Talent Marketplaces and AI-Native Hiring: Sizing to 2028

A research-grade sizing and 2028 forecast for the four marketplaces that now move human labor at scale: classic freelance, on-demand contingent, employer-of-record, and the AI-native data-labor platforms that came from nowhere to outvalue the entire category.

The single most important fact in the talent-marketplace economy in 2026 is that an AI-data-labor company you had probably not heard of two years ago, Mercor, is worth 10 billion dollars, while Upwork, the largest publicly traded freelance platform on Earth, books 788 million dollars of annual revenue and trades for a fraction of that figure - TechCrunch. Mercor raised a 350 million dollar Series C at a 10 billion dollar valuation, a fivefold jump in eight months, on the back of a contractor network it pays more than 1.5 million dollars every single day - CNBC. That is not a recruiting tool. That is a labor marketplace where the buyer is an AI lab, the worker is a human teaching a model, and the matching is done by a machine. The category that the venture market is pricing as the future of work did not exist in its current form before 2023.

The problem this creates for anyone sizing the talent-marketplace economy is that the old map is wrong in two directions at once. The freelance platforms that defined the category for a decade, Upwork and Fiverr, are growing in the single digits and shedding active buyers. The employer-of-record platforms that looked like the next growth engine are locked in a corporate-espionage scandal that has reached a Department of Justice grand jury. And a brand-new category, AI data labor, has quietly become the place where the most money, the most workers, and the highest valuations now concentrate. A buyer, investor, or operator who sizes this market by extrapolating the freelance platforms will miss the entire story. The growth has moved.

This guide sizes the four marketplaces from first principles and forecasts each to 2028, with real funding, real revenue, and real take-rate data. It covers classic freelance (Upwork, Fiverr, Toptal), on-demand contingent and hourly (Instawork, Wonolo), engineering and elite-team networks (Andela, A.Team, Braintrust), employer-of-record (Deel, Remote, Velocity Global, Oyster), and the AI-native data-labor frontier (Mercor, Turing, Scale AI via Outlier, Micro1). It then resolves the question every operator is asking: where is the value actually accruing, and which take-rate model survives the next three years. This is the marketplace companion to our talent technology market map, which maps the vendor stack that sells to recruiters; here we map the platforms that move the labor itself.

Contents

1. How to size a talent marketplace
2. The four-marketplace scorecard
3. Classic freelance: the slow-growth incumbents
4. On-demand contingent: hourly labor at app speed
5. Elite networks: engineers, teams, and the take-rate war
6. Employer of record: the espionage that defines the category
7. AI data labor: the category that ate the valuations
8. The take-rate divergence: zero, premium, and flat
9. Sizing the total addressable market to 2028
10. The value-chain map: who captures the dollar
11. The forecast: four marketplaces, four trajectories
12. What buyers, investors, and operators should do

1. How to size a talent marketplace

A talent marketplace is, at its economic core, a three-sided machine: it aggregates supply (workers), aggregates demand (buyers of work), and inserts itself into the transaction to take a cut. The size of a marketplace is therefore never a single number, and conflating the numbers is the most common error in sizing this economy. There is gross services volume (GSV), the total dollars of work transacted through the platform; there is net revenue, the slice the platform keeps; and there is enterprise valuation, the market's bet on future net revenue. Upwork transacts more than 4 billion dollars of GSV but keeps only 788 million dollars of revenue - Upwork 8-K. Mercor, by contrast, pays out more than 547 million dollars a year to contractors, which is its GSV-equivalent, yet is valued at 10 billion dollars. Sizing the market means tracking all three layers and never substituting one for another.

The second principle of sizing is that the take rate is the entire business model compressed into one number, and the divergence in take rates across this market is the most analytically important signal it produces. A take rate is the percentage of GSV the platform keeps as revenue. Upwork's effective take rate runs in the low teens, Fiverr's runs higher, Toptal's runs above thirty percent, and Braintrust's is structurally zero on the talent side. When two marketplaces serve the same nominal buyer with take rates that differ by a factor of ten, they are not the same business, and they will not converge to the same outcome. The take-rate spread tells you where margin lives, where competitive pressure is fiercest, and which models can survive a price war. We return to this in depth in section 8, because it is the spine of the entire forecast.

The third principle is that AI changes the unit economics of a marketplace in a way that no prior technology shift did, by attacking the supply side directly. Every previous marketplace innovation made matching cheaper or faster. AI does that too, but it also does something new: it lets the marketplace itself perform the work of evaluation that used to require a human recruiter, and in the data-labor case it makes the human worker a complement to a machine rather than a substitute for one. Mercor's matching is done by an AI that interviews candidates and reads their portfolios. This collapses the cost of running the supply side, which is why an AI-native marketplace can scale to billions of dollars of payout with a few hundred employees. Sizing the market without modeling this cost collapse will systematically undervalue the AI-native players and overvalue the labor-intensive incumbents.

The fourth and final principle for reading this map is that the four marketplaces serve genuinely different jobs, and the right way to size each is to start from the job, not the company. Classic freelance sells discrete project work to small and mid-sized buyers. On-demand contingent sells hourly shifts for physical-world labor. Employer of record sells the legal right to employ someone in a country where you have no entity. AI data labor sells vetted human cognition to train machines. These are four different products with four different buyers, four different supply pools, and four different growth curves. A unified "gig economy" number that lumps them together (the often-cited 455 billion dollar figure) is useless for any actual decision because it averages four trajectories that are diverging fast - Business Research Insights. The sections that follow size each separately.

