The AI Trade: Winners, Losers, and Defensible Ground

Part 2 of a three-part series on AI's impact on work. Where the productivity gains actually go, who captures them, and the four positions that still hold. A practitioner's frame for thinking about AI winners and losers without pretending the squeeze isn't real.

The question Part 1 left open

There's a small café across from where I live in Da Nang. Most mornings I sit there, watching two people work who never speak to each other. Inside the café, a few tables over, someone is on a laptop with a Claude Code session open, building software. Across the street, in front of a motorbike repair shop, a mechanic works on a bike on the sidewalk, hands black with grease. Both make a living. In five years one of them will be competing against a million people doing the same thing from anywhere on earth. The other will be competing with the mechanic two streets over.

I'll come back to those two at the end.

Part 1 of this series ended on a question I didn't fully answer: if the productivity gains AI generates don't stay with the person creating them, where do they actually go? The data was clear that they leak outward, to platforms, clients, competitors, the market's redefinition of normal. But "outward" isn't a destination. This part tries to be more specific. Who captures the AI trade, who pays for it, and what positions a working person can still defend without pretending the squeeze isn't real.

One scope note before the analysis. The empirical evidence in this essay is largely from US and EU labor markets, because that's where the macro-data is cleanest. The squeeze itself isn't entirely new. Globalized freelance markets in the 2010s (Upwork, Fiverr, Toptal) already pressured US and EU middle-skill cognitive workers, who found themselves competing against equally-qualified workers from lower-cost regions, including India, the Philippines, and Eastern Europe. AI is the next layer of that same compression, not a fresh shock. Most workers globally were already inside this dynamic. AI extends and intensifies it rather than starting it. What changes with AI is the speed and the geographic distribution of the pressure, not the shape of the trade itself.

Where the money is going

Start with the cleanest signal we have: where capital is moving. Venture flows are an imperfect proxy, but they tell you in real time which bets the people with the most money to lose are willing to make.

In 2025, AI firms attracted 61 percent of all global venture capital, $258.7 billion of $427 billion total, according to OECD analysis of Preqin data. That's roughly double AI's 2022 share. By Crunchbase's classification, the first quarter of 2026 saw AI take roughly 80 percent of all global VC, with four companies (OpenAI, Anthropic, xAI, and Waymo) collectively raising $188 billion, around 65 percent of the entire quarter's global venture. The single largest category inside AI is infrastructure and hosting, which alone pulled in $109.3 billion in 2025.

Two things are worth noticing. First, capital is concentrating, not just shifting. Deal counts are going down, check sizes are going up. Most of the money is going to a small set of frontier labs and the picks-and-shovels companies (the chipmakers, data center operators, energy suppliers) that supply them with compute, chips, and physical infrastructure.

Second, the people who use AI to do their work are not the people receiving this capital. The capital is going to the people who own the AI substrate: the underlying compute, models, and platforms that everything else builds on top of. Most of the savings a freelance developer or copywriter or small company captures by using AI tools, over time, drain toward that substrate. Some get retained by the user, some go to clients through price competition, some leak to competitors. But the structural drift is upward, toward whoever runs the toll booth. The toll booth wins whether the road is going up or down. To be clear about the limit of that claim: the capital concentration is observed and the labor pressure is observed, but the causal chain between them (worker productivity gain → price compression → margin migration upstream) is inferred from the macro pattern, not yet traced end-to-end in any single dataset I've seen.

The three layers of the AI trade

Part 1 used a three-layer model (individual, firm, market) to describe how the productivity paradox shows up at each scale. The same three layers help describe who actually captures the AI trade. From the top of the stack down: infrastructure, platforms, and the small number of top-tier practitioners who manage to extract a premium for using these tools well.

Infrastructure and capital

The cleanest position is at the top of the stack. Foundation labs, the chipmakers, the data centers, the energy supply, the sovereign wealth funds writing checks no traditional VC can match. These businesses don't compete with the people using AI; they sell to them. Every prompt is revenue. As more people get more productive with AI, more compute gets bought. Jevons paradox, which Part 1 borrowed from coal (the observation that more efficient use of a resource often increases total demand for it), shows up here in its purest form: efficiency at the user end produces volume growth at the supply end.

