There is a peculiar contradiction at the centre of today’s startup economy. Entrepreneurs have never had access to more information about markets, consumers and technological change, nor have they ever been able to build and test new products so quickly. At the same time, I continue to see founders approaching entrepreneurship in precisely the opposite direction: they become fascinated by a technology, develop a product around it and only afterwards begin looking for a sufficiently important problem that the product might solve. In my experience, this is one of the most persistent and expensive mistakes a young company can make.
The scale of capital flowing into technology can make that mistake easier to overlook. In the first half of 2026 alone, US startups attracted more than $400bn in venture investment. The previous year, approximately $320bn was invested across more than 15,000 deals, an increase of more than 50% in capital compared with 2024. Yet these extraordinary headline figures conceal an increasingly selective market. Artificial intelligence companies captured roughly 65% of US venture deal value in 2025, while fewer than 500 financing rounds of $100m or more accounted for around two-thirds of all venture dollars invested. Capital is abundant, but it is not evenly abundant: enormous amounts of money are competing for a relatively small group of companies that investors believe have found not merely an interesting technology but an important market at precisely the right moment.
Look at some of the companies attracting attention and the pattern becomes easier to see. Rillet, founded by Nicolas Kopp and Stelios Modes, is rebuilding accounting around AI and automation. Lovable, founded in Stockholm by Anton Osika and Fabian Hedin, allows people to create software by describing what they want in ordinary language. Mercor, founded by Brendan Foody, Adarsh Hiremath and Surya Midha when they were barely out of their teens, began with AI-assisted recruitment and found itself serving an exploding market for doctors, lawyers, engineers and other specialists needed to train and evaluate increasingly sophisticated AI systems. Anduril, founded by Palmer Luckey and a group of co-founders including Brian Schimpf, Trae Stephens, Matt Grimm and Joseph Chen, approached defence as a technology company rather than as a conventional defence contractor.
These businesses have little in common at first sight. One sells accounting software, another allows non-programmers to build applications, another connects scarce human expertise with AI companies, and another builds autonomous defence systems. Yet I believe they illustrate the same entrepreneurial principle: each sits where a large existing market is being disrupted by a technological or institutional change that has made the previous solution increasingly inadequate.
That is where I would encourage founders to look.
Start with the pressure, not the fashion
When I evaluate a startup, one of the questions I find most revealing is what happens to the prospective customer if the company does not exist. If the answer is that the customer continues more or less comfortably with the existing solution, then even impressive technology may face a difficult commercial journey. If, by contrast, the customer is losing money, time, security or competitive position every month that the problem remains unresolved, the economics immediately become more interesting.
Rillet is a useful example precisely because accounting software is hardly a new market. Businesses have been buying accounting systems for decades, and companies such as Intuit, Oracle and NetSuite are deeply embedded in corporate finance departments. The opportunity was not created by discovering that businesses needed accounting software; it emerged because AI made it possible to reconsider how much of the accounting process had to remain manual.
By the time Rillet announced a $100m Series C at a $1bn valuation, the company said it had more than 600 customers and that new annual recurring revenue had doubled within a quarter. The financing itself reportedly came together in roughly 48 hours. What interests me, however, is not the speed of the fundraising. It is what customers were doing. Rillet reported that significant numbers were moving from QuickBooks, NetSuite, Sage Intacct and larger enterprise platforms.
For me, this distinction between adoption and displacement is extremely important. A customer trying a new application demonstrates curiosity; a customer abandoning software that has been embedded in an organisation for years demonstrates something much more valuable. Switching accounting systems involves migration, retraining, risk and internal resistance. If a young company persuades customers to accept those costs, the product is probably solving a sufficiently important problem to overcome inertia.
Lovable illustrates a different kind of disruption. Anton Osika and Fabian Hedin did not discover that people wanted software; what changed was who could plausibly create it. Their product allows users to describe an application in natural language and have AI generate much of the working software. Within months of launch, the company was reporting millions of users and a revenue trajectory that took annual recurring revenue towards the $100m level with extraordinary speed.
The interesting point, in my view, is not simply that AI can make programmers faster. It is that the boundary between people who can create software and people who cannot is becoming less rigid. Whenever technology changes the identity of the producer, rather than merely improving the producer’s productivity, the potential market can change dramatically. Photography did this, publishing did this, online commerce did this, and generative AI may now be doing something similar to software creation.