2. The four-marketplace scorecard

Before profiling each marketplace individually, it helps to see all four scored side by side on the dimensions that determine where value is heading. This scorecard does not rank individual companies; it ranks the four marketplace categories on the strategic characteristics that matter to anyone allocating capital, attention, or procurement budget in 2026. The criteria are chosen from first principles to answer the questions that actually drive decisions: How fast is the category growing (growth velocity)? How much pricing and margin power does the platform hold (margin power)? How defensible is the model against the next entrant (defensibility)? And how much is AI reshaping the category's fundamental economics (AI disruption)?

Each category receives a score from 0 to 10 on each criterion, with the justification in the cell, because a bare number is worthless without the data point behind it. The final score is a weighted average expressing how attractive the category is as a place to build, invest, or watch closely in 2026, where a higher score means the category is where the energy and the value are concentrating. The weights reflect that growth velocity and AI disruption are the forces redrawing this market fastest, so they carry the most weight.

Criteria explained: Growth Velocity (30%) measures how fast the category is expanding in GSV, revenue, and valuation, which is the clearest signal of where value is forming. Margin Power (20%) measures how much take-rate and pricing leverage the platform holds against its supply and demand. Defensibility (20%) measures how hard the model is to replicate, whether through compliance moats, network density, or operational complexity. AI Disruption (30%) measures how profoundly AI is reshaping the category's economics, which carries high weight because it is the force separating winners from losers. A higher final score means the category is more attractive to build in, invest in, or watch closely in 2026. The table is ordered by final score, so the row order is the order of strategic energy in the market.

The scorecard surfaces the central thesis of this entire forecast in a single ranking: the value in the talent-marketplace economy has inverted. The category that did not exist three years ago, AI data labor, sits at the top because it combines the fastest growth with the most profound AI disruption. The category that defined the market for a decade, classic freelance, sits at the bottom because its growth has stalled and AI is now eating into the very work its freelancers sell. Employer of record sits in the strong middle on the strength of a genuine compliance moat, even as its leaders are mired in scandal. On-demand contingent occupies a defensible but unspectacular middle ground. The sections that follow profile each marketplace in the order a capital allocator should think about it, but we begin with the incumbents to establish the baseline that everything else is now outrunning.

3. Classic freelance: the slow-growth incumbents

Classic freelance is the marketplace that taught the world the model, and it is now the clearest case study in what happens to a marketplace when its core value proposition gets commoditized and then partially automated. The fundamental job of a classic freelance platform is to connect a buyer who needs discrete project work (a logo, a website, a marketing campaign, a piece of code) with an independent worker who can deliver it, and to handle the trust, escrow, and payment plumbing in between. For a decade this was a category in secular growth as work went remote and project-based. In 2026 the public leaders are growing in the single digits, and one of them is losing buyers outright.

The numbers tell an unambiguous story of maturity. Upwork posted 788 million dollars in revenue for 2025 on more than 4 billion dollars of gross services volume, with adjusted EBITDA of 226 million dollars and 785,000 active clients - Upwork 8-K. That is a profitable, well-run business, but its growth has flattened into the single digits, and its market capitalization reflects a market that no longer prices it as a growth story. Fiverr reported 430.9 million dollars in 2025 revenue, up 10.1 percent, but its active buyer base fell 13.6 percent to 3.1 million even as spend per buyer rose 13.3 percent to 342 dollars - Yahoo Finance. Fiverr is shedding low-value buyers and retaining higher-value ones, which is a rational defensive move, but it is not the profile of a category in expansion.

The structural reason for the slowdown is that classic freelance sits at the exact intersection of two AI forces pulling against each other, and on balance the forces are a wash at best for the incumbents. On one side, AI is a tailwind: it makes matching better, lets the platform automate proposal drafting and project scoping, and creates demand for a new class of AI-related freelance work. Upwork's own leadership has leaned into AI integration as a growth lever, and AI-skills categories are among its fastest-growing - Yahoo Finance. On the other side, AI is a direct substitute for the lowest tier of freelance work. The 5-dollar logo, the boilerplate copy, the simple translation: these are precisely the tasks a buyer can now do with a model instead of a Fiverr seller. Fiverr's buyer decline is the visible edge of this substitution, as the cheapest tasks evaporate into AI tools.

The freelance platforms market as a whole is still projected to grow, but the growth rate masks a brutal internal redistribution that sizing must account for. The category is estimated at roughly 6.37 billion dollars in 2025, rising to about 7.33 billion dollars in 2026, with longer-run projections reaching above 24 billion dollars by 2033 at roughly an 18 percent CAGR by one widely cited measure - Grand View Research. The honest reading of that forecast is that the category grows because new AI-native entrants and specialized platforms are pulling the average up, while the legacy generalist leaders grow far more slowly. A forecast that applies the category CAGR to Upwork or Fiverr will be wrong; the growth is migrating to platforms that did not exist when those two went public, which is the same pattern we documented in the broader recruiting technology consolidation.

