The Operator Gap

The story

In March 1996, a 38-year-old Caltech engineer named Bill Gross rented an old brick warehouse in Pasadena and made a bet that sounded insane. The web, he said, had made building a business almost free. If you could make a website, you could sell something at 90% gross margins. Ideas were suddenly cheap. So he would stop building one company and start building dozens, in parallel, from a shared floor with a shared back office, hiring a CEO for each and taking the chairman's seat himself. He called it a company factory. Within a year it had 26 companies. Within three it had eToys, CitySearch, NetZero and GoTo.com, the company that invented paid search and, years later, sold to Yahoo for $1.6 billion.

Then the tide went out. Between December 1999 and January 2000 Idealab raised $725 million from outside investors, at the very top, and spent it into companies that had never proven anyone would pay: Eve.com, Modo, Swap.com, MyHome. eToys, the most recognisable brand in online retail, went bankrupt on April 1, 2001. By 2002 Idealab's own investors were suing to liquidate it. Gross survived, stopped taking outside money forever, and did something almost no founder does: he went back through 100 of his own companies and 100 others to find out what he had gotten wrong. His answer was blunt. "I used to think the idea was everything. Over time I came to think that the team, the execution and the adaptability mattered even more than the idea."

He had discovered the operator gap thirty years early.

What Gross saw in 1996, the rest of the world is seeing now, and at a scale that makes his moment look quaint. Lovable went from $1 million to $100 million in annual revenue in eight months and now runs a half-billion-dollar business with 146 people. Polsia lets you type in an idea and, overnight, an AI chief executive plans, codes, markets and supports a company, then emails you a summary in the morning; it crossed $1 million in revenue a month after launch and runs more than a thousand companies at once. Viktor hired itself into two thousand organisations in ten weeks. Every function that Gross once needed a human for, an agent now does at near-zero marginal cost. This is the 1996 moment repeated across engineering, marketing, sales, support, finance and operations, simultaneously.

And underneath the revenue of the tools, the companies they produce are mostly worth nothing. Polsia's own numbers work out to under a thousand dollars of annual revenue per company. Lovable admits most of its 60 million projects are prototypes. Two-thirds of vibe-coded apps ship with security holes. AI-native software companies retain 40 cents of every dollar of revenue a year later, against 82 for ordinary SaaS. The distribution of outcomes has the same shape as the Shopify app store, where the median app makes under a thousand dollars a month, and the App Store, where the top one percent of publishers take 94% of the money. It is the same shape it had in 1999.

Here is what the folklore about 1999 gets wrong, and why it matters. The story everyone tells is that cheap building produced a flood of junk that died. The best research on the period, published in the Journal of Financial Economics, found that 48% of the companies seeking funding in 1999 were still alive five years later, on par with the early auto and television industries, and that survival had nothing to do with how much money a company raised. What killed the other half was strategy: the universal belief that you had to get big fast. Webvan burned $800 million building warehouses. Pets.com never found a business model. Amazon, on the same tools, in the same year, ran its cash conversion cycle like a Swiss watch. Funding did not predict survival. Operators did.

That is the pattern, and it has now run five times. Every time the cost of making something collapses, three things happen in order. Output explodes. Median value stays at zero. And the returns concentrate harder than before, in the small number of people who understand what the new system makes possible before the market does, and who run their loops faster than anyone else. Bezos understood logistics. Page and Brin understood ranking. Evan Spiegel understood the phone camera. MrBeast understood the YouTube algorithm better than YouTube did. None of them won because they could build. Everyone could build. They won because they were a different kind of operator.

The difference this time is the size of the step. Every prior wave cut the cost of one function and left the operator hiring humans for the rest, so leverage was capped by headcount. AI cuts the cost of every function at once. Matthew Gallagher started Medvi with $20,000 and no employees and did $401 million in revenue in his first full year. Midjourney does $18 million of revenue per employee. This is not LeBron James with a better training plan, ten percent stronger. It is LeBron James as a twenty-foot human, playing a sport the rest of the league cannot enter. The best operators are no longer competing with other operators. They are running recursive loops, fifty iterations a week where a normal company runs five, and compounding on every one.

And almost nobody is building for them.

The AI company builders have infinite throughput and no operators. The new AI-native studios have the agents and treat the operator as a hire. The venture funds still give solo-led companies 15 cents of every dollar they raise, even as those companies become a third of all new startups. At the very top of the market, General Catalyst and Thrive have quietly proven the thesis at private-equity scale: buy validated, cash-flowing businesses, install operators and AI, and watch a homeowners-association manager reach $100 million in EBITDA in two years and then buy American Express's travel business for $6.3 billion. Their own analysts say the moat is not the AI. It is "execution speed and operational depth."

There is one precedent that should keep everyone honest. Thrasio also bought only validated businesses, thousands of them, with $3.4 billion. It filed for bankruptcy in February 2024, having paid seven times earnings for what its co-founder now calls vaporware, run hundreds of brands through one central org, and never installed an operator who understood any single one of them. Validated is not the same as selected. Capital was the poison at Thrasio in 2021 exactly as it was at Idealab in 2000. The studio that gets this right will be the one that needs almost no money, because building is free, and spends every dollar it does have on two things: the operators, and the infrastructure they run on.

Rocket Internet's lasting product turned out not to be Zalando or HelloFresh. It was the diaspora of operators who now staff half of European consumer tech. The operator studio's product is the same. The companies are the output. The operators, and the loops they build, are the asset.

That is the opportunity: Gross's 1996 warehouse, rebuilt for a world where the factory floor is free and the only scarce thing left is the person who knows what to make on it.

The studio, as it would be built

A forward narrative. Names, numbers and cadence are placeholders to be argued with.