Two honest caveats. First, the AI infrastructure capex cycle has been challenged through 2025 by questions about ROI on enterprise AI deployments and whether the data center buildout is over-capitalized. If model commoditization continues (open weights, falling inference costs), value migrates down the stack toward platforms and applications. The infrastructure layer is currently the cleanest position. It is not a guaranteed one. Second, this layer isn't accessible to most people reading this. You don't pivot into being a hyperscaler. You can buy the equity through public markets, which is a real option but not a labor-market strategy. The first layer is mostly relevant as a description of where the money goes, not a position you can take.

One thing worth widening before moving on. "Infrastructure winners" isn't only the VC-funded frontier-lab list. Established incumbents that integrated AI into existing distribution (Microsoft via Copilot and Azure OpenAI, Google via Gemini-in-everything, Adobe via Firefly, Salesforce via Einstein, Oracle via embedded models) are capturing significant AI revenue without showing up in venture data, because they aren't venture-funded. Public-equity holders capture some of this diffusely through index-fund concentration in NVIDIA, Microsoft, Google, and Meta, which is how AI gains reach pension systems and sovereign wealth without anyone consciously buying "AI exposure." And non-Western infrastructure (Chinese open-weight models like DeepSeek and Qwen, plus the parallel Chinese AI compute stack) is a separate and growing power center that the US-led VC narrative tends to underweight. The infrastructure layer of the AI trade is broader than the four-lab list above suggests.

Platforms and aggregators

One layer down sit the marketplaces and aggregators, and their economics are subtler but in some ways more revealing. Upwork, since May 2025, charges a variable freelancer service fee of 0 to 15 percent per contract, plus client-side marketplace fees up to 7.99 percent, plus connect fees per proposal, plus a 13.5 percent conversion fee if a client wants to take the relationship off-platform. Real-world blended take runs 11 to 13.5 percent for most freelancers (per GigRadar's agency dataset, with the same 12 to 13 percent range independently confirmed in Jobbers.io's analysis of 2025 user surveys), with total platform costs reaching 14 to 20 percent of gross earnings once you add withdrawal fees and connect spend.

What changes when AI lets ten times as many people offer the same service? The supply side floods, per-unit prices fall, and the platform's volume goes up. The take rate stays the same in percentage terms but covers a much bigger market. Platform revenue grows even though every individual freelancer on it is making less per hour. The platform profits from the same dynamic that pressures the people on it. This isn't a moral failure of platforms; it's structural. Marketplaces capture a slice of every transaction, and AI makes the number of transactions go up.

Descriptively, this is the second-largest layer of capture after infrastructure. Prescriptively, it's mostly closed: building a platform requires capital and patience, and most attempts fail before reaching critical density. The Platforms layer matters here as a description of where the value goes, not as a position the average reader can walk into. The places to actually stand are in the next section.

Top-tier practitioners

The third layer is where the AI-positive narratives mostly live, and the best evidence for it is real but narrower than people usually claim. Two studies anchor the optimistic view. Demirci, Hannane and Zhu (Management Science, 2025) found AI-specialist freelancers extracting 25 to 60 percent rate premiums on Upwork over their non-AI peers. Althoff and Reichardt at Stanford (January 2026) modeled AI substantially reducing wage inequality while raising average wages by about 21 percent, with the gain driven by a "simplification" channel that lets workers across skill levels compete for the same jobs. Both findings are real. Both are also measurements of where the squeeze is sending the surviving share of the labor force, not evidence that the squeeze is benign.

The Demirci premium is a cross-section of freelancers who adopted AI and stayed on the platform; it does not include the freelancers who were displaced, left, or never adopted. The Althoff–Reichardt figure is a long-run model prediction conditional on the assumption that AI's main labor channel is task simplification rather than displacement, which is itself contested. Read carefully, both say the same thing: among workers who remain in cognitive work, the ones who learn to orchestrate AI capture a premium relative to those who don't. That's a real finding about the destination of the trade. It is not a finding that cognitive work is a defensible place to stand. Whether the destination keeps existing depends on questions the data can't answer yet, which is the next section. One thing the data does clarify, though: the premium for hybrid AI-using practitioners isn't a contradiction with the price compression in commodity cognitive work. It's the consequence of it. As the AI-commoditized baseline gets cheaper, the human-judgment overlay above it becomes more visible by contrast. That's the mechanism the Zone 4 walk-through later in this essay will describe.

