The last unpriced trade in AI

The word ‘venture’ changed meaning

Right now, every fund with a chequebook is elbowing into the frontier labs. Close to half of all venture dollars on earth are chasing the same handful of names, at whatever price it takes to get in. OpenAI and Anthropic took around 14% of global funding in 2025. That is where the money is rushing.

But it’s not necessarily where the growth is.

Of the roughly US$280bn that went into North American startups in 2025, about US$191bn went to late-stage and growth rounds. Seed took around US$20bn, and barely grew. Two-thirds of what carries the venture label is now really growth equity: venture-sized fees for private-equity risk. A cheque into a company’s sixth or seventh round is a lot of things.

Venture is not one of them.

The arithmetic is the problem. Entry price decides the return. A good exit is 100x for the seed cheque and barely a double for the late-stage one. The late-stage company is more proven, but by then the market has made up its mind and the price reflects it, so there is nothing left to underwrite. You are buying consensus, at the consensus price.

None of this spares the local market. Australia’s megafunds became the safe place to park money, and in doing so became late-stage themselves. Pay venture fees for late-stage stock and the whole promise of the word falls apart.

The gap between best and rest is widening

AI has removed the two things that used to constrain who could start a company, the building of the product and the running of the company around it. Both are now close to free. What is left as the deciding factor is judgement about product, technology and business model, which is precisely what a seed manager is paid to have. And the founder who benefits most from those falling barriers is often the domain insider who already holds the things that still matter.

Strip those constraints away and three things happen at once. Far more founders start. The winners grow larger and arrive faster than any prior cycle produced. And because building no longer tests whether a company deserves to exist, plenty now launch that should not have, and the failure rate climbs.

The distance between the best companies and the rest widens, in a stage that already carries the widest return dispersion of any private asset class. Whether that lifts returns for the whole category is uncertain. What is not uncertain is the dispersion, and when AI is prising open the widest spread in private markets, the returns available to an outperforming early-stage manager are among the strongest in private markets. The power law that governs venture is about to get steeper, and it is steepest at the first cheque.

Where the largest early-stage returns come from

The scale is what makes the stakes this high. Enterprise AI spend reached roughly US$37bn in 2025, up from US$1.7bn in 2023, on its way to a forecast US$2.5tn in global AI spend in 2026. The build-out is now the main driver of the US economy. Data-centre and AI investment added around a percentage point to GDP growth in the first half of 2025 on Morgan Stanley’s estimate, and on Harvard economist Jason Furman’s it accounted for roughly 92% of all the growth there was, without which the economy would have been close to flat. The most valuable private companies on earth are now AI labs, and AI-native companies are growing at roughly four times the rate of the software that came before them. Whether to own AI is no longer the interesting question. You have to. The question worth arguing about is how.

The answer lives in a pattern that is almost boringly reliable. The largest early-stage returns have tended to cluster around platform shifts. The shift resets the field, incumbents built for the old way cannot retool, distribution comes loose, and whole categories get rebuilt by new entrants. It is the one environment in which a small new company reliably beats a large old one. The internet produced Google and Amazon. Mobile produced Uber and Airbnb. AI is a shift on that scale, and a larger one. Bigger than dotcom, cloud and SaaS put together.

And the value did not accrue to the layer everyone stared at. It accrued to the companies built on top of it. Of that US$37bn in enterprise spend, about three-quarters was bought, not built in-house. The frontier labs take the attention and the giant cheques. The companies above them are the durable prize, still early enough for a venture cheque to own.

The early returns suggest the optimism is earned. Venture in this AI cycle is performing like the SaaS boom of the 2010s, the strongest run of returns in recent memory, rather than the dotcom peak, which for all its noise returned little to the funds that chased it. Three things separate it from a bubble: the returns look like SaaS, the spending is funded by profitable companies out of cash rather than debt, and the word “bubble” is reached for mostly by the investors who sat the cycle out.

The incumbents cannot simply follow. Rebuilding a product around AI means pulling engineers off the revenue that already pays the bills, trading a headcount-based business model for an outcome-based one, and carrying committees and inertia a young company does not. They also raised far more capital, so they need far larger outcomes to justify the effort. It is part of why roughly 36% of the S&P 500 turns over in a decade, and most of the index reshuffles across twenty years. Enterprise software is due for that kind of churn: some incumbents will adapt, and many will not.