Follow the bottleneck
Mercor provides an even more interesting example of why founders need to watch how markets evolve rather than remain attached to the original description of their company. Brendan Foody, Adarsh Hiremath and Surya Midha began the business as teenagers with the broad idea of using AI to improve hiring. That alone would have placed them in an extremely crowded market. What happened next was more consequential.
As AI laboratories raced to improve their models, they discovered that sophisticated systems still required enormous amounts of human expertise for training, evaluation and reinforcement. Suddenly doctors, lawyers, engineers, scientists and other specialists were not simply professionals looking for jobs; they had become inputs into the AI development process. Mercor found itself connecting scarce expert knowledge with technology companies willing to pay for it at scale.
The company’s valuation subsequently climbed into the multibillion-dollar range. The numbers are impressive, but the entrepreneurial lesson is more interesting. Mercor did not need to invent the AI boom. It needed to recognise a bottleneck created by it.
In my experience, technological revolutions create these secondary markets continuously. The smartphone created enormous businesses that did not manufacture phones. Cloud computing created billion-dollar companies that did not own the underlying data centres. Electric vehicles created opportunities in batteries, charging and power management. AI will do the same. Some of its most valuable companies may ultimately be built not around artificial intelligence itself but around the resources, infrastructure and trust mechanisms that AI suddenly requires.
This is why I find experimental projects such as STEEL interesting even before it is clear whether they will become significant companies. STEEL has explored competitive environments in which autonomous AI agents can trade, strategise and compete for real economic rewards. The important point is not whether this particular experiment becomes the next major platform. What interests me is the behaviour being tested: what happens when autonomous agents stop merely answering questions and begin acting strategically in an environment where actions have financial consequences?
The moment agents can buy, sell, negotiate, modify databases or communicate with other agents, a new infrastructure problem appears. How does an agent authenticate itself? What permissions does it possess? Who authorised it? How can its actions be audited? How do two agents establish trust? Who is liable when an autonomous system makes the wrong decision?
Those questions may sound technical today. In my opinion, some of them will become very large markets.
The physical economy is becoming a startup market
Anduril demonstrates the same principle outside conventional enterprise software. Palmer Luckey had already built Oculus before co-founding Anduril in 2017, but the important insight behind the new company was not simply that defence could use better technology. It was that the development model of Silicon Valley — faster iteration, software-defined systems and substantial private capital invested before a government contract was guaranteed — could challenge an industry historically organised around slow procurement cycles and a small number of enormous contractors.
The result has been one of the most striking defence-technology stories of the past decade. Anduril has developed autonomous aircraft, counter-drone systems, surveillance towers and an AI-powered command-and-control platform, while its valuation has climbed into the tens of billions of dollars.
What interests me is the market signal underneath that growth. Western governments are openly acknowledging that conventional defence procurement often cannot keep pace with rapidly changing threats, particularly in autonomous systems, drones and low-cost weapons. US Secretary of the Army Daniel Driscoll has gone as far as publicly calling for “hungry, innovative founders” capable of developing new technologies faster.
For an entrepreneur, an institution with an enormous budget publicly acknowledging that its existing suppliers and processes cannot satisfy a rapidly changing requirement is information worth taking seriously. The customer is effectively describing the gap in the market.
The same principle extends beyond defence. Roughly 80% of the world’s workforce does not spend its working day behind a desk, yet much of the software expansion of the past two decades was designed for office workers. Manufacturing, logistics, construction, agriculture and maintenance still contain enormous amounts of fragmented coordination between people, machines and legacy systems. As robots, sensors and AI agents enter those environments, entirely new operating systems will be required to decide how work is allocated and monitored.
I suspect some of the most valuable enterprise companies of the next decade will emerge here, precisely because the problems are not glamorous. They concern scheduling, maintenance, inspection, safety, procurement and coordination — activities that already consume enormous amounts of money.
Trust is becoming infrastructure
Digital trust offers another example of a familiar problem whose economics are being transformed by technology. For most of the internet’s history, users operated under a loose but generally workable assumption that the person speaking to them, appearing on a video call or sending a message was probably a real person and probably the person they claimed to be. Generative AI is rapidly weakening that assumption by making convincing text, voices, photographs and video inexpensive to manufacture at scale.
The consequences are already measurable. American consumers reported approximately $16bn in fraud losses in 2025. Impersonation scams alone generated more than one million reports and over $3.5bn in reported losses, while investment scams accounted for almost $8bn. Nearly 30% of consumers who reported losing money to fraud said the interaction began on social media, where reported losses reached approximately $2.1bn.