The squeeze on the middle

If the most defensible positions are at the top of the stack, the most exposed are in the middle. This isn't a new pattern. It's a recurrence of one labor economists have been documenting for decades.

The IMF's 2024 working paper on AI and US local labor markets found that the negative employment effects of AI exposure fall mainly on middle-skill workers, in a pattern that mirrors the routine-biased technical change story (the labor-economics finding that earlier waves of automation hollowed out routine middle-skill jobs while leaving the very bottom and the very top relatively intact) Autor, Goos, and others traced for the 1990s and 2000s. Deming, Ong, and Summers (NBER Working Paper 33323, 2025), looking at 124 years of US Census data, identified a sharper recent shift. Through the 2000s, US employment growth was "barbell-shaped," gaining at the top and bottom of the wage distribution and losing in the middle. Since 2016, that pattern has been replaced by what they call a "one-sided ramp": between 2016 and 2022, the share of US employment in low-skill and middle-skill occupations each fell by roughly 2 percentage points, while high-skill occupations grew by slightly more than 4 percentage points. STEM employment specifically rose from 6.5 percent of all US jobs in 2010 to nearly 10 percent in 2024, a 50 percent relative increase. The middle isn't just stagnating relative to the ends. It's compressing inward.

Layer this on top of the freelance data Part 1 covered. Hui, Reshef, and Zhou (2024) found AI-exposed freelance categories on Upwork lost 2 percent of contracts and 5.2 percent of earnings after ChatGPT launched. Teutloff and co-authors (2025) found substitutable skill clusters falling 20 to 50 percent. Demirci's paper found writing roles down roughly 30 percent and basic software development down about 20 percent on a global posting dataset. The picture from above (macro labor data) and from below (platform data) tells the same story.

And the corporate layer is starting to add the same signal. Block (Square's parent, founded by Jack Dorsey) announced a workforce cut of nearly half (10,000 to under 6,000) in February 2026, with Dorsey directly attributing the change to AI in a shareholder letter. IBM laid off roughly 8,000 employees in 2025, primarily from HR functions, replacing them with an internal AI agent called AskHR. CrowdStrike cut 5 percent of its workforce, with the CEO citing AI-driven efficiency. Amazon followed its 14,000 corporate cuts in October 2025 with another 16,000 in January 2026, after spending around $80 billion on AI capex during 2025. Aggregating across companies, roughly 55,000 US tech workers were laid off in 2025 with AI cited as a direct cause, and the share of layoff announcements explicitly attributing the cuts to AI rose from under 8 percent in 2025 to over 20 percent by early 2026. The macro-statistics will keep underselling this for a while, because they always lag the layer-by-layer cuts inside firms. The squeeze is showing up in firm decisions before it shows up in BLS aggregates.

Three caveats before this becomes a class-collapse narrative.

First, this is directional pressure, not a guarantee of mass unemployment. Aggregate employment effects so far are modest. The Dallas Fed found roughly a 1 percent employment decline in the most AI-exposed sectors. The squeeze shows up first in earnings and growth rates, then in entry-level openings, then in employment, with long lags. Most middle-skill workers won't be fired by AI. They'll find that their wages don't grow, their hours get cut as duties consolidate, and the next entry-level person doesn't get hired behind them.

Second, polarization is partly an institutional outcome, not just a technology outcome. The OECD's 2024 paper on AI and wage inequality, looking at 19 OECD countries, finds AI's measured impact on inequality has been muted so far, and where it does push inequality, the effect is moderated by collective bargaining institutions. Countries with stronger labor coordination see narrower wage gaps under the same technological pressure. The technology is a pressure. The institutions decide how much of that pressure reaches the worker. The institutional response isn't only theoretical, and it's not only Western. In April 2026, Hangzhou Intermediate People's Court in China upheld a ruling that a tech company's AI-driven dismissal of a senior worker was illegal, and that AI replacement does not by itself qualify as legal grounds for termination. A separate Beijing arbitration panel reached the same conclusion for a map-data-collection worker in late 2025. Both cases set the precedent that AI cost savings cannot be unilaterally shifted to employees as job loss. China is not the country labor-rights observers usually look to first, which makes the rulings particularly informative about where the institutional question is actually live. Part 3 returns to the institutional and geopolitical layer in depth, including the race-to-the-bottom dynamic between countries with different labor regimes. The takeaway here: the squeeze isn't fated. It's a default outcome of AI plus weak labor-market institutions.