The practical guidance for anyone evaluating classic freelance is to separate the two halves of the category and treat them as different investments. The generalist horizontal platforms (Upwork, Fiverr) are mature cash machines whose growth has normalized and whose lowest-value work is under direct AI threat; they should be valued as profitable incumbents, not as growth stories. The specialized and elite networks that sit on top of the freelance model (covered in section 5) retain genuine growth because they sell vetted, high-value talent that AI cannot yet substitute. The mistake to avoid is treating "freelance" as one monolithic growth category. It is two categories now: a commoditizing base under AI pressure, and a defensible premium layer, and only the second is where value is still forming.

4. On-demand contingent: hourly labor at app speed

On-demand contingent is the marketplace that brought app-speed matching to the physical world, and it is the category whose growth is most constrained by an unglamorous reality: you cannot teleport a warehouse worker. The fundamental job here is to fill an hourly shift, a warehouse pick, a hospitality service slot, a delivery run, a light-industrial role, by matching a local worker to a local employer on demand, often within hours. This is the gig economy in its most literal, physical form, distinct from the digital freelance platforms because the labor and the demand are both geographically bound. That constraint shapes everything about how the category grows and how it should be sized.

The leading platforms have built real scale within that constraint. Instawork connects more than 4 million skilled hourly workers with tens of thousands of businesses, and raised a 60 million dollar Series D in March 2025 led by TCV, bringing total funding to 160 million dollars - Staffing Industry Analysts. Wonolo has raised more than 200 million dollars and serves over 1 million workers across warehouse, delivery, merchandising, event staffing, and administrative roles - TechCrunch. These are substantial businesses, but the funding profile tells the strategic story: the rounds are modest relative to the EOR and AI-data-labor categories, and the latest Instawork raise is explicitly earmarked for AI capabilities, signaling that even the physical-labor marketplaces see software intelligence as their next lever.

The defensibility of this category is real but local, which is the key insight for sizing it. A digital freelance marketplace competes globally; a single Upwork serves the whole world. An on-demand contingent marketplace competes city by city, because liquidity (enough workers and enough shifts in a given metro to match reliably) has to be built market by market. This is the same density moat that powered ride-hailing and food delivery, and it makes the category genuinely defensible in any market where a platform achieves liquidity, but it also caps growth velocity, because each new market is a fresh liquidity problem rather than a free expansion. This is why the contingent platforms score well on defensibility but only moderately on growth velocity in the scorecard.

AI enters this category as an efficiency layer rather than a transformation, which is the honest assessment of where it stands in 2026. Instawork's investment in AI is aimed at optimizing matching, predicting shift demand, scheduling, and reducing no-shows, the operational frictions that determine whether a physical-labor marketplace is profitable - Instawork. These are valuable improvements, and they widen the margin and the reliability moat, but they do not change the fundamental product the way AI changes data labor or screening. A warehouse shift still requires a human in the warehouse. AI makes the marketplace that fills the shift better; it does not replace the shift. The category is therefore being optimized by AI, not redefined by it, which is a meaningfully less disruptive dynamic than the one reshaping the front of the funnel.

For anyone sizing or investing in on-demand contingent, the guidance is to value liquidity and reliability above raw growth rate, because in a density-driven marketplace those are the durable assets. The platforms that win are the ones that achieve deep liquidity in their core markets and convert it into reliability (the shift gets filled, the worker shows up, the employer trusts the platform), which compounds into pricing power and retention. The category will grow steadily with the broader shift toward flexible hourly work, but it will not produce the parabolic curves of the AI-native marketplaces, because the physical constraint caps it. The reasonable position is to treat on-demand contingent as a defensible, steady-growth category where the prize goes to operational excellence and market density rather than to technological leaps.

5. Elite networks: engineers, teams, and the take-rate war

Elite networks are the part of the marketplace economy that sits above commodity freelance, selling vetted, high-skill talent (often engineers) at a premium, and they are the clearest battleground for the take-rate war that defines the entire industry's economics. The fundamental job here is curation: instead of giving a buyer access to millions of freelancers of unknown quality, an elite network promises that everyone in the pool has been screened, so the buyer pays more for confidence. The three models we examine, Toptal, Andela, and Braintrust, take radically different positions on the central question of how much to charge for that curation, and the contrast between them is the most instructive take-rate comparison in the market.

Toptal is the premium bootstrapped incumbent, and its economics are the proof that a high take rate can build a durable business without venture capital. Founded in 2010, Toptal markets access to the top 3 percent of freelance talent and reached roughly 167 million dollars in annual recurring revenue entirely bootstrapped, having never raised outside venture funding - GetLatka. Toptal's take rate runs well above thirty percent of what the client pays, which is the premium it extracts for rigorous vetting, and that fat margin is precisely what let it self-fund. The model has not been without drama: Toptal was the subject of a high-profile founder-versus-investor legal battle in which a jury found in 2025 that a tech investor had orchestrated a takedown plot against the company - CNBC. The durability of Toptal through that conflict is itself evidence of how much pricing power a genuine curation moat confers.

Andela took the opposite geographic bet and a venture-funded path, building a pan-African engineering network that became a unicorn. Founded in 2014, Andela trains and places software engineers from Africa into global companies, and reached a valuation exceeding 1.5 billion dollars in a 200 million dollar Series E backed by SoftBank Vision Fund 2 and the Chan Zuckerberg Initiative - TechCrunch. Andela's thesis is labor arbitrage with quality: tap a vast, underutilized engineering talent pool in emerging markets and connect it to global demand at rates below US engineering costs but above local ones, capturing the spread. The model competes directly with Toptal on the demand side, which is why the two became entangled in litigation, with Toptal alleging Andela poached employees and trade secrets - Justia. The litigation between the two elite networks is itself a signal of how contested the high-skill segment has become.

