In any gold rush, the first to get rich are always those selling shovels and water.

As we look back at the AI commercial wave at the end of 2025—cutting through the surface glamour and noise—we’re startled to discover this truth: the shovel-sellers have already struck it rich, while the gold-diggers haven’t even found the vein.

While the world fixates on OpenAI’s next-generation model or NVIDIA’s next chip, a quiet redistribution of wealth and power is unfolding deep within the AI tide.

The most surprising story of 2025 is this: the giants striving to build “Artificial General Intelligence” (AGI) still haven’t found a reliable business model. Yet the companies supplying them with training data, writing scoring rubrics, and building test environments—the ones once dismissed as “doing the dirty, grunt work of AI”—are now raking in enormous profits.

In the shadows of these tech titans, a wealth-creation boom reminiscent of the 19th-century Gilded Age is quietly underway. A cohort of people in their twenties, armed with startups that didn’t even exist a few years ago, is rocketing into the billionaire club at an unprecedented pace. They rarely make headlines; you’ve probably never heard their names—Mercor, Surge AI, Handshake, Scale AI…

These companies are generating wealth at a historically rare speed: valuations in the tens of billions, revenues doubling overnight, founders becoming the youngest billionaires ever. This is the shovel business hidden beneath the gold mine—and it may be redefining who the real winners of the AI revolution truly are.

This article offers a comprehensive overview of the little-known winners of this AI gold rush and delves into the more concealed flows of wealth behind this technological revolution. You’ll see something highly counterintuitive: the new winners aren’t the spotlight-seeking prophets, but the “infrastructure engineers” rooted in real-world needs. They’re redefining value creation and capture at a breathtaking pace and in unexpected ways.

Finally, we’ll distill five defining traits of these AI-era winners and reveal five surprising insights about the future of technology, the essence of business, and the restructuring of societal division of labor hidden beneath this wealth feast.

When Teenagers Start Running Their Own AI Companies

01 Undercurrent: Data Deserts and the Gold Rush

In 2018, when Google researchers published a paper titled “Everyone Wants to Do the Model Work, Not the Data Work,” they exposed a harsh reality: in AI development, data preparation was seen as “an unfortunate but necessary chore” to be done as quickly and cheaply as possible.

Seven years later, everything has flipped.

Data is no longer just fuel—it’s a scarce strategic resource. Frontier AI labs have already “sucked dry” all easily accessible text, images, and code from the internet. The era of breakthroughs simply by scaling up data volume is over.

The real bottleneck has emerged: to make AI genuinely useful in specialized domains—like law, healthcare, or finance—it no longer needs massive generic datasets, but small, high-quality, expert-curated data tailored to specific tasks.

According to internal industry estimates, AI labs alone spent over $10 billion on training data in 2025, with the vast majority coming from just five or so companies. Paradoxically, these big spenders haven’t yet monetized AI themselves—but their data suppliers are already reporting stable profits.

“It’s human expertise from every corner of the world,” one practitioner described the current demand. From nuclear engineering to animal training, from medical diagnosis to carpentry—any human skill that can be codified into a “scoring rubric” has become hot commodity.

And new billionaires are being minted from this trend.

Peak Report: What Kind of Wealth Transfer Is Stablecoin Really Enabling?

02 Rise of the New Elite: The Lightning Strike of Data Giants

· Mercor: $10B Valuation at Age 22

In 2019, 19-year-old Brendan Foody and two high school friends founded Mercor with a simple idea: help startup friends hire software engineers overseas. The platform launched in 2023, using AI to screen résumés and conduct interviews, quickly hitting $1 million in annual revenue.

The turning point came in early 2024. Scale AI—the then-star of AI data labeling—approached them: it needed 1,200 software engineers. Scale needed these engineers to produce code for training data, as OpenAI, Anthropic, and others raced to teach AI how to code.

Foody sensed a bigger opportunity. When the recruited engineers began complaining about Scale withholding wages (the company was being sued in California for wage theft), he decided to cut out the middleman. Mercor started dealing directly with AI labs, offering expert-level data services.

In September 2025, Foody announced Mercor’s annualized revenue had hit 500million,callingit“thefastest?growingcompanyinhistory.”Itslatestfundingroundvaluedthecompanyat10 billion. At just 22, Foody and his co-founders became the youngest self-made billionaires ever.

· Surge AI: The Silent Industry Dominator

Unlike Mercor’s high-profile rise, Surge AI’s ascent happened almost in silence. Founder Edwin Chen, formerly of Google, Twitter, and Facebook, was shocked by the poor quality of data from existing vendors—riddled with errors from non-experts working for pennies.

He decided to build his own solution. Surge created the Data Annotation Tech platform, characterized by smaller scale, sharper focus, and stricter quality control. It pays around $30/hour—far above industry standards.