Third, the polarization story above isn't the only serious account of what AI is doing to work. Two flanks deserve naming. The pessimistic flank: Daron Acemoglu ("The Simple Macroeconomics of AI," 2024) argues that AI's productivity gains will be smaller and displacement larger than the routine-biased technical change framework predicts, because AI climbs the cognitive ladder rather than just substituting for routine work. If he's directionally right, the "top-tier judgment-layer practitioner" position (the worker whose value lies in judgment, taste, and accountability rather than the production output itself, expanded in Zone 4 below) compresses next, not last. The optimistic flank: David Autor's recent work and Erik Brynjolfsson's J-curve thesis suggest that mid-skill cognitive work may re-expand under AI complementarity the way clerical work expanded after computing rather than disappearing. If that flank is right, the squeeze is transitional rather than terminal.

My read: Acemoglu is overweighting near-term displacement and underweighting deployment friction (hallucination rates that still demand human oversight, capex skepticism, the slower-than-promised path to autonomous agentic AI). Three signals will tell us if I'm wrong over the next 18 months. First, enterprise tolerance for hallucination: does it drop (indicating real autonomy) or stay flat (indicating continued oversight)? Second, the price of inference for frontier-capability tasks: does it crash the way commodity-token pricing has, or stay sticky? Third, entry-level openings in judgment-layer roles: do they dry up the way commodity cognitive work has, or hold? If those move toward Acemoglu, Zone 4 has a shorter half-life than this essay treats it. If they move toward Autor and Brynjolfsson, the squeeze itself is shorter than treated. The four-zone frame should be read with both edges in mind.

Four zones that still hold

The four zones below intersect with the three layers above but answer a different question. The layers describe where AI value is captured at the macro level (infrastructure, platforms, top-tier practitioners). The zones describe what positions are available to take from the ground level. Zone 4 (Hybrid) is where most top-tier-practitioner premiums actually live; Zone 1 (Control) overlaps with platform and owner economics. The zones aren't a separate model. They're the macro picture rotated to the angle a working person stands at.

The map, before the walk

What follows is a working frame I keep returning to in conversations with operators, not a categorical taxonomy. The zones overlap. A doctor with AI tools (Zone 4) can also own a clinic (Zone 1). A massage therapist (Zone 2) operates in physical space (Zone 3). The point isn't crisp membership, it's the shape of the tradeoffs each kind of position involves.

If the most pressure lands in the middle of cognitive work, the question is where the pressure is lower. Not zero. Lower. Four positions, each with its own kind of compromise. Here's the map first:

The four zones differ on where AI substitution risk sits, what the upside looks like, and what kind of pressure each one trades for the other.

Now the walk through each.

Zone 1: Control and capital

Owning the result, not just the work. Founders, product owners, people who run systems instead of operating inside them. Productivity gains stay with you instead of leaking outward. If your AI tools cut delivery cost by 40 percent, your margin goes up; you're not racing a competitor to lower the price by 40 percent because, if you're well-positioned, you don't have to.

The honest tradeoff: this position requires capital, risk, and tolerance for variance. Most people who try fail. Founder selection bias (where successful founders are overrepresented in advice-writing because the failures aren't around to talk) applies double here. The founders writing about this stuff are by definition the ones who survived long enough to write about it. The other failure mode worth naming is the thin-wrapper trap: a founder whose product is mostly a UI layer over someone else's model loses the toll-booth position fast, because the substrate underneath them captures the real value while their margins get arbitraged by the next thin wrapper.

Zone 2: Human-to-human work

Massage therapists, coaches, therapists, eldercare workers, trainers. Anyone whose value depends on physical presence and trust. AI can imitate intelligence reasonably well at this point. It can't yet imitate the experience of being in the room with another person who is paying specific attention to you. The 2025 ILO update on generative AI exposure put roughly a quarter of global jobs in some exposure category, with core human-presence services among the least affected, because the value the buyer is paying for is the human presence, not a simulation of it.