It starts the way Idealab did: with a founder who has already run the experiment on himself. Steve built a physical consumer brand out of Los Angeles with a team that would have been a rounding error a decade ago, running sales, ops, content, legal and fundraising through AI loops instead of headcount. He is not the plumber who vibe-coded an app. He is the operator who found out, by doing it, that the constraint had moved. The studio is the answer to a question he could not stop asking: if one person can run a company this way, what happens when you find ten of them and give them something already proven to work?

The first thing the studio does is refuse. It does not take ideas. Every week it turns down founders with decks, because decks are the one input the world now has infinite supply of. What it looks for is a product that has already cleared the CarsDirect test, the four cars sold on a Thursday night in 1999: real customers, real money, and, the part Thrasio skipped, real retention. A meal-prep app at $30K MRR that has kept 70% of its customers for a year. A niche B2B tool stuck at $400K ARR because its founder is exhausted. A consumer brand with a loyal base and no growth engine. These are everywhere. Carta's data says a third of new companies are now solo-led and they get 15% of the capital. The AI company builders are minting thousands more every week and abandoning nearly all of them. The studio's deal flow is the long tail that everyone else has priced at zero.

The second thing it does is find the operators, and this is the hard part, because the profile does not exist on LinkedIn. Not an engineer. Not a marketer. Someone who already runs something, however small, at ten times the normal iteration rate, and who understands a customer the way Bezos understood a warehouse. The studio finds them where they are visible: solo founders whose revenue-per-person is absurd, indie operators posting their loops in public, the person inside a company who quietly automated their whole department. It pays them like founders, with equity across the portfolio rather than in one company, so their incentive is the same as the studio's: make every company on the floor compound. Gross hired a CEO per company and took the chairman's seat. The studio does the same, but the CEO comes with a swarm.

The third thing, and the one the AI-native studios have right, is the floor itself. The studio does not rent Lovable seats or run companies inside Polsia. It builds and owns its agent infrastructure: the orchestration layer, the data layer, the shared services agents that Alloy describes running across a portfolio, security reviewed centrally so no company ships the 48-day hole Lovable did. This is Idealab's brick warehouse with the shared back office, rebuilt as software. When a new company comes in, it does not start from zero. It plugs into loops that already run: acquisition, conversion, retention, support, finance, each with an explicit iteration target and a dashboard the operator reviews every morning the way Polsia's founder reads his overnight email.

A day on the floor looks like this. An operator opens the morning report on three companies. Overnight, the agents ran forty ad variants on one, drafted and sent a win-back sequence to lapsed customers on another, and flagged that the third's churn ticked up among one cohort. The operator spends the morning on the one decision the agents cannot make: whether the churn is a pricing problem or a product problem, which requires knowing the customer. By noon the agents are running the test. By the next morning there is an answer. Fifty loops a week, where the company's original founder ran five. The operator is not doing the work. The operator is deciding what the work is, and the deciding is the whole game.

The studio stays small on purpose. Idealab had 26 companies in year one and nearly died of it. The first cohort here is four companies and two operators. Not because ambition is low but because capital discipline is the survival variable in every precedent in this document, and because the model is unproven until the first cohort shows what an operator plus a swarm does to a $30K MRR product in twelve months. If the answer is 3x, the model is interesting. If the answer is 10x, the model is the story of the decade, and the studio raises then, on its own numbers, the way Gross should have in 1999 and did not.

By year three the portfolio is fifteen to twenty companies, each run by an operator who owns a piece of all of them, on infrastructure that has compounded through every loop it has ever run. The exits are mostly M&A, because that is where exits are now, and because a company that runs on the studio's floor at 60% margins is exactly what the AI rollups at the top of the market want to buy. Some companies the studio simply keeps, because a profitable company with no employees does not need an exit.

By year five the companies are not the point. Rocket's lasting product was the diaspora of operators who now run European consumer tech. The studio's lasting product is the operator bench and the playbook of loops, which is also its moat, because a competitor can copy the agents but cannot copy the people who know what to point them at.

That is the studio. Not a factory for ideas, which the world no longer needs. A factory for the one thing the world has never been able to manufacture: judgment, running at machine speed.

1. The claim

AI has collapsed the cost of building a company to near zero. It has not collapsed the cost of building a good one. The scarce input is now the operator: the person who knows the market, the customer, and the systems well enough to point the tools at the right thing and run fast iterative loops on it.

Three things follow:

1. The tools are real businesses; the companies inside them mostly are not. Lovable, Polsia and Viktor all print revenue. The projects, companies and tasks they produce are overwhelmingly prototypes, tests, and sub-$1K-ARR shells.

2. The operator gap is a step function, not a slope. Prior platform shifts cut the cost of one function (publishing, distribution, infrastructure). AI cuts the cost of every function at once. The best operators do not get 10% better; they run a different kind of company. Returns concentrate harder than in any prior wave.

3. The opportunity is a studio that industrializes elite operators on validated products. Not "anyone can build a company." The opposite: ruthless selection of products that already work, paired with a small bench of operators who can 100x them, on infrastructure the studio owns.

This is not a new pattern. Every time the cost of making something has collapsed, output exploded, median value flatlined, and a studio model appeared within a few years to capture the operator premium. Bill Gross built Idealab in 1996 on exactly this logic. Section 5 treats it as the prototype.

2. The build-cost collapse

Building is no longer the constraint. Every layer of company-building now has a tool printing revenue at a speed no prior software category has matched.

The structural signals underneath:

• Solo-founded companies rose from 23.7% of new startups in 2019 to 36.3% in H1 2025 (Carta Solo Founders Report).

• 84% of AI coding tool users in 2026 have no engineering background (Value Add VC).

• A full solo-founder stack runs $300–$500/month against $80K–$120K/month in equivalent salaries (AI Business); solo-built operations report 60–80% operating margins vs 10–20% for traditionally staffed businesses (Grey Journal).