The strategy worked. Surge has been profitable since inception, and in 2024 its revenue surpassed 1billion,exceedingScaleAI’sreported870 million. Chen still owns roughly 75% of the company. Reuters reported that Surge is now seeking a 1billionraiseata15 billion valuation.

· Handshake AI: Revaluing the University Network

When Garrett Lord founded Handshake in 2014, his goal was to build a “Linkedin for college students.” Over 1,000 university career centers paid to join, and companies could recruit from a network of 20 million alumni and graduate students.

In early 2025, Lord noticed legacy data vendors scrambling for “experts”—and realized his platform already housed a vast pool of highly educated talent. He quickly incubated Handshake AI internally.

The pivotal moment came in June 2025: Meta poached Scale AI’s CEO and acquired a 49% stake. Competing labs, fearing Scale was no longer neutral, rushed to find alternatives. Demand flooded into Handshake AI.

“It was like someone smashed a billion-dollar pi?ata over every data startup’s head,” Lord said. Demand tripled overnight. By November, Handshake AI’s annual revenue exceeded $150 million—surpassing its decade-old core business.

“We grew from 3 to 150 people in five months,” Lord recalled. “One Monday, we onboarded 18 people. We ran out of desks.”

Venture capitalists and tech founders are rushing to catch up—this person’s value in the AI era is skyrocketing.

03 Application Breakthrough: From Tools to Titans

If Mercor and Surge are the “shovel sellers,” another class of companies is trying to build entirely new machines with AI. They’re not feeding models data—they’re using models to reshape entire vertical industries.

In 2025, these “AI-native applications” shattered the old order of SaaS (Software-as-a-Service), solving domain-specific pain points and achieving astonishing valuation leaps.

· Cursor: The Alchemists of Code

The brightest star in this category is undoubtedly the founding team of Cursor (parent company Anysphere). Michael Truell, Sualeh Asif, and others are MIT alumni who only graduated in 2022.

Their product, Cursor, is essentially an AI-assisted code editor—but by 2025, it had become an indispensable tool for programmers. More importantly, Cursor enjoys a unique “data flywheel”: the code users generate while using it is precisely the “golden data” AI labs crave most. Reports say both OpenAI and xAI have expressed acquisition interest.

In November 2025, after a 2.3billionfundinground,Cursorreachedastaggering27 billion valuation. The net worth of its four young founders surged collectively into the billionaire ranks.

In San Francisco, 30-year-old Winston Weinberg still shares an apartment with his co-founder Gabe Pereyra.

Despite living like ordinary workers, their balance sheets tell a different story. Their AI legal software startup, Harvey, underwent three funding rounds in 2025, with its valuation soaring from 3billionatthestartoftheyearto8 billion by year-end.

Faced with astronomical wealth, Weinberg remains remarkably grounded: “Sure, there are billions on paper—but it’s still just paper wealth.”

· Perplexity: The Wisdom Engine Challenging Google

In search, Perplexity is shaking Google’s foundation. Its 31-year-old CEO, Aravind Srinivas, founded this AI search engine company in 2022.

Unlike traditional search, Perplexity aims to deliver direct answers instead of links. By the end of 2025, its valuation stood at approximately $20 billion. Interestingly, Srinivas claims he doesn’t care much about money, preferring a simple life: “Seeking wisdom is far more important than seeking wealth.”

These three companies represent another path to AI commercialization: not infrastructure, but super-applications. They prove that if you dominate a vertical deeply enough, AI can generate value rivaling the oil giants of old.

04 The New Billionaires: Young, Fast, Homogeneous

A December 2025 New York Times report revealed an even more startling phenomenon: founders of these data companies are joining the billionaire ranks at an unprecedented speed.

· Age Records Shattered

Mercor’s three founders, aged 22, became the youngest self-made billionaires in October 2025 after their funding round. Cursor’s founding team, around 25, saw their company valued at $27 billion in November 2025.

· Speed of Wealth Creation Breaks All Records

For contrast: Elon Musk took years to go from millionaire to billionaire. Many of these AI founders went from zero to $1 billion in net worth in just 2–3 years.

Take Mira Murati, 37, who founded Thinking Machines Lab in February 2025; by June, it was valued at 10billion—withouthavingreleasedanyproduct.IlyaSutskever’sSafeSuperintelligencehita32 billion valuation in 2025, also pre-product.

· Highly Homogeneous Group

These new billionaires show striking homogeneity: overwhelmingly male, mostly from elite schools (MIT, Stanford, etc.), concentrated between ages 20–40.

University of Washington professor Margaret O’Mara warns this boom is amplifying “homogeneity”—people of similar backgrounds, ages, and genders now occupy the top of the wealth pyramid.

The influencer era is ending, but everyone can still succeed at the top of their own ladder.