The tradeoff is in the ceiling. These roles don't scale. Income is bounded by the number of hours one body can show up. The floor is more stable than in cognitive work; the ceiling is much lower. The counter-case: presence-based work that gets mediated through a screen (tele-therapy, online group coaching, app-based wellness) erodes the premium fast, because the thing being sold stops being presence and starts being content. The protection is in the room, not in the role title.

Zone 3: Physical world

Construction, repair, agriculture, logistics, manufacturing. Work that has to happen in physical space, with tools, on actual objects. AI accelerates digital output. The physical world remains constrained by atoms, time, and the slow business of moving things.

This zone bought time but didn't buy immunity, and the time it bought is finite. Amazon now operates roughly a million robots in its warehouses, mostly non-humanoid, with a generative AI fleet-coordination system Amazon claims is cutting fleet travel time by about 10 percent. Agility Robotics' Digit is in paid pilots with GXO and Amazon. Figure AI's robots are testing at a BMW plant. Gartner's January 2026 forecast says fewer than 100 companies will move humanoid robots beyond proofs of concept by 2028, and fewer than 20 will reach production scale in supply chain operations by then. The honest range, with broad disagreement among forecasters: two to ten years before meaningful displacement reaches the easier sub-segments. Tightly structured environments (warehouse picking, palletizing, predictable conveyor work) sit at the near end of that range. Outdoor unstructured work (sidewalk repair, open-field agriculture, varied-site construction, anything requiring real-time adaptation to unpredictable physical contexts) sits much further out, plausibly decades. The direction of travel is clear. The buffer differs by an order of magnitude across what we're calling "physical work," and Zone 3 is not one zone but a gradient.

Zone 4: Hybrid roles

The most interesting zone, and the one I'd bet most heavily on for cognitive workers reading this, is the hybrid. A doctor with AI diagnostic tools. A consultant with AI research and analysis. A designer with AI generation and a personal brand. A trainer with AI-driven performance analytics. A lawyer with AI research and a client base who chose them specifically.

What unites these is that AI doesn't substitute for the human; it amplifies them. The judgment, accountability, taste, and relationship layer is still legibly human to the buyer. The production layer underneath is now AI-accelerated. They get faster, deliver more, and can charge more, not less, because their unique value is more visible against an AI-commoditized baseline.

What this looks like in practice, from my own work: most of what I do now is delegate implementation to AI agents while I hold the architecture, the design choices, and the accountability. The agent writes the code; I decide what's worth building, where the risks sit, and whether the output is right. My role shifted from author of the work to author of the decisions about the work. The skill that used to define my contribution (writing things myself) became less important than the skill that organizes the AI doing it. The Zone 4 premium is real today. So is the constant pressure to keep climbing as the boundary between the human judgment layer and the AI production layer keeps moving downward.

Two honest critiques of that anecdote and this whole zone. First, the same survivorship lens I applied to Demirci and Althoff in §3 applies to me here. I'm one data point from someone who happens to be surviving in the position I'm describing as defensible. I don't see the people who tried this same shift to architect-and-orchestrator and gave up, took a salaried implementation role, or lost their client base to someone faster. Treat the anecdote as one practitioner's lived experience, not as evidence the position is durable for the next person who tries it. Second, the Acemoglu risk from §4 lands hardest in this zone. If AI's comparative advantage keeps climbing the cognitive ladder, Zone 4 is a temporary plateau, not a defensible position. The premium is real today. Whether it's still here in five years is the open empirical question of this whole essay, and the data won't settle it for a while.

Self-locating in the map

Four questions to find your default zone, in rough order of usefulness. None of these have a single right answer; the point is that your gut response usually tells you where you're standing without you having to score it.

1. What did your effective hourly rate do over the last 18 to 24 months, controlling for your AI adoption? If you're using AI tools more and your real hourly is flat or down, you're closer to the squeeze than the premium, regardless of what the role is called on your tax return. If your hourly is climbing faster than your input hours, you're in or near a defensible zone.

2. If the AI tools you use today disappeared, would your value proposition survive? In Zones 1, 2, and 3, mostly yes (the tools are recent additions to a position that already had its own basis). In Zone 4, often no. The production layer the AI now does was the work, and what's left is just judgment without the throughput. That fragility is the cost of the premium.