Braintrust is the radical experiment that pushes the take-rate question to its logical extreme: zero on the talent side. Braintrust is a decentralized talent network where talent keeps 100 percent of what they earn and the platform charges only the client a fee (currently around 15 percent), with the network governed by a token, BTRST, held by its community - Braintrust FAQ. The structural bet is that by taking zero from talent (versus the 20 to 40 percent that traditional platforms extract), Braintrust attracts the best workers and aligns incentives through community ownership, with referral programs paying out in BTRST tokens driving roughly half of new acquisition - Modular Capital. The model is unproven at the scale of the incumbents, and the token mechanism adds complexity and volatility, but it represents the most aggressive possible answer to the take-rate question and a genuine challenge to the premium model.

A.Team rounds out the elite-network picture with a fourth model: selling assembled teams rather than individuals. A.Team emerged from stealth with 60 million dollars in funding, backed by Insight Partners, Tiger Global, and Spruce House, having built a members-only network of top engineers, product managers, designers, and marketers - PR Newswire. The insight behind A.Team is that buyers often do not want a single freelancer; they want a functioning team that can ship a product, so A.Team's product is team formation rather than individual matching. This is a genuinely different unit of sale, and it sits at the higher-value end of the elite-network spectrum because an assembled, vetted team commands more than the sum of its individual members. The four models together (Toptal's premium, Andela's arbitrage, Braintrust's zero take, A.Team's team formation) map out the full strategic space of how to monetize curated talent, and the next section examines why their take rates are diverging rather than converging.

6. Employer of record: the espionage that defines the category

Employer of record (EOR) is the marketplace category with the deepest moat and, in 2026, the most lurid headlines, and the combination is not a coincidence. The fundamental job of an EOR is to let a company hire a worker in a country where it has no legal entity, by acting as the legal employer of record itself: it owns the local entity, runs the payroll, handles the taxes and benefits, and assumes the compliance liability, while the client directs the worker's actual work. This is a deceptively hard product, because it requires owning or operating compliant legal infrastructure in every country it serves, which is why the category has a genuine moat and why its leaders have raised and grown so aggressively. It is also why the competition turned vicious enough to involve the federal government.

Deel is the runaway leader, and its scale is the headline. Deel raised a 300 million dollar Series E in October 2025 at a 17.3 billion dollar valuation, co-led by Ribbit Capital alongside Andreessen Horowitz and Coatue, after surpassing 1 billion dollars in annual recurring revenue - TechCrunch. Deel processes 22 billion dollars in payroll annually across 150-plus countries, serving more than 35,000 customers and 1.5 million workers, and reported three consecutive years of profitability with its first 100-million-dollar revenue month in 2025 - Deel. By every operating metric, Deel is the dominant EOR, and its valuation places it second only to the AI-native marketplaces in the entire talent economy. That dominance is exactly what made the competitive conflict at the center of the category so explosive.

The defining drama of the EOR category is the corporate-espionage war between Deel and Rippling, which has escalated into a Department of Justice criminal probe. Rippling sued Deel in March 2025, alleging that Deel cultivated a Rippling employee as a corporate spy who searched Rippling's systems for the term "Deel" an average of 23 times a day over four months to harvest competitive intelligence; the employee was caught in a sting operation and confessed in a sworn statement in an Irish court - Rippling. Deel countersued in Delaware in April 2025, alleging Rippling infiltrated Deel's platform and ran a campaign of frivolous regulatory complaints and misleading marketing. The conflict moved from civil to criminal when, per a Wall Street Journal report, the DOJ opened a criminal investigation and a grand jury issued subpoenas seeking documents related to the alleged spying operation - PYMNTS. By late 2025, Rippling obtained bank records it says show funds flowing from Deel through the account of a Deel executive's spouse to the confessed spy - TechCrunch.

The espionage saga is not merely gossip; it is a structural signal about the category that a serious analyst must read correctly. Two facts explain why the EOR competition turned this ugly. First, the EOR moat is built on infrastructure and customer relationships rather than on a defensible technology, so competitive intelligence about which customers are considering switching is enormously valuable, more valuable than in a category where the product itself is the moat. Second, the category is winner-take-most: the platform with the most countries, the most workers, and the best compliance record compounds its advantage, so the stakes of the head-to-head fight between the two leaders are existential. The espionage was the visible symptom of a category where the prize is large, the moat is relationship-based, and the leaders are locked in a duel for it. That dynamic also raises the legitimate question of governance risk for any buyer placing a multi-year compliance bet on a vendor under federal criminal investigation.

Beyond the two protagonists, the EOR category has a credible second tier that the scandal makes more relevant, not less. Remote raised a 300 million dollar Series C at a valuation exceeding 3 billion dollars, led by SoftBank Vision Fund 2, positioning itself as the privacy-conscious, ownership-of-infrastructure alternative - TechCrunch. Velocity Global rebranded as Pebl in September 2025, repositioning from a pure EOR into an AI-powered hiring platform with flat per-employee pricing around 599 dollars per month across 185-plus countries - WhichPayroll. Oyster, founded in 2020, has raised more than 150 million dollars and competes on a self-service, transparent-pricing model also around 599 dollars per employee per month - Oyster comparison. The consistency of that ~599-dollar flat per-employee fee across the challengers is itself a critical data point for the take-rate analysis, because it reveals where EOR pricing is settling.