05 Industry Earthquake: From Body Shops to the Expert Economy

The transformation in the data industry reflects a fundamental shift in AI training paradigms.

Early data work—like labeling images for autonomous driving—was labor-intensive and low-skill. Amazon Mechanical Turk paid thousands of workers mere cents per image to label cats and dogs. Scale AI’s Remotasks platform had workers draw boxes around stop signs and traffic cones.

ChatGPT changed the game. Its use of Reinforcement Learning from Human Feedback (RLHF) required evaluating answer quality—not just simple labeling. Judging medical advice requires medical training; assessing legal arguments demands legal knowledge.

This gave birth to the expert data economy. It’s no longer about “body shops” stacking cheap labor, but sourcing experts with advanced degrees and professional credentials.

Surge AI boasts Fields Medal-winning mathematicians, Supreme Court litigators, and Harvard historians. Mercor highlights its Goldman Sachs analysts and McKinsey consultants. Handshake AI mobilizes PhDs, master’s students, and professionals from over 1,000 universities.

· The Rubric Arms Race

AI companies discovered that to truly master complex skills, models need extremely detailed “scoring rubrics”—breaking “doing well” into dozens of measurable criteria.

Creating one rubric can take over 10 hours and include more than ten distinct standards. OpenAI’s medical benchmark includes nearly 50,000 criteria, with different combinations applied to different prompts. Insiders say labs order tens of thousands to hundreds of thousands of rubrics per training run—totaling millions of individual standards.

“It needs to be hyper-granular,” explains Mercor’s Foody. To build a consulting rubric, you start by creating a taxonomy of all consulting industries, then subtypes of consulting within each, then report formats for each subtype.

· Explosive Demand for Environments

Rubrics alone aren’t enough—you also need “gyms” for AI to practice: reinforcement learning environments.

These vary wildly: “Sometimes they’re clones of DoorDash or Salesforce, but often they’re custom enterprise environments,” says Alex Ratner, CEO of Snorkel AI, whose company originally built annotation software but now also offers synthetic data services.

Demand for both environments and rubrics is surging simultaneously. AI firms are attempting brute-force approaches—hiring thousands of lawyers, consultants, and specialists to encode human knowledge into machine-readable checklists.

Why Smart People Are Fleeing Social Media

06 Bubble or Future? Extreme Risk Concentration

Behind these $10B+ valuations lies alarming customer concentration.

On a podcast, Mercor’s Foody compared his client concentration to NVIDIA’s: the latter gets 61% of its revenue from just four customers. Such dependency is industry-wide.

· Historical Warnings Are Stark:

In August 2020, Australian data labeling firm Appen was valued at over 4.3billion;todayit’sunder130 million—a 97% drop. At its peak, 80% of its revenue came from just five clients: Microsoft, Apple, Meta, Google, and Amazon. Losing any one was existential. Today’s market is similarly concentrated—any shift could trigger a collapse.

· Fierce Competition Breeds Open Hostility

Rivalry has sparked public animosity among CEOs. Surge AI’s Chen called competitors “body shops.” Mercor’s Foody shot back that Surge and Scale are “old-school crowdsourcers.” Handshake’s Lord mocked rivals still spamming TikTok to find experts.

· The Meta-Scale Shockwave

In June 2025, Meta poached Scale’s CEO and bought a 49% stake—triggering an industry earthquake. Competing labs feared Scale was no longer neutral and that data might flow to Meta. The move reshuffled the entire sector overnight, doubling demand for neutral vendors like Handshake AI.

When the top 0.001% owns three times the wealth of the bottom 50% of the global population…

07 The Core Paradox: AGI Dreams vs. Data Dependence Reality

The deepest paradox lies here: AI labs’ ultimate goal is to create “Artificial General Intelligence”—capable of autonomously performing any task without domain-specific training.

But reality is moving in the opposite direction.

“AI labs believe the right future is: as performance improves, the need for human data will decline until it disappears entirely,” says Professor Daniel Kang of the University of Illinois. “But the opposite is happening.”

Labs are spending more on data than ever, and progress is coming from increasingly specialized datasets. Kang predicts: if this trend continues, acquiring high-quality human data in each discrete domain will become the main bottleneck for AI advancement.

This sets up two competing bets about AI’s future:

One is the labs’ “AGI vision”: a single breakthrough solves everything.

The other is the “AI as general-purpose technology” view: like the steam engine or the internet—powerful but limited, requiring specialized training and continuous updates in every application domain.

Data companies are clearly betting on the latter. SignalFire investor Ryan Wexler believes frontier models won’t “suddenly hit that magical general point.” So they’re investing in companies like Centaur AI, which provides specialized data for verticals like healthcare.

08 A New Economic Landscape: The Entire Economy as a Reinforcement Learning Environment

The grandest narrative is taking shape: the entire economy may become one giant reinforcement learning environment.