3. If your top three clients or revenue sources disappeared tomorrow, how long would it take to replace them at your current rate? Long replacement time signals platform/marketplace dependency; that's middle-of-the-squeeze territory, where the buyer is interchangeable and so are you. Short replacement time signals direct relationships, which is where Zone 4 and Zone 1 actually live.

4. Would your work survive 50 percent cheaper inference and noticeably better models 12 months from now? If yes, you're in a slower-decaying zone. If no, the position is more exposed than its current pricing suggests.

The point of these isn't to score yourself onto one of four buckets. It's to notice which kind of pressure your current position is most exposed to, which adjacent position has lower exposure, and what would have to change for the move to be possible. The zones are coordinates. The questions are how you read your own coordinates.

Where the zones break

A taxonomy that doesn't admit its limits is just a sales pitch. Three places this one weakens are worth being explicit about.

First, pockets of efficiency are temporary. A month ago I built an HR-process automation for a client. The HR person genuinely captured free time from it, in the actual extra-hours-on-the-couch way. It was a real win. It was also fragile. As soon as the asymmetry becomes visible to management, to a competitor, or to the next client looking at the price, the gain gets repriced. AI advantages compound for a while and then get arbitraged. The four zones are positions where the arbitrage runs slower, not positions where it doesn't run.

Second, there's no zone you can fully relax in. Owners face risk and capital exposure. Service workers face physical strain and income ceilings. Physical workers face the slow approach of robotics. Hybrid workers face permanent re-skilling. The choice is which kind of compromise suits your temperament, capital, and life stage. "Defensible" in this context means slower-decaying, not invulnerable. Some kinds of pressure compound at lower rates than others, and that difference matters over a working life even though none of these zones lets you stop running.

Third, the four-zone frame doesn't account for jurisdiction. Where you live changes which zones are available and how much pressure each one absorbs. A hybrid worker in a country with strong labor protections and high collective bargaining coverage faces a different version of the squeeze than the same worker in a low-coordination economy. The full geopolitical version, including the race-to-the-bottom dynamic between countries that protect labor and those that push efficiency hardest, is Part 3's territory.

Where the gains went, and what's left

Part 1 closed on the question, productive for whom. The picture this part lets us paint, finally, is sharp enough to act on. AI winners and losers map cleanly onto the layers and zones above: the gains pool in the substrate (compute, frontier labs, platforms), the squeeze lands in the middle of cognitive work, and the zones describe four imperfect places to stand instead. They aren't safe positions. They're positions where the pressure decays slower.

Back to the two people in the café from the opening. The clean dichotomy was that the developer competes globally and the mechanic competes locally. The four zones complicate it both ways. The developer who treats AI as a hybrid (Zone 4) and holds judgment over the work, or who pushes into Zone 1 and owns the result, isn't really competing with a global million; they're competing with the much smaller pool in the same hybrid or owner position. The mechanic is in Zone 3, and his buffer is significantly longer than a warehouse worker's. Outdoor unstructured repair sits well behind structured logistics on the robotics deployment curve. He's not watching robots arrive next year. He's watching them arrive at warehouses next year, with his own work much further out. Same scene, different time horizons. The question isn't who's winning. The question is who am I actually competing with, and over what, and the answer keeps moving.

The harder thing this part doesn't fully answer is what to do with the realization that even the winning positions are still positions inside the system, not outside it. The hybrid doctor earns more and works faster and harder. The owner captures the gains and runs an infinite backlog. The physical worker is in the slowest zone today and watching robots arrive next year. The human-to-human worker has the lowest substitution risk and the lowest ceiling. Each zone slows the rate at which downward pressure reaches you. None of them lets you step outside it. Part 3 takes that observation seriously and asks what an accelerating system without an off-switch is actually accelerating toward, and whether there's any version of "enough" that doesn't get rewritten every six months.

For now, the practical content of this essay is small and not very comforting. The AI trade has clear winners. By and large, they're not the people doing the work. The middle of cognitive work is being squeezed in a way the macro statistics will keep underselling for a while, because they always do during transitions. There are positions that hold up better than others, and the four zones are one way to think about which. None of them is safe. Some of them are slower-moving. Pick the kind of pressure you can carry, build something durable inside it, and watch what the structure does next (the three signals from earlier are a good place to start), because the answer to "productive for whom" isn't a number on a spreadsheet. It's a position. You're already in one.