• Bessemer's top "Supernova" AI companies reach $40M revenue by end of year one and $125M by year two (Wealthy Tent, citing Bessemer).

• Anthropic CEO Dario Amodei put 70–80% odds on the first billion-dollar company with a single human employee appearing in 2026 (Grey Journal).

This is Gross's 1996 "90% gross margin if you can make a website" moment, repeated across every function at once.

3. Output explosion, value flatline

The tools work. The companies they produce mostly do not. The distribution of outcomes is a power law that AI is making steeper, not flatter.

Polsia: the per-company number

Polsia's own Product Hunt listing claimed 500+ companies and $450K+ ARR on autopilot (Product Hunt), under $1K ARR per company. Reviewers report the product "does not consistently deliver on that vision," with early users citing task quality, reliability and limited control over autonomous actions, and conclude it is better for experimentation than for workflows where mistakes affect revenue or customers (Salesforge review). One review names the structural gap directly: the validation step Polsia skips is the one every good startup runs before any build tool gets involved (Preuve AI review).

Lovable: churn, traffic and the prototype problem

• Lovable has acknowledged most projects are prototypes or tests (Catalaize).

• Barclays flagged that site traffic fell ~40% from its 2025 peak while ARR quadrupled; Vercel's v0 fell 64% and Bolt 27%. The analysts questioned the economics of month-to-month subscribers. Lovable has never disclosed logo churn, only net dollar retention above 100% (Product Growth, Aakash Gupta).

• Unit economics are fragile: every generation costs inference paid to Anthropic and OpenAI, and Wix's Base44 entered 2026 with non-GAAP gross margin near zero (Visionary Talks).

• Security: a single Lovable-built app exposed 18,000 users' data in Feb 2026; a March 2026 API vulnerability let any free account read other users' source code and credentials, open for 48 days (Product Growth).

Vibe-coded companies: the failure pattern

• 65% of scanned vibe-coded apps carry security vulnerabilities; 58% have at least one critical flaw (Value Add VC).

• 36% of vibe coders skip QA entirely and rely on reprompting instead of debugging (ICSE-SEIP '26, via Museum of Vibe Coding). The documented "vibe coding hangover": a product built fast hits real scale and architectural incoherence makes every fix break something else.

• Veracode's 2025 GenAI Code Security Report found 45% of AI-generated code failed security checks across 100+ models; Replit's agent deleted a production database during a code freeze (Gronkiewicz).

The cause of death has not changed

CB Insights' analysis of 431 failed VC-backed companies found poor product-market fit is the #1 root cause at 43%; "ran out of capital" is the symptom (Preuve AI). Harvard's Shikhar Ghosh found ~75% of venture-backed startups never return cash. Carta recorded 966 shutdowns in 2024, up 25.6%, 74% at pre-seed or seed (Preuve AI).

The capital gap

Solo-led companies were 30% of startups founded in 2024 but received only 14.7% of cash raised in priced rounds (Carta). The market is producing more single-operator companies than the funding model knows how to underwrite. That mismatch is where a studio earns its return.

Read together: the tools moved who can build. They did not move what wins. Exactly as in 1999, when anyone could make a website and eToys still went to zero.

4. Historical precedent

The same pattern has run at least four times since 1995, and the data on each wave is more interesting than the folklore. The folklore says "cheap building produced a flood of junk that failed." The data says something sharper: cheap building produced a flood of plausible companies, most of which survived or failed on the quality of their operators, and returns concentrated in a top percentile that got narrower with each wave.

The internet, 1995–2005: it was never a building problem

U.S. internet usage went from 0.8% of the population in 1990 to 43% by 2000; the Nasdaq rose from 549 in January 1990 to 5,048 on March 10, 2000, then fell 77% to 1,108 by October 2002 (arXiv, Nasdaq bubble analysis). More than 5,000 internet firms failed after the start of 2000 (Springer, survival of public internet firms); over 100,000 dot-com employees lost jobs between October 2000 and July 2001 (strategy+business).

But the most rigorous study of the era, by Goldfarb, Kirsch and Miller at Maryland's Smith School using a random sample of 1999 venture-seeking companies, found something the folklore misses: the five-year survival rate was 48%, on par with the early auto, tire, television and penicillin industries, and survival was unrelated to the receipt or amount of private funding. Their conclusion, published in the Journal of Financial Economics, was that there may have been too little entry, not too much (Smith School). A separate study of ~2,000 B2B e-marketplaces found 55% still active two years after the Nasdaq bottom (strategy+business).

What killed the other half was strategy, not tooling: the "get big fast" doctrine that prioritised growth over unit economics and was near-universal until early 2000 (Smith School). Webvan burned $800M+ building its own infrastructure before shutting in June 2001; Pets.com went bankrupt nine months after its $82.5M IPO with no working model; Boo.com burned $135M in 18 months (Quartz). Amazon, eBay and Google ran the same tools with different operators.

The read-across to 2026 is exact. Funding did not predict dot-com survival; operators did. Today's version: subscribing to Lovable or Polsia will not predict survival either.

Mobile, 2008+: the 1% takes 91–94% of revenue

The App Store carries roughly 2.2–2.5 million apps, ~95% free to download; developers have earned over $550B cumulatively since 2008 (Axis Intelligence). Sensor Tower found that the top 1% of monetizing publishers took 94% of U.S. App Store revenue (Sensor Tower), and across 900,000 publishers on both stores, the top 1% (9,000 publishers) captured 72 billion of 87 billion installs in 1H 2022, about 79% (Sensor Tower). Distribution to a billion pockets was free. The return went to the 1%.