One case illustrates the change particularly well. An employee of a multinational company was persuaded to transfer approximately $25m after joining a video conference in which the other apparent participants, including the company’s chief financial officer, were digitally generated impersonations. A fraud that once might have required compromising accounts or bribing insiders could instead exploit the victim’s most basic sensory evidence: he could see and hear the people apparently authorising the transaction.
Christopher Mufarrige, director of the Federal Trade Commission’s Bureau of Consumer Protection, has summarised the broader economic consequence succinctly: “Fraud undermines that foundation.” Competitive markets depend on participants having some reasonable basis for believing the information and identities presented to them.
In my opinion, this is the beginning of an infrastructure market rather than merely another cybersecurity niche. Banking approvals, recruitment, marketplaces, customer support, corporate communications and online reviews were designed when convincing impersonation was relatively difficult and expensive. If AI permanently changes that cost, future digital systems will need to establish whether an individual is human, whether an AI agent is authorised to represent that person, whether digital content has a verifiable origin and whether a requested transaction is legitimate.
This is another reason I would encourage founders to pay attention to apparently defensive industries such as identity, fraud prevention and compliance. They may sound less exciting than creating the next consumer AI application, but when technological change destroys an assumption on which billions of transactions depend, rebuilding that assumption can become an enormous business.
Sometimes the most interesting startup is solving the least interesting problem
The same logic applies much further down the technology stack. Modern software companies depend on thousands of external services, APIs and packages that constantly change. Engineers who have worked on major cloud platforms have estimated that unnoticed changes to external APIs or software packages can account for more than 30% of service downtime in some environments.
There is nothing particularly exciting about an API changing and breaking somebody else’s software, but the economic cost is real. A system capable of identifying the change, scanning affected codebases, locating potential failures and preparing the appropriate fixes before customers experience an outage would eliminate a measurable operational expense.
Based on my experience, I am generally more interested in a startup capable of removing 30% of an expensive recurring failure than in one capable of producing a spectacular demonstration without an equally clear economic outcome. Technology can make a product possible, but the economic problem determines whether the product becomes necessary.
This is also where I think founders can learn something from the extraordinary growth of companies such as Rillet, Lovable and Mercor without attempting to copy them. None of these companies succeeded because somebody selected “AI” from a list of fashionable sectors. Rillet attacked expensive manual work inside an established business function; Lovable lowered the barrier separating an idea from working software; Mercor found a scarce resource that a rapidly growing industry suddenly needed.
The technologies differ, but the entrepreneurial pattern is remarkably consistent.
Fundraising is the consequence, not the evidence
Startup culture understandably pays enormous attention to financing rounds because they are visible and easy to compare. A company raises $100m, becomes a unicorn or announces a multibillion-dollar valuation and the number immediately becomes part of its public identity.
I think founders should be careful about treating these events as validation in themselves.
A financing process may take 48 hours because the evidence took years to create. By the time sophisticated investors compete aggressively for a company, they may already be looking at hundreds of customers, accelerating recurring revenue, strong retention and evidence that users are replacing incumbent products rather than merely experimenting with a new one. The financing is therefore a lagging indicator of what has already happened inside the business.
When I look at a young company, I would rather understand what happens after the initial excitement. What is retention after six or twelve months? What is churn? How expensive was the customer to acquire? What is the lifetime value of that relationship? How much cash does the company burn to generate each additional dollar of recurring revenue? Does the gross margin improve with scale? Most importantly, is the customer adding another interesting tool to an already crowded technology stack, or is the startup replacing something the customer previously considered necessary?
A billion-dollar valuation does not answer those questions, nor does an enormous total addressable market. Almost any founder can produce a slide describing a market worth tens or hundreds of billions of dollars. There are more than eight billion people who communicate, learn, work, shop, age and require healthcare, but the existence of those people does not constitute a business model. A billion-person theoretical market is of little value if customers disappear after three weeks.
Building has become cheaper; being wrong has not
Artificial intelligence has dramatically reduced the cost and time required to create software. Small teams can now produce prototypes in weeks that would once have required substantially more engineers and capital. I regard this as one of the most positive developments for entrepreneurship in decades, but it creates a paradox: when building becomes easier, founders are tempted to build before they have properly established what should be built.