At Mercor, the customer support team doesn’t just handle tickets AI can’t solve—they also update scoring rubrics so AI can handle similar issues next time.

“If you zoom out,” Foody says, “it feels like the entire economy is becoming a reinforcement learning environment.”

Turing’s CEO predicts: AI data annotators will become the most common job on Earth in the coming years, with billions participating in evaluating and training models. Invisible Technologies describes itself as a “digital assembly line” that can mobilize experts “from anywhere on Earth” to generate data.

The deepest irony? The labs promising superintelligence and world-changing breakthroughs are still burning cash searching for a business model—while the “shovel sellers” providing training data, writing rubrics, and building test environments have already built highly profitable businesses.

In this AI gold rush, NVIDIA sells “picks” (chips), while these data companies sell “shovels” (data). They don’t bet on who will find gold—they sell tools to all prospectors.

The Only Way to Win Is Not to Compete: Peter Thiel’s $750B Startup Assembly Line Success Formula

09 Summary: Five Defining Traits of This Cycle’s AI Winners

Looking back at the winners of the 2025 AI wave, we can identify five clear characteristics:

Unbelievably Young Founding Teams

22-year-old founders with $10B valuations, 25-year-old coding-tool billionaires, 30-year-old legal-AI moguls… Youth has become a competitive advantage, not a liability. They grew up in the AI era, intuitively understand the technology, carry no legacy baggage, and make decisions swiftly.

From “Grunt Work” to Core Competency

Data labeling, rubric writing, environment building—once seen as AI’s “dirty work”—are now critical competitive moats. High-quality expert data isn’t optional; it’s the prerequisite for AI usefulness in professional domains.

Stability of B2B Models

Unlike consumer-facing AI apps (like chatbots), these data companies operate on a business-to-business (B2B) model, serving AI labs with clear budgets and urgent needs. Though client concentration is high, demand is real and immediate.

Platformization of Expert Networks

Winners aren’t simple brokers—they’ve built platforms that aggregate global expert networks. Handshake’s 20M alumni, Surge’s domain expert pool, Mercor’s engineer community—these network effects are hard to replicate quickly.

Betting on “AI as Ordinary Technology”

The most successful players are fundamentally betting that AI won’t suddenly become general intelligence, but will instead behave like other technologies—requiring specialization, continuous training, and constant updates in every vertical. Their entire business rests on this assumption.

When the top 0.001% owns three times the wealth of the bottom 50% of the global population…

10 Surprises: The Real Flow of Wealth Hidden Beneath the Hype

The biggest surprise of the 2025 AI gold rush isn’t the speed of technical progress—but the path of wealth distribution.

Surprise #1: Shadows Beneath the Spotlight

The media darlings—OpenAI, Anthropic, Google DeepMind—are still figuring out monetization, while the companies rarely in headlines are already profitable with jaw-dropping valuations.

Surprise #2: Monetization of Expertise

Once, expertise was monetized through consulting, publishing, or teaching. Now, it’s monetized by being encoded into AI training data. A Wall Street analyst’s financial modeling approach or a top lawyer’s argument framework can become a million-dollar “scoring rubric.”

Surprise #3: Higher Education’s Unexpected Outlet

University career platform Handshake accidentally became an AI data hub; elite alumni networks turned into expert reservoirs. The higher education system is emerging as a key supplier of AI training data.

Surprise #4: Redefining Work

“Data annotation” no longer means low-wage repetitive labor—it might involve nuclear engineering assessments, medical diagnosis validation, or legal clause analysis. This is a new form of knowledge work: on-demand, remote, project-based, globally priced.

Surprise #5: The Digital Mirror of the Entire Economy

The grandest vision is crystallizing: digitizing, rule-ifying, and making trainable every workflow in the economy. Every business process, every professional judgment, every tacit skill could be broken down into scoring rubrics—becoming AI training material.

When historians look back on 2025, they may identify a fascinating inflection point: the year when the dream of creating intelligence began yielding to the reality of training it; when alongside the grand narrative of AGI, a more pragmatic, immediately profitable ecosystem quietly took root.

Those who got rich first in the AI wave weren’t the prophets of the future—but the craftsmen equipping the prophets. They don’t debate the philosophical question “Will AI replace humans?” Instead, they solve the engineering problem: “How do we teach human knowledge to AI?”

The 2025 AI gold rush reminds us: on the journey to the stars, never overlook those building the rocket fuel factories. They may not stand on the launchpad—but their work determines how far the rocket can fly.

A new map of wealth is being drawn—and the first people to mark their territories on it have surprised everyone.

Sources:

The Verge: "AI Data Gold Rush" (December 15, 2025)

The New York Times: "The New Billionaires of the A.I. Boom" (December 29, 2025)

PitchBook Database

In any gold rush, the first to get rich are always those selling shovels and water.