Shopify's app ecosystem: median under $1K/month

The Shopify App Store has ~13,000 listed apps. The top 1% clear $1M+ ARR, the top 10% clear $100K+ ARR, and the median listed app earns under $1,000 per month, with "a meaningful chunk" earning nothing (Week One Labs). Developers have collectively earned $1.5B since inception (Uptek). Note the shape: it is the Polsia distribution (median <$1K) with a fatter head, because Shopify apps have had 15 years for operators to compound.

The Amazon aggregators, 2018–2024: buying validated output without operators

This is the precedent that most directly tests the thesis, because the aggregators did one thing right that this thesis also proposes: they bought only validated, cash-flowing businesses. Thrasio raised $3.4B, was valued at $6B–$10B, and filed Chapter 11 on February 28, 2024, eliminating $495M of debt to emerge as a company "focused on a handful of top-performing brands" (MDS, Aura). The sector raised $6B in 2021; funding fell 88% the following year; Benitago went bankrupt and Apollo sought a buyer for Perch (PYMNTS).

Co-founder John Hefter's post-mortem: Thrasio began buying at 2x EBITDA and, to hit growth targets, was paying 7x for "Chinese vaporware garbage" by the peak (Marketplace Pulse). Creditors asked how the company "lost over $3 billion in value in less than two years" (CNBC). The restructuring backed "a narrower company built around fewer brands," not a return to the playbook (MDS).

Three lessons for an operator studio:

1. Validated is not enough. Every Thrasio brand had revenue. Selection has to test the loop (retention, margin, defensibility), not the top line.

2. Centralised ops without per-company operators fails. Thrasio ran hundreds of brands through one org and could not run "fewer things well."

3. Capital was the poison, exactly as at Idealab in 2000. Cheap money forced volume; volume forced overpaying; overpaying killed the model.

The pattern, stated plainly

1. A platform drops the cost of making by an order of magnitude.

2. Output explodes. Median quality and median revenue stay near zero (Shopify apps: median <$1K/month; Polsia: <$1K ARR per company).

3. Value migrates to what is still scarce: judgment, market knowledge, distribution, and mastery of the new system's loops.

4. Returns concentrate harder each wave (App Store: 1% takes 94%).

5. A studio model appears within a few years to industrialise step 3. The ones that survive select on operators and stay capital-disciplined (Idealab post-2001, Hexa, YC). The ones that die select on assets and raise too much (Idealab 1999–2000, Thrasio, most MCNs).

5. Idealab as the prototype

Bill Gross founded Idealab in March 1996, at the exact moment the web made building cheap, on the premise that ideas were abundant and execution capacity was the bottleneck. Thirty years later it is the longest-running studio in the world, and its record is the best available evidence for both halves of this thesis: the operator premium is real, and a studio that forgets it gets destroyed.

The model

Idealab was built as a company factory, not a fund. Gross himself came up with most of the ideas, Idealab put in the first capital, hired a CEO for each company, and Gross took the non-executive chairman seat. A shared back office handled HR, accounting and infrastructure for every company in its early stage, so each new venture started with an operating spine rather than building one (Fuhrman, Seeking Alpha, 2016). Within its first year it had 26 businesses in incubation, from three-person teams to CitySearch, which already employed over 1,000 people across 14 cities (University of Michigan case study, 1997).

The thesis in Gross's own words: Idealab made only internet companies at first because it was "an incredible new medium that had unbelievably high gross margins. If you could make a Web site, you could sell something online and you could make margins of 90 percent or higher" (TechCrunch, 2007). That is the 1996 version of "anyone can build an app now."

The record

A 30% exit rate across 150 companies is extraordinary against a base rate where roughly 75% of venture-backed startups never return cash. Note what the winners have in common: GoTo.com invented paid search, the auction-based model that directly inspired Google AdWords in 2000; CitySearch established the local-directory category and merged with Ticketmaster Online in 1998; NetZero pioneered ad-supported free dial-up (Grokipedia). These were not "anyone could build this" companies. They were new mechanisms run by operators who understood the medium better than the market did.

What broke: the 2000 collapse

Idealab is also the cautionary tale. In December 1999 and January 2000 it raised $725M in preferred stock from outside investors, then targeted its own IPO. By October 2000 the IPO was stalled (CNN Money, Oct 2000). By January 2001 Eve.com, Modo.com, Myhome.com and Swap.com had all folded, and eToys had missed its December sales targets and laid off 70% of staff. A former employee of one failed company told Forbes: "they spent so much money immediately, it was out of the picture to buy a brand. We ran out of cash and couldn't support the sites that existed" (Forbes, Jan 2001).

eToys is the clearest case. IPO on January 4, 1999 at $20, closed the first day at $76, was called "the benchmark against which all other toy sites are measured," and went bankrupt on April 1, 2001. KB Toys bought the remaining assets for $5M (Wikipedia).

By December 2002 the preferred investors were suing to liquidate Idealab, alleging it had been used as a "private piggy bank," with the company valued at around $380M against their $725M (Forbes, Dec 2002). Idealab survived, stopped raising outside money entirely, and pivoted to "atoms" businesses: solar, robotics, 3D printing (Marketplace).

What Gross concluded

After the crash Gross ran a systematic review of 100 Idealab companies and 100 outside startups to find what separated success from failure. His conclusion: "I used to think the idea was everything but, over time, I came to think that maybe the team, the execution and the adaptability mattered even more than the idea" (Gross, TED, via CB Insights). At Stanford he reduced failure to two causes, team problems and running out of money, and said the most successful entrepreneurs are those with the best execution and the most persistence (Stanford eCorner).

His operating lessons from the 40 failures, per First Round Review:

• Every company needs four skill sets in large doses: entrepreneur (idea), producer (ships and sells), administrator (process), integrator (people). A company with only the entrepreneur burns out; without the administrator, "the wheels come off through growth."