Entry price, not layer

The AI stack has five layers. At the base is the physical compute that runs everything. Above it sit the foundation models, then the data that feeds them, then the applied infrastructure that connects them to software, and at the top the applications that do a job for a customer. Each layer already has companies too big for a seed cheque to enter. But none of the layers is settled, least of all the applications, where new categories keep opening. What separates an investable company from a priced-out one is not the layer it sits in. It is the entry price.

Physical compute takes most of the money and all of the headlines. NVIDIA and the hyperscalers build the chips, data centres and power, financed by balance sheets, sovereign funds and public debt.

The foundation models, the general-purpose systems everything else is built on, are the work of a few labs. OpenAI and Anthropic have filed to go public at valuations already near US$850 billion and US$965 billion, reached in private markets too late for a venture cheque to matter. SpaceX, merged with xAI, listed in June 2026 at around US$1.75 trillion, the largest IPO ever, and the shares have already fallen below their price. The value was made in private, and it went to the holders who were early.

Two layers in the middle are easy to miss. The data layer is where information is stored, prepared and labelled, and where evaluation and training data is sold back to the frontier labs. Databricks and Scale AI are among its larger names. The applied infrastructure layer is the platforms and tooling that applications are built and run on. Vercel deploys them, Supabase gives them a database and backend, and OpenRouter connects them to the models through a single API. Every AI product leans on tools like these, so the layer grows in lockstep with the applications above it. Both have their giants, and both still hold companies early enough that a first cheque can own a real position outright.

The applications are the finished products that do one job for one customer, and this is the least settled layer of all. There are as many of them as there are jobs to do, and today’s winners keep being displaced. It is where most of the still-affordable entries sit, because it is where new companies are built fastest and priced last. The next Nvidia is being funded right now, at a stage the public market cannot touch.

You cannot buy the future of AI once the market has agreed it is the future. You own it at the first cheque, or you pay full price for it later.

How AI raises the value of early judgement

AI consumes whatever can be measured.

It took the sourcing, the screening and the diligence and made them free. It hands the same conclusion, and the same edge, to everyone. The same deal at the same price. The asymmetric upside vanishes, and with it the thing venture exists to find: the outliers.

Consensus thinking and artificial intelligence are a dangerous pair for venture. Trained on the past, a model is superb at telling you what a strong Series A looks like in 2026. At the first cheque it is close to useless: no revenue, no cohort, nothing to pattern-match on. Point it at that stage and it hands you the same answer as everyone else.

And the companies that return a fund never resemble the ones before them, so the data cannot see them. They are the “weird” deals, the ones a seed-stage investor has to fight for in the room.

The edge has moved back to the beginning. Back the person, early, before the data. The judgement that counts holds two unstable things at once: a feature of the company odd enough to keep everyone else away, and a read on where the category is going and why this team could be the ones to break it open. That is where 100x returns have always lived, and it is where they still are. It is the one place AI cannot erase the edge.

Venture climbed the ladder and left the first cheque behind

The opening exists because the industry drifted away from it. Until 2021, near-zero interest rates flooded venture with cheap capital, and funds bid company valuations up to levels that could not hold. When rates rose, many of the funds that had overpaid went under. The bigger funds mostly survived, but often by moving up-market into later, higher-priced rounds.

The famous early bets that made this country’s venture names were written by small, hungry funds. Those funds grew, and growth forced them to climb, until the small cheque that built the reputation was one they could no longer write. Canva was not backed by the fund now associated with it. It was backed by a small seed fund that shared the name before it grew up and moved on.

Some read the shrinking population of funded seed companies as proof the stage is dying. The truer reading is that the supply of capital is shrinking faster than the supply of companies worth backing. Sound seed-stage founders go unfunded while the money crowds into the thematic of the day.

And where money does reach seed, it chases the wrong signal. Founders with the right pedigree, a stint at a famous scale-up, a degree from a top university, and the right investor connections, raise at several times the valuation a comparable company commands with only a working product, paying customers and real domain expertise to show for it. Large funds have to deploy large sums, which creates structural pressure to pay premium prices for pedigree regardless of whether the company has customers or revenue.

Investors pay up for the CV and underprice the business. For a patient investor, that is the opportunity. The return stayed where the capital left, and the substance is cheapest exactly where the credential-chasers are not looking.

And the best place to write it is Australia

Geography is the last variable, and it favours one place. Australia turns less capital into more than any developed market. Australian venture has compounded a five-year pooled return of 24.4%, almost double the US. Median seed valuations run about half the US for a company of the same calibre, so the entry multiple is structurally better before a single deal is picked.