The EOR market is sized at roughly 5.59 billion dollars in 2025, rising toward 5.97 billion in 2026, with longer-run forecasts ranging from a conservative 6.9 percent CAGR (reaching about 6.79 billion by 2028) to an aggressive 14.8 percent CAGR (reaching nearly 20 billion by 2036) - Business Research Insights. The wide range reflects genuine uncertainty about how much of global hiring shifts to the EOR model, but even the conservative case is healthy growth, and Deel's individual trajectory (from zero to 1 billion dollars of ARR) shows the demand is real. The category's defensibility is the highest of the four marketplaces because the compliance-and-entity moat is genuinely hard to replicate, which is why it scores so well on defensibility even as its flat-fee pricing limits its margin-power score.

7. AI data labor: the category that ate the valuations

AI data labor is the marketplace that did not exist three years ago and now commands the highest valuations and the fastest growth in the entire talent economy, and understanding it is the single most important thing for anyone forecasting this market. The fundamental job is new: to recruit, vet, and deploy human experts (engineers, doctors, lawyers, scientists, linguists) to generate and evaluate the data that trains and aligns AI models. The buyer is an AI lab. The worker is a credentialed human teaching a machine. And the matching, crucially, is performed by AI itself. This is a labor marketplace whose existence depends on the AI boom, and whose growth tracks the capital pouring into frontier models.

Mercor is the defining company, and its trajectory is the most extraordinary in the talent economy. Founded in 2023 by Brendan Foody, Adarsh Hiremath, and Surya Midha, three high-school debate teammates who became Thiel Fellows and, in 2025, among the youngest self-made billionaires, Mercor built an AI system that interviews candidates, assesses their skills from transcripts and portfolios, and matches them to AI-training work - Wikipedia. Its 350 million dollar Series C at a 10 billion dollar valuation in October 2025 was a fivefold jump in eight months, led by Felicis with Benchmark and General Catalyst, on the back of a run rate north of 450 million dollars and contractor payouts exceeding 1.5 million dollars a day to a network of roughly 30,000 contractors, with customers including OpenAI and Anthropic - CNBC. That single 10-billion-dollar figure outvalues the entire public freelance-platform category, which is the data point that reframes the whole market.

Turing is the second pillar of the category, built on the same insight from a different starting point. Turing raised 111 million dollars in a Series E at a 2.2 billion dollar valuation in March 2025, having become a key coding-data and engineering-talent provider to OpenAI and other frontier-model builders, with roughly 300 million dollars in annualized revenue and around a year of profitability - TechCrunch. Turing originated as an engineering-talent marketplace and pivoted toward AGI infrastructure as it became clear that supplying expert human data to AI labs was a far larger and faster-growing business than placing engineers at companies. That pivot is itself the central lesson of the category: the most valuable thing a talent marketplace can sell in 2026 is not labor to companies but cognition to models.

Scale AI, through its Outlier platform, is the incumbent giant of the category and the cautionary tale. Scale AI reached an approximately 29 billion dollar valuation when Meta agreed to take a 49 percent non-voting stake for 14.3 billion dollars in June 2025, with founder Alexandr Wang moving into a senior role at Meta - Wikipedia. Scale coordinates a network of more than 100,000 contractors doing data work through subsidiaries including Outlier (for generative-AI data, staffed by credentialed experts) and the now-wound-down Remotasks - FinancialContent. Micro1 represents the same pattern at smaller scale, having built an AI recruiter named Zara and raised at a 500 million dollar valuation in 2025 - TechCrunch. The cluster of Mercor, Turing, Scale/Outlier, and Micro1 is now the center of gravity of the talent-marketplace economy by valuation.

The first-principles lesson of this category is the most important strategic insight in the entire forecast, so it bears stating precisely. When intelligence becomes cheap and abundant, the bottleneck shifts to the human judgment that trains and aligns the intelligence, and a new labor marketplace forms around that bottleneck. The reason these companies are worth more than the freelance platforms is not hype alone; it is that they sit on the critical path of the most capital-intensive technology buildout in history. Every frontier lab needs expert human data, the demand is enormous and growing with each model generation, and the AI-native matching lets these platforms scale supply far more cheaply than a human-recruiter marketplace ever could. The same first-principles dynamic, cheap intelligence creating new categories rather than merely improving old ones, is the one we traced through the recruiting workflow in our state of AI in recruiting analysis.

A research house owes its readers the skeptical reading of this category, not only the bullish one, because the risks are real and concentrated. The first risk is demand concentration: a handful of frontier labs are the buyers, so any slowdown in model-training spend, or a shift toward synthetic data, would hit the category hard. The second is labor risk: the human side of AI data labor has drawn intense scrutiny over wages and conditions, with lawsuits against Scale AI and Outlier over alleged underpayment, contractor misclassification, and psychological harm from disturbing content, and an active US Department of Labor investigation into Scale AI's compliance with the Fair Labor Standards Act - The Register. The third is valuation risk: Mercor's traction is concentrated in an unusual demand pocket, and whether the model generalizes beyond AI-lab data work is genuinely unproven. The honest position is that the category is the most exciting and the most fragile of the four, with extraordinary growth resting on a narrow and ethically fraught demand base.