• Many failures were products one to two years ahead of their market; staying lean until the market ripened would have saved them.

• Test the core proposition as cheaply as possible before building anything. CarsDirect had 90 days and $80K to find out if anyone would buy a car online. It sold four cars the first night, at a $4,000 loss each, and shut the site off. That company later sold to Hellman & Friedman for $600M.

• Decide fast: 80% right and quick beats 100% right and slow.

• Fire non-A players immediately.

What transfers to an AI-operator studio

1. The shared spine works. Idealab's central back office let each company start with operating infrastructure. The 2026 equivalent is owned agent infrastructure: the studio's own orchestration, data and tooling layer rather than rented seats on Lovable or Polsia.

2. Idea-first was the fatal bias. Gross generated most ideas himself and funded them before validation. The crash companies (Eve, Modo, Swap, eToys) were plausible ideas with no proven demand and expensive execution. Gross's own post-mortem moved him to execution and team. This thesis takes that one step further: select only products that have already cleared demand, then apply the operator.

3. The four-role model collapses into one operator plus agents. In 1996 the producer, administrator and integrator had to be separate hires. In 2026 an elite operator with agents can cover producer and administrator, which is exactly why the operator's judgment becomes the whole variance.

4. Capital discipline is the survival variable. Idealab nearly died from raising $725M at the top and spending it into unvalidated companies. Gross never raised outside money again. A studio built on cheap AI execution should need far less capital per company, and should treat that as the point, not as room to spend.

5. The CarsDirect test is the selection filter. 90 days, $80K, does the customer pay. With AI the same test costs days and hundreds of dollars, which means the studio can run it on every candidate before committing an operator.

Rocket Internet: execution-only, on validated models

If Idealab proved the studio can generate and operate ideas, Rocket Internet proved the opposite half of this thesis: that execution alone, applied to models already validated elsewhere, is enough to produce multiple public companies. It is the closest historical analogue to "validated products + industrialised operators."

The Samwer brothers' first company, Alando, an eBay clone, sold to eBay in under 100 days in 1999. From 2007 Rocket turned that one-off into a repeatable clone-localise-scale system, and over the following decade launched 100+ companies in 100+ countries (Startuprad.io, Netfigo). Its Berlin headquarters grew to thousands of employees providing shared services across the network, with standardised tech stacks, centralised recruitment and product and marketing templates that cut time-to-scale and initial burn (GrowthShareMatrix, Business Model Canvas Template).

Sources: Netfigo, Business Model Canvas Template.

Rocket itself IPO'd in October 2014 at €42.50 (€6.7B), delisted in October 2020 at €18.57, and was then forced by Elliott Management into a €35-per-share self-tender in 2022 (Startuprad.io). Its lasting output, per the same analysis, is "the scaled companies and operators it produced" and "a diaspora of founders and funds": the Zalando, Delivery Hero and HelloFresh alumni network now staffs Wolt, Flink, Getir, Choco and most of European consumer tech (The Big Byte). Rocket's real product turned out to be operators.

Where the clone model broke: when local competitors knew the market better, or when the original entered directly (Wimdu vs Airbnb) (Netfigo). Execution on a validated model still loses to a better operator on the same model. That is the thesis's own constraint stated as a Rocket failure.

What Idealab and Rocket prove together

Other studios refined one piece each: Hexa/eFounders (2011, Paris) proved a narrow lane (B2B SaaS only) with a repeatable playbook produces serial unicorns (Front, Aircall, Spendesk); Atomic (2012) and Science (2011) proved the hired-CEO model in consumer (Hims, Dollar Shave Club); Flagship Pioneering proved a studio can own the underlying platform (Moderna). Each is a partial answer to the question Idealab first posed in 1996: when building is cheap, what is the studio's actual asset?

6. The step function

The elite operator is not 10% or 50% better than the average one. The gap is a step function, because AI removes the cost of every function simultaneously and the gain compounds through iteration rate.

Why this wave is different

Every prior platform cut the cost of one function. The web cut publishing. AWS cut infrastructure. The App Store cut distribution. Shopify cut storefronts. In each case the operator still had to hire humans for everything else, so leverage was capped by headcount and hiring speed.

AI cuts engineering, marketing, sales, support, finance, ops and research at once. Viktor's own framing is the tell: "You stop doing the work and start reviewing it" (Viktor). When execution across every function is a reviewing job, the only remaining variable is the quality and speed of the reviewer's judgment.

The compounding mechanism

A company is a set of loops: acquire, convert, retain, learn, ship. The operator's output is the number of high-quality iterations per week across those loops.

• An average operator with AI runs maybe 2–3x more iterations than before, because they use the tools as faster typing.

• An elite operator redesigns the loops themselves so agents run them continuously and the human only intervenes at decision points. Polsia's nightly cycle (evaluate, decide, execute, report by morning) is a crude version of this (Teamday).

If one operator runs 5 iterations a week and another runs 50, and each iteration compounds at even 2%, the gap after a year is not 10x. It is closer to 10x in rate multiplied by the compounding, which is why revenue-per-employee figures now span two orders of magnitude between Midjourney ($18M) and a conventional SaaS company (~$200K–$300K).

The LeBron framing

Giving LeBron a better training program makes him 10% better. Giving him a new supplement makes him 20% better. That is what prior tools did for operators. This wave is LeBron becoming a 20-foot human: the rules of the game he plays no longer apply to anyone else on the court. The best operators are not outcompeting other operators. They are running a different sport.

Evidence of operators at the top of the step

What these have in common is not the tools. Everyone has the tools. Each of these operators picked a narrow market they understood, built the loop, and ran it faster than anyone else in that market could.

The iteration model

The gap is not output per cycle. It is cycles per unit time, compounded — which is why a linear gain in tooling produces a non-linear gap in outcomes.