The outcomes are there too. Measured per dollar invested, Australia produces more decacorns than any market on earth, and more unicorns than the US, the UK, Germany or China. The operators who built the last generation, Canva, Airwallex and Eucalyptus, are recycling capital and expertise into the next.

The one thing the market still lacks is seed capital itself: Australia invests far less at the first cheque than its opportunity warrants, which is precisely the gap a disciplined local fund exists to close. The money bunches at the top: in the first half of 2026 the five largest rounds took four-fifths of all the venture capital raised here, while the sub-scale rounds that are most of the market got only a sliver of the dollars.

The shape of the journey has shifted, too. Revenue arrives faster than ever, but growth capital has become scarcer and slower to reach: Australian founders now raise their first cheque younger than at any point on record, while the median company does not reach a Series B for about a decade. Both point the same way for an early backer, because more of the value is now created, and held, in the long private years a seed investor owns the company. The prize is open. The capital is scarce.

The window is shorter, the prize is larger, and both point early

AI changed the physics of company formation, and every change points to entering earlier. Companies become legible in months, not years: Slack needed two and a half years and around 650 people to reach US$100m in revenue. Lovable reached comparable milestones in eight months with around 45, and Cursor passed US$1 billion in annual revenue in under two years. Those companies run at roughly six times the revenue per employee of classic SaaS, so a single seed cheque can now carry a company a long way on a fraction of the old capital, sometimes all the way to profitability.

And the exit arrives sooner. The incumbents have become buyers, the platforms have become buyers, and both are buying earlier. A company can now be acquired for US$100m before its Series B, rather than spending a decade in the IPO queue.

Once a company is legible, its price reflects it. The window to back the person before the data exists is closing faster than it ever has.

The prize at the other end has grown by an order of magnitude. A decade ago a billion-dollar exit made a career. Today the biggest outcomes are worth ten times that or more, and the returns that reach that scale are only found in the dark, before the company has a price the market has agreed on.

There is a moment, before any of the machinery, when a company is just a person and the investor is just a human, sitting at a table with the laptops closed, talking. That is where the engine Georges Doriot built in 1946 still runs exactly as it was meant to. A billion-dollar fund cannot meet a company there with the same stakes.

When a giant does write a seed cheque, it is buying an option: a small stake and a seat at the table, so it can put real money in later, once the company is proven and the price reflects it. Its conviction, and its capital, arrive after the risk is gone. A dedicated seed fund goes in before, when that first cheque is the bet that has to carry the fund.

What the opportunity lives inside

The stage is only half the answer. The other half is what you back at it, because the same flood that widened the funnel filled it with noise. AI is a cycle, and a loud one. The dangerous version of the trade is the horizontal promise, that every company and every workflow is transformed at once. The returns are not there. They are in the verticals, in the large, labour-heavy industries software never fully captured and can now reach for the first time.

The pattern is not new: cloud storage made 40,000 photos on a phone normal, and GPS reshaped a sense of direction humans had spent millennia building. Those founders had unfashionable ideas in unglamorous categories, and their earliest backers earned outsized returns.

Two things make those verticals larger than they have ever been. The first is what the software now does. Old software sped the worker up, but a trained human still drafted the contract, triaged the patient note, reconciled the ledger. AI does the work itself, and that moves which budget it is paid from. Old software took a slice of the software line, a fraction of what an industry spends. A vertical AI company is paid out of the wage bill, the labour and operating cost of the task it now performs. That is why the market it can address is five to ten times what a prior software cycle could reach.

The second is that the incumbents’ oldest defence is failing. Customer lock-in used to rest on how painful it was to move off a legacy system. AI agents now handle that migration cheaply and quickly, so an AI-native product can win on the merits rather than waiting for inertia to lift.

And it is early, which is the part the noise obscures. Using an AI tool is now common, but deep adoption is early: even in the most AI-exposed work, computer and mathematical roles, penetration sits at about a third. Procurement caution, cultural inertia, and the time it takes an organisation to trust an AI-driven workflow are real brakes, releasing slowly rather than all at once. The window is opening for new entrants, not closing.

The companies that return this cycle are built by operators who spent ten years inside a hard, overlooked industry and are now rebuilding it. The insider who has lived the problem, over the newcomer trying to out-code someone who has. Domain knowledge a model cannot fake, relationships that predate the company, the trust that lands the first customer in a market that is brutal to sell into. An LLM can never dig you a moat.

This is the discipline the flood makes essential. AI has become every would-be founder’s cofounder, and a collapsing barrier to entry does not produce an influx of good ideas. It produces an influx of people, most of them vibe-coding a product that was never a business. “Build it and they will come” is a riskier assumption than ever.