As we look back at the AI commercial wave at the end of 2025—cutting through the surface glamour and noise—we’re startled to discover this truth: the shovel-sellers have already struck it rich, while the gold-diggers haven’t even found the vein.

While the world fixates on OpenAI’s next-generation model or NVIDIA’s next chip, a quiet redistribution of wealth and power is unfolding deep within the AI tide.

The most surprising story of 2025 is this: the giants striving to build “Artificial General Intelligence” (AGI) still haven’t found a reliable business model. Yet the companies supplying them with training data, writing scoring rubrics, and building test environments—the ones once dismissed as “doing the dirty, grunt work of AI”—are now raking in enormous profits.

In the shadows of these tech titans, a wealth-creation boom reminiscent of the 19th-century Gilded Age is quietly underway. A cohort of people in their twenties, armed with startups that didn’t even exist a few years ago, is rocketing into the billionaire club at an unprecedented pace. They rarely make headlines; you’ve probably never heard their names—Mercor, Surge AI, Handshake, Scale AI…

These companies are generating wealth at a historically rare speed: valuations in the tens of billions, revenues doubling overnight, founders becoming the youngest billionaires ever. This is the shovel business hidden beneath the gold mine—and it may be redefining who the real winners of the AI revolution truly are.

This article offers a comprehensive overview of the little-known winners of this AI gold rush and delves into the more concealed flows of wealth behind this technological revolution. You’ll see something highly counterintuitive: the new winners aren’t the spotlight-seeking prophets, but the “infrastructure engineers” rooted in real-world needs. They’re redefining value creation and capture at a breathtaking pace and in unexpected ways.

Finally, we’ll distill five defining traits of these AI-era winners and reveal five surprising insights about the future of technology, the essence of business, and the restructuring of societal division of labor hidden beneath this wealth feast.

When Teenagers Start Running Their Own AI Companies

01 Undercurrent: Data Deserts and the Gold Rush

In 2018, when Google researchers published a paper titled “Everyone Wants to Do the Model Work, Not the Data Work,” they exposed a harsh reality: in AI development, data preparation was seen as “an unfortunate but necessary chore” to be done as quickly and cheaply as possible.

Seven years later, everything has flipped.

Data is no longer just fuel—it’s a scarce strategic resource. Frontier AI labs have already “sucked dry” all easily accessible text, images, and code from the internet. The era of breakthroughs simply by scaling up data volume is over.

The real bottleneck has emerged: to make AI genuinely useful in specialized domains—like law, healthcare, or finance—it no longer needs massive generic datasets, but small, high-quality, expert-curated data tailored to specific tasks.

According to internal industry estimates, AI labs alone spent over $10 billion on training data in 2025, with the vast majority coming from just five or so companies. Paradoxically, these big spenders haven’t yet monetized AI themselves—but their data suppliers are already reporting stable profits.

“It’s human expertise from every corner of the world,” one practitioner described the current demand. From nuclear engineering to animal training, from medical diagnosis to carpentry—any human skill that can be codified into a “scoring rubric” has become hot commodity.

And new billionaires are being minted from this trend.

Peak Report: What Kind of Wealth Transfer Is Stablecoin Really Enabling?

02 Rise of the New Elite: The Lightning Strike of Data Giants

· Mercor: $10B Valuation at Age 22

In 2019, 19-year-old Brendan Foody and two high school friends founded Mercor with a simple idea: help startup friends hire software engineers overseas. The platform launched in 2023, using AI to screen résumés and conduct interviews, quickly hitting $1 million in annual revenue.

The turning point came in early 2024. Scale AI—the then-star of AI data labeling—approached them: it needed 1,200 software engineers. Scale needed these engineers to produce code for training data, as OpenAI, Anthropic, and others raced to teach AI how to code.

Foody sensed a bigger opportunity. When the recruited engineers began complaining about Scale withholding wages (the company was being sued in California for wage theft), he decided to cut out the middleman. Mercor started dealing directly with AI labs, offering expert-level data services.

In September 2025, Foody announced Mercor’s annualized revenue had hit 500million,callingit“thefastest?growingcompanyinhistory.”Itslatestfundingroundvaluedthecompanyat10 billion. At just 22, Foody and his co-founders became the youngest self-made billionaires ever.

· Surge AI: The Silent Industry Dominator

Unlike Mercor’s high-profile rise, Surge AI’s ascent happened almost in silence. Founder Edwin Chen, formerly of Google, Twitter, and Facebook, was shocked by the poor quality of data from existing vendors—riddled with errors from non-experts working for pennies.

He decided to build his own solution. Surge created the Data Annotation Tech platform, characterized by smaller scale, sharper focus, and stricter quality control. It pays around $30/hour—far above industry standards.

