools on globally. Remote collaborators and agents make small teams less sensitive to location.

Then reality reappears.

Capital, compute, research talent, and major platforms remain highly concentrated. Manufacturing clusters matter because suppliers, technicians, and tacit knowledge live near one another. Defence and space depend on national policy and procurement. Energy prices shape which AI and industrial businesses make sense. Regulation can create a protected market, a useful standard, or a dead end.

So the next startup map will not be flat. It will have more viable starting points and more specialized centers of gravity. A founder can begin almost anywhere, but the company will still need to connect to the place where its stubborn constraint is cheapest to solve.

That may be a robotics cluster, a port, a hospital network, a defence test range, a university lab, a community of small manufacturers, or a city where one industry’s awkward knowledge has accumulated for decades. The ecosystems that win will not simply host the most AI meetups. They will connect new founders to specific customers, equipment, expertise, and permission.

How to spot the opportunities now

This transition creates a lot of noise. A useful opportunity usually contains one newly cheap capability and one stubbornly expensive constraint.

Find the expensive handoff or the expert’s full calendar. Look for the moment where a skilled person copies, translates, checks, diagnoses, prices, or approves something while customers wait. Part of that judgment may be made repeatable. The expert should design and supervise the system, and the company should own the ugly exceptions.

Look for physical work that has become measurable. Better cameras, sensors, simulation, and robot control make tasks addressable that were previously too variable to automate. Start with one painful task and a clear economic threshold. General intelligence can wait.

Look for a market whose gate just moved. A new standard, public procurement route, test program, payment rail, or model capability can turn an old idea into a timely company. The opportunity is often in the boring infrastructure required by everyone walking through the new gate.

Look for a narrow group already buying the work. AI lowers the cost of serving small categories and translating specialist needs into a completed outcome. A market can be narrow in identity and global in revenue. If models improve, delivery cost can fall, while the founder still owns quality, liability, and the customer relationship.

The signal is the same in each case: one capability just became cheap while a stubborn constraint still carries the margin.

What founders should do with the moment

Start from your unfair context, not from the model. The useful question is not “What can AI do?” It is “What can people in this market finally buy because AI changed the cost of doing it?”

Keep the first company small enough to learn. Use AI to reduce fixed commitments before you have evidence. Sell early. Stay close to the customer. Track the decisions and corrections that make the system specific to your market. Spend saved payroll on the constraints that are actually scarce: trust, access, distribution, data rights, testing, and time with the problem.

Know where the machine stops. Put names against consequential decisions. Verify work that can hurt a customer. Bring experts in before the cost of being wrong exceeds the cost of asking. A tiny company still needs grown-up accountability.

And do not confuse leverage with isolation. On Product Hunt, solo entry is rising while teams still dominate the highest-ranked launches. AI can replace a queue. It cannot fully replace a cofounder who sees the problem differently, a colleague who tells you the product is wrong, or a relationship built over years of showing up.

For investors and ecosystem builders, the job is to make the new attempts better, not merely more numerous. Provide first-customer access. Open test environments. Build founder networks around hard domains rather than generic ambition. Create small financing instruments for experiments and different capital for factories. Teach evaluation, security, sales, and financial discipline alongside model use. Measure survival, customer value, and learning speed, not demo-day volume.

The IMF’s Dan Katz wrote in June 2026 that AI’s economic outcome is an adjustment problem: productivity can lower costs and expand demand, but only if new tasks and firms grow fast enough to offset what disappears. New company formation is one of the ways society converts a technical capability into shared economic value. It is not automatic. We have to build the bridges.

More people get to take a serious shot

Every industrial revolution changes who owns productive capacity.

This one is putting a surprising amount of it into a browser window and charging by the month. That is obviously not the whole factory. It is giving millions of people tools to research, design, code, sell, and operate at a level that previously required permission from an employer, a technical gatekeeper, or an investor.

Most of those people will not build famous companies. Plenty will make bad products. Some will discover that they preferred the job. Others will build a useful, profitable business that sends a child to university, gives a partner room to stop working nights, hires five people locally, or brings an overlooked service to customers halfway around the world.

A smaller number will combine cheap intelligence with difficult science, physical production, exceptional distribution, or unusual trust and build companies that could not have existed before this cost curve moved.

We will spend years arguing about the first one-person unicorn. The bigger change will happen more quietly: millions of people getting a credible first shot at building something of their own.

The tools have widened the door. The next economy depends on who gets through it, what they choose to build, and whether the rest of the ecosystem helps the good attempts survive.