Model a company as a product that improves by a factor of g per learning cycle, with an operator running r cycles per year:

V_A(t) / V_B(t) = (1 + g)^{(r_A − r_B)t}

The exponent carries the difference in cycle rate, so the gap widens geometrically in time even when g is small and identical for both operators. At a modest g of 3% per cycle, an operator running 50 cycles a year is 4.4x better after year one against 1.2x for an operator running five — a 3.8x gap. By year three it is 84x against 1.6x, a 54x gap. Nothing about the operator's talent changed in that model. Only the rate did.

Three findings say the rate is genuinely moving, and that it moves for some people and not others.

The unit of autonomous work is doubling roughly every seven months. METR measured the length of software task a frontier model can complete unattended at 50% reliability across six years of models and found a doubling time of about seven months (METR, March 2025). That is the input to r rising without headcount rising.

But adoption alone does not raise r. Google's 2025 DORA study of nearly 5,000 technology professionals found 90% using AI at work and more than 80% believing it made them more productive. AI adoption correlated positively with delivery throughput — and still correlated negatively with delivery stability. The gains landed in teams with loosely coupled systems and fast feedback loops; teams without them saw little or nothing (DORA, 2025). The report's own summary is the operator-gap claim in someone else's data: AI amplifies what is already there.

And the operator cannot feel the difference. METR's randomized trial put 16 experienced open-source developers on 246 real issues in repositories they already maintained. With AI tools they were 19% slower. They had expected to be 24% faster, and after finishing they still believed AI had sped them up by about 20% (METR, July 2025).

That last result is usually read as evidence against AI leverage. It is the strongest evidence for the operator gap. It shows that raw tool access does not raise the cycle rate, that it can lower it, and that the person holding the tool has no reliable sense of which is happening. The model above has a floor nobody talks about: if g is zero, iteration is noise, and a faster loop just produces noise faster. Positive g requires a measurement discipline — knowing what to test, what counts as a result, and when to kill a direction — that is an operator property, not a tool property.

So the correct form of the claim is narrower than "AI makes builders faster." It is: AI raises the ceiling on r, and the operator determines how much of that ceiling gets used and whether g stays above zero. That is a selection problem, which is what a studio is for.

Caveat on the model: it assumes improvements compound rather than substitute, which holds while a product is being found and breaks once a company's constraint moves to distribution, regulation or capital. Two of the case studies below hit exactly that wall.

Operator case studies

Four companies reached outcomes that used to require a full team. In each, the operator already had the domain before the tools arrived.

The baseline matters. WhatsApp ran 450 million users on 32 engineers and sold for $16 billion in 2014, so extreme output per head predates generative AI. What changed is the clock: Stripe's data on the top 100 AI companies puts the median time to $1M annualized revenue at 11.5 months (nazr). The ceiling did not move as much as the time to reach it.

What the four have in common is the part the company-builder pitch leaves out. Maor Shlomo was a repeat founder with a prior data company before Base44. Matthew Gallagher was a performance marketer who already knew direct-response acquisition before he touched a model. Cursor and Lovable were built by engineers on the frontier of the exact problem they were selling into. None of them is a first-time operator who was handed a builder and produced a company. The tool converted existing operator judgment into output at a rate that used to be impossible. It did not supply the judgment.

Medvi is also the cautionary case, and the thesis is stronger for saying so. Its revenue figures are self-reported and unaudited. Since early 2026 the company has drawn an FDA warning letter alleging misbranding (February 20, 2026), a California class action over affiliate spam (March 20, 2026), reporting by Business Insider identifying more than 800 apparently fake doctor accounts used in its advertising, and earlier reporting that before-and-after photos on its site were AI-generated (Fortune). Gallagher disputes that the FDA letter applies to his primary site.

Read correctly, that is not a counterexample — it is the thesis. A two-person company running a very high cycle rate in a regulated category compounded its distribution edge and its compliance exposure at the same speed. Leverage is symmetric. It amplifies the operator's judgment, including the parts of it that are wrong, which is the argument for selecting operators rather than arming them.

All revenue figures except the Base44 acquisition price are company-reported and unaudited.

7. The studio model

The studio is a selection engine plus an operator bench plus owned infrastructure. It is not an idea factory and not a fund.

The current field (all launched or reframed since 2024)

What they have in common: they are all throughput plays. Build many things fast, on the bet that AI-native execution wins. None of them centers the operator as the asset, and none publishes per-company revenue.

The wider landscape: five categories

The field is bigger than the "AI venture studio" label. Five models are competing to capture the same shift, and they differ on two axes: what the input is (ideas vs. validated businesses) and what the asset is (tools vs. operators).

Where this thesis sits

The operator-first studio sits in the empty cell: validated input + operator asset + studio-scale capital.

• The AI company builders have the throughput and none of the operators.

• The AI-native studios have the agents and treat operators as hires, not the asset.

• The AI rollups have the operator + AI thesis exactly right, but at private-equity scale: $100M+ acquisitions of services firms, backed by $1B vehicles, with the AI stack often supplied by a lab (OpenAI inside Thrive). One analysis of the GC portfolio puts the moat plainly: it is "distribution and proprietary data, not the AI itself," and the competitive advantage is "execution speed and operational depth" (Capital & Clarity). That is the operator thesis, proven at the top of the market.

Nobody is running the rollup logic at the bottom of the market: products and small companies at $10K–$1M ARR, too small for GC or Thrive, too far along for Forum or IAIG, and starved of capital per Carta. The combination that does not yet exist is best-in-class agent infrastructure (the AI-native studios' asset) + real operators as the core asset (the rollups' insight) + validated-only selection (the rollups' input) + capital-light checks (the studio's scale).