The strategy worked. Surge has been profitable since inception, and in 2024 its revenue surpassed 1billion,exceedingScaleAI’sreported870 million. Chen still owns roughly 75% of the company. Reuters reported that Surge is now seeking a 1billionraiseata15 billion valuation.

· Handshake AI: Revaluing the University Network

When Garrett Lord founded Handshake in 2014, his goal was to build a “Linkedin for college students.” Over 1,000 university career centers paid to join, and companies could recruit from a network of 20 million alumni and graduate students.

In early 2025, Lord noticed legacy data vendors scrambling for “experts”—and realized his platform already housed a vast pool of highly educated talent. He quickly incubated Handshake AI internally.

The pivotal moment came in June 2025: Meta poached Scale AI’s CEO and acquired a 49% stake. Competing labs, fearing Scale was no longer neutral, rushed to find alternatives. Demand flooded into Handshake AI.

“It was like someone smashed a billion-dollar pi?ata over every data startup’s head,” Lord said. Demand tripled overnight. By November, Handshake AI’s annual revenue exceeded $150 million—surpassing its decade-old core business.

“We grew from 3 to 150 people in five months,” Lord recalled. “One Monday, we onboarded 18 people. We ran out of desks.”

Venture capitalists and tech founders are rushing to catch up—this person’s value in the AI era is skyrocketing.

03 Application Breakthrough: From Tools to Titans

If Mercor and Surge are the “shovel sellers,” another class of companies is trying to build entirely new machines with AI. They’re not feeding models data—they’re using models to reshape entire vertical industries.

In 2025, these “AI-native applications” shattered the old order of SaaS (Software-as-a-Service), solving domain-specific pain points and achieving astonishing valuation leaps.

· Cursor: The Alchemists of Code

The brightest star in this category is undoubtedly the founding team of Cursor (parent company Anysphere). Michael Truell, Sualeh Asif, and others are MIT alumni who only graduated in 2022.

Their product, Cursor, is essentially an AI-assisted code editor—but by 2025, it had become an indispensable tool for programmers. More importantly, Cursor enjoys a unique “data flywheel”: the code users generate while using it is precisely the “golden data” AI labs crave most. Reports say both OpenAI and xAI have expressed acquisition interest.

In November 2025, after a 2.3billionfundinground,Cursorreachedastaggering27 billion valuation. The net worth of its four young founders surged collectively into the billionaire ranks.

In San Francisco, 30-year-old Winston Weinberg still shares an apartment with his co-founder Gabe Pereyra.

Despite living like ordinary workers, their balance sheets tell a different story. Their AI legal software startup, Harvey, underwent three funding rounds in 2025, with its valuation soaring from 3billionatthestartoftheyearto8 billion by year-end.

Faced with astronomical wealth, Weinberg remains remarkably grounded: “Sure, there are billions on paper—but it’s still just paper wealth.”

· Perplexity: The Wisdom Engine Challenging Google

In search, Perplexity is shaking Google’s foundation. Its 31-year-old CEO, Aravind Srinivas, founded this AI search engine company in 2022.

Unlike traditional search, Perplexity aims to deliver direct answers instead of links. By the end of 2025, its valuation stood at approximately $20 billion. Interestingly, Srinivas claims he doesn’t care much about money, preferring a simple life: “Seeking wisdom is far more important than seeking wealth.”

These three companies represent another path to AI commercialization: not infrastructure, but super-applications. They prove that if you dominate a vertical deeply enough, AI can generate value rivaling the oil giants of old.

04 The New Billionaires: Young, Fast, Homogeneous

A December 2025 New York Times report revealed an even more startling phenomenon: founders of these data companies are joining the billionaire ranks at an unprecedented speed.

· Age Records Shattered

Mercor’s three founders, aged 22, became the youngest self-made billionaires in October 2025 after their funding round. Cursor’s founding team, around 25, saw their company valued at $27 billion in November 2025.

· Speed of Wealth Creation Breaks All Records

For contrast: Elon Musk took years to go from millionaire to billionaire. Many of these AI founders went from zero to $1 billion in net worth in just 2–3 years.

Take Mira Murati, 37, who founded Thinking Machines Lab in February 2025; by June, it was valued at 10billion—withouthavingreleasedanyproduct.IlyaSutskever’sSafeSuperintelligencehita32 billion valuation in 2025, also pre-product.

· Highly Homogeneous Group

These new billionaires show striking homogeneity: overwhelmingly male, mostly from elite schools (MIT, Stanford, etc.), concentrated between ages 20–40.

University of Washington professor Margaret O’Mara warns this boom is amplifying “homogeneity”—people of similar backgrounds, ages, and genders now occupy the top of the wealth pyramid.

The influencer era is ending, but everyone can still succeed at the top of their own ladder.