8. The take-rate divergence: zero, premium, and flat

The take rate is where the abstract economics of these marketplaces become concrete, and the most analytically important pattern in the entire market is that take rates are diverging rather than converging. In a maturing market, theory predicts that take rates should compress toward a common competitive level as platforms compete for the same supply and demand. The opposite is happening. Take rates across the talent-marketplace economy now span from literally zero to well above thirty percent, and the spread is widening, because the four marketplaces are pursuing fundamentally different value-capture strategies that reflect their different moats. Understanding the divergence is the key to forecasting which models survive.

At one extreme sits Braintrust's zero-percent talent take rate, a deliberate structural choice rather than a temporary promotion. By taking nothing from the worker and charging only the client (around 15 percent), Braintrust bets that superior worker economics will attract superior talent and that community token ownership will align everyone's incentives - Braintrust. The strategic logic is that in a market where supply quality is the scarce resource, the platform that pays talent the most wins the best talent, which wins the best clients, which is a virtuous cycle. The risk is that zero take on the talent side leaves the platform dependent entirely on the client fee and the token economy, a thinner and more volatile revenue base than a traditional two-sided cut. Braintrust is the boldest single bet in the market on the proposition that take rate should approach zero.

At the other extreme sits Toptal's premium take rate of well above thirty percent of the client's spend, sustained by a genuine curation moat and validated by the fact that Toptal built a 167-million-dollar bootstrapped business on it - GetLatka. Toptal's bet is the inverse of Braintrust's: that buyers will pay a large premium for the confidence that everyone in the pool is genuinely elite, and that this confidence is worth more than the savings from a cheaper platform. The fat take rate is what funds the rigorous vetting that creates the confidence, so the high price and the quality are causally linked rather than in tension. The risk is that as AI-driven vetting (the kind Mercor and Micro1 have built) becomes commoditized, the cost of credible curation falls, which could erode the premium that Toptal's take rate depends on.

In the middle sits the flat per-employee fee of the EOR category, settling around 599 dollars per worker per month at the challengers Velocity Global (Pebl) and Oyster, with Deel's scale economics pushing its effective per-worker cost down further - WhichPayroll. The EOR fee is structurally different from a percentage take rate: it is a flat cost per employee regardless of that employee's salary, which means it captures a declining percentage of value as salaries rise. This is a deliberate strategy that reflects the EOR moat: the value EOR delivers is compliance and infrastructure, which costs roughly the same whether the employee earns 40,000 or 400,000 dollars, so flat-fee pricing matches cost to value. The strategic consequence is that EOR economics improve with worker seniority (a high-salary worker generates the same fee at the same cost) and compress under competition, which is why the ~599-dollar figure has become a competitive anchor.

The AI-data-labor marketplaces occupy a fourth position that does not map cleanly onto any of the others, and that novelty is part of why they command such valuations. Mercor charges an hourly finder's fee and matching rate on top of what it pays contractors, capturing a spread on high-value cognitive work rather than a fixed percentage or flat fee - Marketing AI Institute. Because the work is high-value (expert humans paid well for specialized judgment) and the buyer is a well-capitalized AI lab with urgent demand, the platform can sustain healthy margins without the price sensitivity that compresses take rates in commodity freelance. The take-rate divergence, then, is really a map of moats: zero take where community alignment is the moat, premium take where curation is the moat, flat fee where compliance is the moat, and spread capture where AI-native matching of scarce expertise is the moat. The forecast in section 11 follows directly from which of these moats proves most durable.

9. Sizing the total addressable market to 2028

Having profiled the four marketplaces, we can now assemble a defensible total-addressable-market view and project it to 2028, with the explicit caveat that the four categories must be summed carefully because they overlap at the edges and use different sizing conventions. The broad gig economy is often cited at roughly 455 billion dollars for 2025, but that figure includes ride-hailing, delivery, and other categories outside the scope of this analysis, so it is the wrong number for sizing talent marketplaces specifically - Business Research Insights. The disciplined approach is to size each of our four categories from its own most credible source and present them side by side, because their growth rates differ so sharply that any blended number would mislead.

The four categories sized independently for 2025 to 2026 give a clear picture of relative scale and, more importantly, relative momentum. Classic freelance platforms are roughly 6.4 to 7.3 billion dollars and growing in the high teens at the category level but single digits at the incumbent level - Grand View Research. Employer of record is roughly 5.6 to 6.0 billion dollars and growing at a 7 to 15 percent CAGR depending on methodology - Business Research Insights. AI data labor, sized through its closest market proxy of data collection and labeling, is roughly 4.9 billion dollars in 2025 and projected to reach the high twenties of billions by the early 2030s at a CAGR near 29 percent, by far the fastest of the four - Coherent Market Insights. On-demand contingent is harder to size cleanly because it blends into traditional staffing, but the funded-platform layer is a low-single-digit-billions market growing steadily.

The chart makes the central thesis of the forecast visible: AI data labor crosses over the freelance and EOR categories within the forecast window, moving from the smallest of the four in 2025 to the largest or co-largest by 2028 if its near-29-percent CAGR holds. This is the quantitative expression of the valuation inversion. The freelance category is larger today, but it grows slowly at the incumbent level. The data-labor category is smaller today but grows roughly four times faster, so it overtakes within three years on these trajectories. A sizing exercise that ignored this crossover, and many do, because data labor is often filed under "AI infrastructure" rather than "talent," would completely miss where the marketplace economy is heading.