Two cautions from the neighbors. ChartMogul's analysis of 3,500 software businesses found AI-native companies posting a median net revenue retention of 48% and gross retention of 40%, against 82% NRR for conventional SaaS, the "AI tourist effect" of curiosity sign-ups that never activate (SaaS Library). Validation has to mean retained revenue, not sign-ups. And the rollup analysts note that none of the GC or Thrive platforms has operated through a recession and the margin claims are self-reported (Capital & Clarity). The operator + AI thesis is early everywhere, including at the top.

The operator-first design

1. Selection: validated products only. The studio does not take ideas. It takes products that have already cleared the CarsDirect test: someone has paid. Sources of deal flow: solo founders who have hit $10K–$100K MRR and stalled; products inside Lovable/Polsia-style platforms that show real retention; small businesses with a working loop and no leverage. This is the population Carta shows is starved of capital.

2. The operator bench. A small number of people, not a large team. The profile is not "technical" or "non-technical"; it is someone who already runs a business on AI at 10x the normal iteration rate and understands the customer. Comp is equity-heavy across the portfolio, so an operator is paid like a founder, not a consultant. This is the Idealab CEO-hire model with the hire pool redefined.

3. Owned infrastructure. The studio runs its own agent orchestration, data layer and tooling. Renting seats on Lovable or Polsia means renting a zero-margin product with undisclosed churn and a security record. The Idealab shared back office, rebuilt as agents.

4. The loop as the unit of work. Each company is run as a set of recursive loops (acquire, convert, retain, iterate) with an explicit target iteration rate. The operator's job is to raise that rate. This is the mechanism behind the step function in Section 6.

5. Capital. Small checks, because execution is cheap. The Idealab failure mode was $725M raised at the top and spent into unvalidated companies. The AI studio's advantage is that it should need almost none of that.

6. Portfolio cadence. Few companies at a time per operator, with the portfolio growing only as operators are found. Idealab's 26 companies in year one is the wrong number; Forum's 20 over several years with 63% follow-on is closer.

Section 8 covers why the capital structure follows the operator team. Operator sourcing and the deal-flow pipeline — where the studio finds people with a proven hit rate, and on what terms — remain the open design questions.

8. What this does to venture capital

Venture capital is already consolidating into fewer, larger managers. The open question is not whether the industry shrinks in firm count — that is happening now — but what the surviving capital buys. The argument here is that it stops buying single companies and starts buying operator teams that ship many.

The consolidation is measured, not predicted

US VC firms closed $66.1B across 537 funds in 2025, and $62.4B across 288 funds in the first five months of 2026 alone (PitchBook-NVCA, via Value Add VC). More money, far fewer vehicles.

The distribution inside that is the real story. Twelve firms took more than half of all US venture capital raised in the first half of 2025; the top 30 took 74% (PitchBook). By Q1 2026, 90.9% went to established firms, and funds over $1B now absorb 68.3% of every dollar raised, with 12 firms taking close to three-quarters of all commitments (PitchBook, Q2 2026). Founders Fund's $4.6B in 2025 was more than the $1.8B raised by all 44 first-time managers combined. Emerging-manager vehicles fell to 177 closes in 2025, the fewest since 2015, and the median fund now takes about 15 months to close — the longest in over a decade.

Deployment concentrated the same way. $340B went into US VC-backed companies in 2025 while deal count fell 15%; the top 1% of companies took a third of all capital and the bottom half took 7% (SVB State of the Markets 2026). Only about 3% of seed companies now reach Series A within 12 months.

The arbitrage

Put those two facts side by side. LPs are allocating as though finding out whether a company works still costs $5–10M and takes three years. Base44 found out in six months for nothing. The price of a shot on goal collapsed; the capital structure that funds shots on goal did not reprice — it got more concentrated and slower.

That is the gap a studio sits in. The measured studio numbers are directionally supportive: the Global Startup Studio Network's study of 258 studio-created startups reported 53% IRR against 21.3% for traditional startups, 5.8x MOIC against 1.57x, zero-to-seed in 10.7 months against 36, 72% converting seed to Series A against 42%, and exits at an average company age of 3.85 years against 6.6 (GSSN).

Those figures need a health warning, and the thesis should carry it openly: the study was published in 2020, the sample is self-selected and self-reported, it predates this tooling wave entirely, and studios that folded are unlikely to have answered a survey. Treat it as the direction of an effect, not its size.

Why the fund follows the team

The portfolio math changes when both the cost and the duration of a shot fall. A seed fund holding 15–20% needs roughly one in twenty companies to return the fund. A studio holding 30–50% at formation, on a cost basis one order of magnitude lower, needs closer to one in ten — and gets that ratio from an operator whose hit rate is the thing being underwritten rather than a founder met once at a demo day.

Underwriting shifts from a company to a person with a track record across attempts. That is closer to how LPs already underwrite a GP than to how a seed fund underwrites a founder, which is why the two structures converge.

What actually shrinks

Not venture capital. The middle of it.

The claim to defend is therefore not "VC shrinks." It is that the layer whose edge was picking one company at a time gets compressed between capital that funds infrastructure and capital that funds operators. A fund that backs a team shipping ten companies is not a new asset class. It is what a seed fund becomes when selection at the company level stops being a source of edge.

9. Risks and open questions

The thesis fails if any of the following turn out to be true. Each is listed with what would confirm it.

What would falsify the thesis

• Polsia or a peer publishes per-company data showing a meaningful fraction of AI-run companies reaching real revenue without a human operator.

• Revenue-per-employee dispersion narrows over the next two years instead of widening.

• A throughput-first studio (IAIG, Alloy) produces multiple $10M+ companies with no identifiable elite operator behind each.

Open questions

• What is the right equity split between studio, operator and original founder?

• Where does deal flow come from at scale: platforms, solo founders, or acquisition?