The declining cost of producing software does not eliminate the cost of producing the wrong software. In my view, customer discovery therefore becomes more important in the AI era, not less. Before committing heavily to a product, founders should identify the assumption on which the business depends and make a serious attempt to disprove it. They should speak to people who supposedly experience the problem, understand how those customers solve it today, establish what the existing solution costs, identify who controls the budget and determine whether the organisation has previously tried to solve the problem.
I am particularly cautious about the question, “Would you use this?” People are generous with hypothetical enthusiasm. Existing behaviour is considerably more informative. If a potential customer is already paying consultants, employing staff, maintaining spreadsheets or buying several imperfect products to manage the same problem, demand has already revealed itself.
Founder-market fit matters for the same reason. A technically brilliant entrepreneur entering an unfamiliar industry may still be at a disadvantage against somebody who has spent ten years living with the problem. Domain expertise produces invisible advantages: knowing which customers really control budgets, which regulations matter, which assumptions are false, where incumbents are vulnerable and which apparently important problems nobody will actually pay to solve.
Mercor is interesting partly because its founders were extraordinarily young and therefore appear to contradict this principle. I would argue that the case actually demonstrates something slightly different: domain expertise is not always measured in years. A founder can compensate for limited historical experience through unusually rapid learning, proximity to the emerging market and a willingness to change the company’s direction when the evidence changes. The relevant quality is not age but the speed at which somebody develops an accurate model of the market.
The founder’s job is not to predict the future
Venture investing is often described as an exercise in predicting the future, but I do not think that description is quite right. The future contains too many variables to predict reliably. What entrepreneurs and investors can do is identify pressures that are already visible and consider what happens if they continue.
Populations are ageing. Artificial intelligence is becoming more capable. Computing demand is increasing. Digital identity is becoming less reliable. Industrial automation is accelerating. Governments want cheaper and faster defence technologies. Companies are beginning to deploy autonomous software systems into real economic processes. None of these observations requires a heroic prediction; they are visible today.
The entrepreneurial question is what becomes necessary if those pressures intensify. That, in my opinion, is where some of the most compelling startup opportunities originate. They do not necessarily come from sitting in a room attempting to imagine something nobody else has ever considered. Quite often they come from looking carefully at what millions of consumers, thousands of companies or entire governments are already struggling to do and recognising that the existing solution will not be sufficient for much longer.
The companies discussed here make that point in very different ways. Palmer Luckey and his co-founders at Anduril recognised that defence procurement and technological development were moving at incompatible speeds. Nicolas Kopp and Stelios Modes saw an accounting industry in which increasingly sophisticated companies were still dependent on labour-intensive processes and software designed for an earlier era. Anton Osika and Fabian Hedin recognised that generative AI could change not only how quickly software was written but who was capable of creating it. Brendan Foody, Adarsh Hiremath and Surya Midha followed the development of AI closely enough to see that scarce human expertise itself was becoming a critical resource. Experiments such as STEEL are now asking what happens when autonomous agents move from assisting humans to competing and transacting inside economic environments.
I would not assume that every one of these models will ultimately produce the winner in its category. That is not the point. What makes them useful to study is that each represents an attempt to build directly into a structural change rather than to manufacture a market around a fashionable technology.
The mythology of entrepreneurship celebrates the visionary who sees something nobody else can see. Such founders certainly exist. Based on my experience, however, there is another kind of entrepreneurial insight that is at least as valuable: seeing the same evidence everybody else can see, understanding its economic implications more clearly and acting before everybody else does.
With more than $400bn flowing into US venture-backed companies in only six months, there is plainly no shortage of capital for businesses that investors believe can capture exceptional opportunities. Nor is there a shortage of technology. What remains scarce is the combination of a genuinely important problem, the right moment to solve it and a founder who understands that problem deeply enough to build something customers cannot easily abandon.
For that reason, I would encourage founders to spend less time asking what investors currently find exciting and more time studying where economic pressure is accumulating. Look at the customer already spending money badly, the institution publicly admitting that its existing system is inadequate, the industry developing a new bottleneck, the demographic change that cannot be reversed, the technology that has made an old solution obsolete or the physical constraint that is becoming increasingly expensive.
In my view, the startup market is already providing an extraordinary amount of information about what needs to be built. The advantage will increasingly belong not to the founders who generate the greatest number of ideas, but to those who can distinguish a fashionable technology from a structural change, recognise the expensive problem that change creates and build the solution before the rest of the market fully understands why it has become necessary.
By Uri Poliavich