05 Industry Earthquake: From Body Shops to the Expert Economy

The transformation in the data industry reflects a fundamental shift in AI training paradigms.

Early data work—like labeling images for autonomous driving—was labor-intensive and low-skill. Amazon Mechanical Turk paid thousands of workers mere cents per image to label cats and dogs. Scale AI’s Remotasks platform had workers draw boxes around stop signs and traffic cones.

ChatGPT changed the game. Its use of Reinforcement Learning from Human Feedback (RLHF) required evaluating answer quality—not just simple labeling. Judging medical advice requires medical training; assessing legal arguments demands legal knowledge.

This gave birth to the expert data economy. It’s no longer about “body shops” stacking cheap labor, but sourcing experts with advanced degrees and professional credentials.

Surge AI boasts Fields Medal-winning mathematicians, Supreme Court litigators, and Harvard historians. Mercor highlights its Goldman Sachs analysts and McKinsey consultants. Handshake AI mobilizes PhDs, master’s students, and professionals from over 1,000 universities.

· The Rubric Arms Race

AI companies discovered that to truly master complex skills, models need extremely detailed “scoring rubrics”—breaking “doing well” into dozens of measurable criteria.

Creating one rubric can take over 10 hours and include more than ten distinct standards. OpenAI’s medical benchmark includes nearly 50,000 criteria, with different combinations applied to different prompts. Insiders say labs order tens of thousands to hundreds of thousands of rubrics per training run—totaling millions of individual standards.

“It needs to be hyper-granular,” explains Mercor’s Foody. To build a consulting rubric, you start by creating a taxonomy of all consulting industries, then subtypes of consulting within each, then report formats for each subtype.

· Explosive Demand for Environments

Rubrics alone aren’t enough—you also need “gyms” for AI to practice: reinforcement learning environments.

These vary wildly: “Sometimes they’re clones of DoorDash or Salesforce, but often they’re custom enterprise environments,” says Alex Ratner, CEO of Snorkel AI, whose company originally built annotation software but now also offers synthetic data services.

Demand for both environments and rubrics is surging simultaneously. AI firms are attempting brute-force approaches—hiring thousands of lawyers, consultants, and specialists to encode human knowledge into machine-readable checklists.

Why Smart People Are Fleeing Social Media

06 Bubble or Future? Extreme Risk Concentration

Behind these $10B+ valuations lies alarming customer concentration.

On a podcast, Mercor’s Foody compared his client concentration to NVIDIA’s: the latter gets 61% of its revenue from just four customers. Such dependency is industry-wide.

· Historical Warnings Are Stark:

In August 2020, Australian data labeling firm Appen was valued at over 4.3billion;todayit’sunder130 million—a 97% drop. At its peak, 80% of its revenue came from just five clients: Microsoft, Apple, Meta, Google, and Amazon. Losing any one was existential. Today’s market is similarly concentrated—any shift could trigger a collapse.

· Fierce Competition Breeds Open Hostility

Rivalry has sparked public animosity among CEOs. Surge AI’s Chen called competitors “body shops.” Mercor’s Foody shot back that Surge and Scale are “old-school crowdsourcers.” Handshake’s Lord mocked rivals still spamming TikTok to find experts.

· The Meta-Scale Shockwave

In June 2025, Meta poached Scale’s CEO and bought a 49% stake—triggering an industry earthquake. Competing labs feared Scale was no longer neutral and that data might flow to Meta. The move reshuffled the entire sector overnight, doubling demand for neutral vendors like Handshake AI.

When the top 0.001% owns three times the wealth of the bottom 50% of the global population…

07 The Core Paradox: AGI Dreams vs. Data Dependence Reality

The deepest paradox lies here: AI labs’ ultimate goal is to create “Artificial General Intelligence”—capable of autonomously performing any task without domain-specific training.

But reality is moving in the opposite direction.

“AI labs believe the right future is: as performance improves, the need for human data will decline until it disappears entirely,” says Professor Daniel Kang of the University of Illinois. “But the opposite is happening.”

Labs are spending more on data than ever, and progress is coming from increasingly specialized datasets. Kang predicts: if this trend continues, acquiring high-quality human data in each discrete domain will become the main bottleneck for AI advancement.

This sets up two competing bets about AI’s future:

One is the labs’ “AGI vision”: a single breakthrough solves everything.

The other is the “AI as general-purpose technology” view: like the steam engine or the internet—powerful but limited, requiring specialized training and continuous updates in every application domain.

Data companies are clearly betting on the latter. SignalFire investor Ryan Wexler believes frontier models won’t “suddenly hit that magical general point.” So they’re investing in companies like Centaur AI, which provides specialized data for verticals like healthcare.

08 A New Economic Landscape: The Entire Economy as a Reinforcement Learning Environment

The grandest narrative is taking shape: the entire economy may become one giant reinforcement learning environment.