The crossover projection demands a skeptical stress test, because a forecast built on the fastest-growing line is the most fragile, and a research house must say so. The data-labor trajectory rests on continued frontier-model training spend, and there are two scenarios that would flatten it. The first is a shift toward synthetic data, where labs generate training data with models rather than buying human-labeled data, which would directly shrink the demand for human data labor. The second is a consolidation of demand into a few labs that bring data work in-house, as Meta's acquisition of a Scale AI stake hints at, which would compress the independent-marketplace opportunity. Against these risks, the bullish case is that each new model generation requires more sophisticated human evaluation (reasoning, agentic behavior, domain expertise) that is harder to synthesize, sustaining demand. The honest forecast carries both scenarios: the base case shows data labor overtaking the incumbents, but the bear case shows it plateauing if synthetic data or demand consolidation bites, and buyers and investors should watch those two variables above all others.

10. The value-chain map: who captures the dollar

To forecast where value accrues, it helps to trace a single dollar of labor through each marketplace and see who captures what, because the structure of value capture is what ultimately determines valuation. The four marketplaces split the dollar very differently, and the differences explain both the take-rate divergence of section 8 and the valuation inversion of section 9. Mapping the value chain is the bridge between the descriptive sizing and the forward-looking forecast, because it shows not just how big each category is but how much of each category's gross volume converts into platform value.

In classic freelance, the dollar of work splits roughly into the worker's earnings minus a low-teens platform cut, which is why Upwork keeps 788 million dollars out of more than 4 billion dollars of GSV - Upwork 8-K. The platform captures a modest, competitive slice because the work is relatively commoditized and the switching cost is low, so any attempt to raise the take rate risks losing supply or demand to a cheaper rival. This is the structural reason classic freelance generates large GSV but modest platform value, and why the public market prices the incumbents as profitable but mature. The value chain is long (many low-value transactions) and the platform's share of each is thin.

In employer of record, the dollar split is inverted in an instructive way: the worker captures the full salary, and the platform captures a flat fee on top that is unrelated to the salary size. This means the EOR platform's share of total spend is small for high-salary workers and larger for low-salary ones, and the platform's value comes from volume and from the compliance liability it absorbs rather than from a percentage of wages. Deel's 22 billion dollars of payroll throughput generating 1 billion dollars of ARR implies a blended take of roughly four to five percent of payroll volume - Deel. The EOR captures a thin slice of a very large flow, but the slice is defensible because the compliance moat is hard to replicate, which is why the category sustains a high valuation on a modest take.

In premium networks and AI data labor, the value chain is short and the platform's share is fat, which is the structural source of their superior economics. Toptal captures thirty percent or more of a high-value client dollar because curation is the scarce resource and buyers pay for it - GetLatka. AI-data-labor platforms capture a healthy spread on expert cognition because the buyer (an AI lab) is well-capitalized and urgent, and the supply (credentialed experts willing to do data work) is scarce and AI-matched at low marginal cost - Marketing AI Institute. The short value chain (fewer, higher-value transactions) combined with a fat platform share is exactly the profile the venture market rewards, which is why these categories command the highest valuations per dollar of GSV. The value-chain map thus predicts the valuation inversion: capital flows to the marketplaces that capture the largest, most defensible share of the most valuable dollars, and in 2026 those are the premium and AI-native models, not the commodity incumbents.

11. The forecast: four marketplaces, four trajectories

The forecast to 2028 follows directly from the sizing, the take-rate divergence, and the value-chain map, and it resolves into four distinct trajectories that anyone allocating capital or attention should hold separately rather than blending into a single "future of work" narrative. The four marketplaces are not converging toward a common outcome; they are diverging toward four different ones, driven by their different moats and their different exposures to AI. Stating each trajectory precisely is the payoff of the entire analysis.

Classic freelance: managed maturity with a bifurcating base. The forecast for Upwork and Fiverr is continued single-digit revenue growth, sustained profitability, and a steady erosion of the lowest-value work as AI substitutes for commodity tasks, offset by growth in AI-related and higher-value freelance categories. The incumbents will not collapse (they are profitable and entrenched), but they will not lead either, and the public market will continue to price them as mature cash generators rather than growth stories - Yahoo Finance. The premium and specialized layer of freelance (Toptal, A.Team) will outgrow the generalist base, so the right forecast is bifurcation: a slow-or-shrinking commodity tier and a faster premium tier within the same nominal category. Anyone forecasting "freelance" as one line will be wrong; the line is splitting.

On-demand contingent: steady density-driven expansion. The forecast for Instawork and Wonolo is continued market-by-market liquidity growth, improving margins as AI optimizes matching and scheduling, and gradual consolidation as the platforms that achieve deep liquidity in their core metros absorb or outlast weaker regional players - Instawork. This is the most predictable of the four trajectories because the physical constraint caps both the upside and the downside: AI cannot replace a warehouse worker, so the category is insulated from the substitution risk that threatens freelance, but it is also denied the parabolic growth of the AI-native models. The forecast is a healthy, unspectacular grind toward larger, more profitable, more consolidated regional marketplaces.