• Can operator skill be measured before committing a company to someone?

• Is the studio's product the companies, the operators, or the infrastructure?

10. Sources

Idealab and Bill Gross

• First Round Review: Lessons Learned from Bill Gross' 35 IPOs/Exits and 40 Failures

• Peter Fuhrman: Bill Gross, America's "Mr. Unicorn" (Seeking Alpha, 2016)

• University of Michigan case study: idealab (1997)

• Forbes: Incubators on Life Support (Jan 2001)

• Forbes: Angry Investors Seek to Liquidate Idealab (Dec 2002)

• CNN Money: Idealab IPO stalled (Oct 2000)

• TechCrunch: Bill Gross Is Into Atoms (2007)

• Stanford eCorner: 2 Reasons Companies Can Fail

• Wikipedia: eToys.com

• Grokipedia: Bill T. Gross

AI company builders and AI employees

• TechCrunch: Lovable hits $500M annualized revenue (Jun 2026)

• Product Growth: How Lovable hit $100M ARR in 8 months

• Visionary Talks: Lovable revenue, dated

• Catalaize: How Lovable reached 2.3M users

• Aakash Gupta on Lovable churn (X, Mar 2026)

• Tim Frin: How Polsia builds and runs companies with AI agents

• Teamday: Polsia solo founder hits $1M ARR

• Product Hunt: Polsia

• Salesforge: Polsia review

• Preuve AI: Polsia review

• EU-Startups: Viktor raises Series A after $15M run rate in 10 weeks

• Accel: Welcome, Viktor

• Viktor: Series A announcement

One-person companies and solo founders

• Carta: Solo Founders Report 2025

• BuildMVPFast: One-person unicorn 2026

• Wide Journal: AI solopreneur 2026 (Medvi)

• Grey Journal: 7 solo founders building $1M+ AI businesses

• AI Business: Where solo founders hit a wall

• Forbes: OpenAI called the one-person AI startup (Apr 2026)

Failure data

• Preuve AI: Startup failure statistics 2026

• Museum of Vibe Coding: Vibe coding for startups (May 2026)

• Gronkiewicz: Why vibe-coded startups fail

• Value Add VC: Vibe coding for non-technical founders

AI venture studios

• Alloy Partners: Inside the first agentic venture studio

• AI Insider: IAIG raises $6M pre-seed (Aug 2026)

• Forum Ventures AI Studio

• TFSF Ventures ranking (press release, Mar 2026)

• Shape: What is an AI venture studio

AI rollups and the wider landscape

• Capital & Clarity: The General Catalyst behind $1.5B of AI roll-ups (Apr 2026)

• Capital Founders: Founder's guide to AI-enabled roll-ups (Jul 2026)

• Newcomer: Inside the VC roll-up craze (Dec 2025)

• L40: AI rollups in 2026

• AIMultiple: AI rollups funding and trends

• Waveup: Top 10 venture studios 2026

• SaaS Library: B2B SaaS trends 2026 (ChartMogul NRR data)

Historical precedent (Section 4 and Rocket Internet)

• Smith School, University of Maryland: Research sheds new light on dot-com bust (Goldfarb, Kirsch, Miller)

• strategy+business: Lessons of the Last Bubble (2007)

• arXiv: Dissecting the dot-com bubble in the 1990s NASDAQ

• Springer: Business model characteristics and survival of public internet firms

• Quartz: 7 companies that died in the dot-com bubble

• Sensor Tower: 94% of U.S. App Store revenue from top 1% of publishers

• Sensor Tower: Top 1% of publishers, 1H 2022

• Axis Intelligence: App Store statistics 2026

• Week One Labs: Shopify app revenue benchmarks 2026

• Marketplace Pulse: Death by Valuation, the Amazon aggregator autopsy

• CNBC: Thrasio loses CEO in bankruptcy (Apr 2024)

• MDS: What happened to Thrasio

• PYMNTS: Thrasio's bankruptcy spotlights aggregators' decline

• Startuprad.io: Rocket Internet and Oliver Samwer

• Netfigo: Oliver Samwer profile

• The Big Byte: Delivery Hero and HelloFresh leadership networks

Iteration rate and developer productivity

• METR, Measuring AI Ability to Complete Long Software Tasks — task time-horizon doubling every ~7 months (March 2025)

• METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — RCT, 16 developers, 246 issues, 19% slowdown (July 2025)

• Google Cloud / DORA, 2025 State of AI-assisted Software Development — ~5,000 respondents; throughput up, stability down

Operator case studies

• TechCrunch, 6-month-old solo-owned Base44 sells to Wix for $80M cash (June 2025)

• nazr, How many employees does it take to reach $100M ARR now? — Lovable, Cursor, Gamma headcounts; Stripe time-to-$1M data

• Forbes, AI and $20,000 helped one man build a $1.8B telehealth startup (April 2026)

• Fortune, A $1.8 billion startup with just 2 employees — now the allegations are piling up — FDA letter, class action, fake-endorsement reporting

Venture capital structure

• PitchBook, 12 firms collected over 50% of all venture cash in the first half of 2025

• Value Add VC (PitchBook-NVCA data), Emerging Manager VC Funds 2026 — fund counts, established-firm share, mega-fund share

• SVB, 2026 State of the Markets, via SaaStr — deployment, deal count, capital concentration by company

• GSSN, Disrupting the Venture Landscape — 258 studio startups; IRR, MOIC, time-to-milestone (2020)

• PitchBook, Q2 2026 US VC Fundraising and Returns Report — mega-fund share, firm concentration

Still unsourced: Hexa, Atomic and other studio economics in Section 7; GSSN figures are 2020, self-reported and pre-AI, and should be re-cut against post-2024 studio data before this goes to anyone external.