At Mercor, the customer support team doesn’t just handle tickets AI can’t solve—they also update scoring rubrics so AI can handle similar issues next time.

“If you zoom out,” Foody says, “it feels like the entire economy is becoming a reinforcement learning environment.”

Turing’s CEO predicts: AI data annotators will become the most common job on Earth in the coming years, with billions participating in evaluating and training models. Invisible Technologies describes itself as a “digital assembly line” that can mobilize experts “from anywhere on Earth” to generate data.

The deepest irony? The labs promising superintelligence and world-changing breakthroughs are still burning cash searching for a business model—while the “shovel sellers” providing training data, writing rubrics, and building test environments have already built highly profitable businesses.

In this AI gold rush, NVIDIA sells “picks” (chips), while these data companies sell “shovels” (data). They don’t bet on who will find gold—they sell tools to all prospectors.

The Only Way to Win Is Not to Compete: Peter Thiel’s $750B Startup Assembly Line Success Formula

09 Summary: Five Defining Traits of This Cycle’s AI Winners

Looking back at the winners of the 2025 AI wave, we can identify five clear characteristics:

Unbelievably Young Founding Teams

22-year-old founders with $10B valuations, 25-year-old coding-tool billionaires, 30-year-old legal-AI moguls… Youth has become a competitive advantage, not a liability. They grew up in the AI era, intuitively understand the technology, carry no legacy baggage, and make decisions swiftly.

From “Grunt Work” to Core Competency

Data labeling, rubric writing, environment building—once seen as AI’s “dirty work”—are now critical competitive moats. High-quality expert data isn’t optional; it’s the prerequisite for AI usefulness in professional domains.

Stability of B2B Models

Unlike consumer-facing AI apps (like chatbots), these data companies operate on a business-to-business (B2B) model, serving AI labs with clear budgets and urgent needs. Though client concentration is high, demand is real and immediate.

Platformization of Expert Networks

Winners aren’t simple brokers—they’ve built platforms that aggregate global expert networks. Handshake’s 20M alumni, Surge’s domain expert pool, Mercor’s engineer community—these network effects are hard to replicate quickly.

Betting on “AI as Ordinary Technology”

The most successful players are fundamentally betting that AI won’t suddenly become general intelligence, but will instead behave like other technologies—requiring specialization, continuous training, and constant updates in every vertical. Their entire business rests on this assumption.

When the top 0.001% owns three times the wealth of the bottom 50% of the global population…

10 Surprises: The Real Flow of Wealth Hidden Beneath the Hype

The biggest surprise of the 2025 AI gold rush isn’t the speed of technical progress—but the path of wealth distribution.

Surprise #1: Shadows Beneath the Spotlight

The media darlings—OpenAI, Anthropic, Google DeepMind—are still figuring out monetization, while the companies rarely in headlines are already profitable with jaw-dropping valuations.

Surprise #2: Monetization of Expertise

Once, expertise was monetized through consulting, publishing, or teaching. Now, it’s monetized by being encoded into AI training data. A Wall Street analyst’s financial modeling approach or a top lawyer’s argument framework can become a million-dollar “scoring rubric.”

Surprise #3: Higher Education’s Unexpected Outlet

University career platform Handshake accidentally became an AI data hub; elite alumni networks turned into expert reservoirs. The higher education system is emerging as a key supplier of AI training data.

Surprise #4: Redefining Work

“Data annotation” no longer means low-wage repetitive labor—it might involve nuclear engineering assessments, medical diagnosis validation, or legal clause analysis. This is a new form of knowledge work: on-demand, remote, project-based, globally priced.

Surprise #5: The Digital Mirror of the Entire Economy

The grandest vision is crystallizing: digitizing, rule-ifying, and making trainable every workflow in the economy. Every business process, every professional judgment, every tacit skill could be broken down into scoring rubrics—becoming AI training material.

When historians look back on 2025, they may identify a fascinating inflection point: the year when the dream of creating intelligence began yielding to the reality of training it; when alongside the grand narrative of AGI, a more pragmatic, immediately profitable ecosystem quietly took root.

Those who got rich first in the AI wave weren’t the prophets of the future—but the craftsmen equipping the prophets. They don’t debate the philosophical question “Will AI replace humans?” Instead, they solve the engineering problem: “How do we teach human knowledge to AI?”

The 2025 AI gold rush reminds us: on the journey to the stars, never overlook those building the rocket fuel factories. They may not stand on the launchpad—but their work determines how far the rocket can fly.

A new map of wealth is being drawn—and the first people to mark their territories on it have surprised everyone.

Sources:

The Verge: "AI Data Gold Rush" (December 15, 2025)

The New York Times: "The New Billionaires of the A.I. Boom" (December 29, 2025)

PitchBook Database