Employer of record: durable growth under a governance cloud. The forecast for EOR is continued solid growth (the conservative 7-percent and aggressive 15-percent CAGRs both imply a meaningfully larger market by 2028), led by Deel's scale, with the flat-per-employee fee compressing slowly under competition toward and possibly below the ~599-dollar anchor - Business Research Insights. The wild card is the Deel-Rippling litigation and the DOJ probe, which introduce genuine governance and reputational risk for the category leader and could, in an adverse outcome, reshape the competitive order in favor of Remote, Pebl, Oyster, or Rippling's own EOR offering - PYMNTS. The base case is durable category growth with Deel leading; the risk case is a leadership reshuffle if the criminal probe produces charges. The moat is the strongest of the four, but the leader carries a legal overhang the others lack.

AI data labor: the fastest growth and the highest fragility. The forecast for Mercor, Turing, Scale/Outlier, and Micro1 is the steepest growth curve of the four, with the category potentially overtaking freelance and EOR in absolute size by 2028 on its near-29-percent CAGR, sustained by relentless frontier-model training demand - Coherent Market Insights. But it is also the most fragile forecast, exposed to the synthetic-data risk, the demand-concentration risk, and the labor-and-regulatory risk crystallized in the DOL investigation of Scale AI - The Register. The base case is that data labor becomes the largest talent marketplace by valuation and possibly by size, on the strength of being on the critical path of the AI buildout. The bear case is a plateau if synthetic data displaces human labeling or if a few labs internalize the work. The honest forecast assigns the highest expected value and the widest variance to this category: it is simultaneously the most likely to dominate and the most likely to disappoint, and that combination is exactly what its 10-billion-dollar valuations are pricing.

12. What buyers, investors, and operators should do

Everything in this forecast converges on a set of decisions that buyers, investors, and operators actually control, and the purpose of research-grade sizing is to sharpen those decisions rather than merely to catalog the market. None of these audiences controls the AI-training boom, the EOR litigation, or the pace of synthetic-data progress. But each controls where to place capital, which marketplace to source labor from, and which model to build, and the four-trajectory forecast translates directly into guidance for each.

For the buyer of labor, the guidance is to match the marketplace to the job rather than defaulting to the most familiar platform, because the four marketplaces are now genuinely specialized. A buyer needing discrete project work should use classic freelance but should expect to do the commodity tier with AI tools directly and reserve the platform for vetted, higher-value work where human judgment still beats a model. A buyer hiring internationally should treat EOR as the default for compliant global employment, while pricing the governance risk of choosing a vendor under federal investigation into the decision and favoring vendors with clean records and exportable data. A buyer needing expert human data for an AI system should go straight to the AI-native platforms, which are purpose-built for it, rather than trying to assemble that supply through a generalist freelance marketplace. The mistake to avoid is using one marketplace for every job; the specialization is now real enough that the wrong marketplace means worse supply at a worse price.

For the investor, the guidance is to read the valuation inversion as a signal but to underwrite the fragility behind it. The capital has voted that AI data labor is the most valuable marketplace category, and the first-principles case (cheap intelligence shifts the bottleneck to the human judgment that trains it) is genuinely sound, which is why the category tops the scorecard. But the same analysis that justifies the excitement demands underwriting the synthetic-data risk, the demand-concentration risk, and the labor-and-regulatory risk before paying a 10-billion-dollar multiple - TechCrunch. In the more settled categories, the EOR moat is the most durable and the most underappreciated relative to its scandal-driven headlines, while classic freelance should be valued as mature cash flow rather than growth. The disposition that separates a good allocator from a poor one is to buy the durable moat at a fair price (EOR) and the fragile hypergrowth only at a price that respects its variance (data labor).

For the operator building a marketplace, the guidance is that the take-rate divergence is a strategy menu, not a problem to be solved, and the right choice depends entirely on which moat is achievable. If the achievable moat is community and alignment, the Braintrust zero-take model is coherent and can attract superior supply. If the achievable moat is curation, the Toptal premium-take model is proven and can be bootstrapped to scale on its fat margins. If the achievable moat is compliance and infrastructure, the EOR flat-fee model captures a thin but defensible slice of a huge flow. And if the achievable moat is AI-native matching of scarce, high-value cognition, the data-labor spread-capture model is where the largest valuations live. The fatal error is to pick a take-rate strategy that does not match the moat, charging a premium take rate without a curation moat, or running a zero-take model without a community moat, which leaves the platform with neither margin nor defensibility.

The single insight that ties this entire forecast together is the one a buyer, investor, or operator can act on immediately: the talent-marketplace economy is no longer one market with a few leaders; it is four markets diverging toward four outcomes, and the value has inverted from the platforms that move commodity labor to the platforms that move scarce, AI-relevant cognition. The freelance incumbents that defined the category are now its slowest-growing tier. The EOR leaders that looked like the next engine are growing fast but fighting in court. And a category that did not exist three years ago, AI data labor, now commands the highest valuations and the steepest growth in the entire economy, resting on a demand base that is as enormous as it is concentrated and fragile. The decision-maker who internalizes that inversion, and who matches marketplace to job, capital to moat, and take rate to defensibility, will navigate the next three years deliberately. The one who still sees a single "gig economy" will be navigated by it. For the workflow-level view of how AI is reshaping hiring itself, read our companion state of AI in recruiting, and for the vendor stack that sells to recruiters, our talent technology market map.

This forecast reflects the talent-marketplace economy as of March 2026. Valuations, funding, take rates, and the Deel-Rippling litigation are changing rapidly; verify current figures and legal status before any procurement or investment decision. The AI-data-labor category in particular carries unusual variance, and its trajectory depends on frontier-model training spend and the trajectory of synthetic data, which should be monitored closely.