And the buyers are no easier to win: around 95% of enterprise AI pilots still deliver no measurable impact on the bottom line, so most of what gets deployed never proves its worth. The company worth backing clears both bars: a real business under the AI, revenue a buyer can measure inside the first billing cycle, and a moat that compounds with use. Everything else is a demo with a valuation.

In venture, the manager is the return

In venture, which manager you back matters more than in any other asset class. The gap between the top funds and the bottom is brutal, far wider than in buyout or public markets, and a single position can decide which end a fund lands on.

That gap is not luck. The best managers keep winning, because the best founders take lower prices to work with the firms they rate. Pick the right firm and the odds move with you. The return depends on which firm you back.

And not just any firm, a dedicated seed fund

The largest funds have always written seed cheques, so it can look like the same trade. It is not. A million-dollar cheque cannot move a multi-billion-dollar fund whatever it does, so those funds are price-insensitive at the stage. For them the seed cheque buys a look at the later round that actually returns the fund, and a seat at the table: an option, not a commitment. If the fund later declines to follow on, its pass becomes a damaging public signal: the market reads a prominent backer walking away as a verdict on the company. The pattern shows up in the data: seed companies whose prominent backer re-invests graduate at 51%, against 27% when that backer passes.

A dedicated seed fund is a different instrument. When the fund is small, a single winner can return the whole of it, so the seed outcome is the entire return, and the reason to underwrite each first cheque properly is total. The evidence favours the shape. Early-stage returns run ahead of later-stage, smaller funds outperform larger ones, and a portfolio of around fifty companies rarely loses money while leaving the upside uncapped. In the end it is an argument about alignment, and alignment is not sentimental. Trust only exists where someone has something to lose. A fund whose entire return rides on the first cheque has everything to lose. A large fund buying a look does not.

Alignment wins the right to back a founder. What earns the return afterward is a second strength, and it is a different skill from picking. The best early managers do more than pick the company. They work alongside it, opening the door to a first enterprise customer or a senior hire the founder could not reach alone. That hands-on record is also what earns a place in the strongest rounds, because the founders worth backing tend to choose an investor who believed early over one who arrived once the price had moved.

And the manager built for this cycle

Two more traits separate the managers who will capture this from the ones who will not. One is proximity. In a market like Australia the generalist funds that dominate early stage rarely carry a dedicated AI mandate or the depth to judge these companies, while the offshore funds that have the depth lack the local networks to find or back them, so the manager with both AI conviction and domain fluency on the ground is often the only serious buyer at the table.

The other is how the firm itself is built. The old venture model scaled by adding analysts, associates and layers of committee, each one adding cost and thinning the attention any company received. The manager built for this cycle has rebuilt around AI instead, running the screening, research and monitoring that once needed a team of ten or twenty on infrastructure a small, senior team maintains, so the people making the call are senior and the evaluation behind each cheque, from unit economics to the honest question of whether one outcome can return the fund, is disciplined and consistent rather than passed down a chain.

And the system compounds: every deal it screens and every company it monitors feeds back in, so the firm’s read on which founders win, which markets move and which moats hold sharpens with each cheque. Having crossed the AI adoption curve themselves is also what makes their read on an AI-native founder worth trusting.

Where the case could break

None of this is a sure thing. The safety nets are thinner than they were, so a wrong pick costs more, and liquidity is tight while parts of the market run warm. The winners are not yet clear. Moats may prove shorter-lived as the models improve, and the defensibility playbook everyone recites, proprietary data, workflow embedding, accumulating context, is more theorised than proven. Regulation in healthcare, law and finance is still being written, and can cut for the compliant entrant or against the whole category.

None of this argues against the stage. It argues for discipline: enter early and cheaply, refuse to overpay, follow the winners, hold enough names that one can carry the fund, and treat every first cheque as if it were the only one. The risks do not close the opportunity. They raise the price of getting it wrong, which is another way of saying they reward the manager who gets it right.

The bet that is left

Exposure to AI is easy, and mostly already priced. What the moment rewards is scarcer: the judgement to back a founder before the data arrives, in a vertical hard enough that a model cannot fake the way in, through a fund small enough that the cheque is the entire game and aligned enough to treat it that way. Get that combination right and the widest dispersion in private markets works for you instead of against you.

So the exposure worth having in this cycle is a single early cheque, written with conviction into a founder the market has not yet agreed on, by a manager whose whole fund depends on being right about them.

Everything else is paying full price for a story that has already been told.