Last month a former lab colleague sent me a screenshot of her weekly earnings. She holds a doctorate in computational biology; she now spends six hours a day writing reasoning traces for a frontier model at roughly $85 an hour, through a platform whose name she is contractually barred from disclosing. Two time zones away, the same pipeline pays Kenyan reviewers about $2 an hour to flag violent content. Same product. Same model. A spread of forty-to-one, and neither worker is an employee.
That ratio is the most honest number in the AI economy, and it is almost entirely absent from how crypto prices the sector. Over the past quarter I have watched a cluster of "AI data" tokens re-rate on a narrative โ decentralized annotation, expert marketplaces, DePIN for human feedback โ with no disclosed take rate, no worker retention data, and no verification layer worth the name. The market is pricing a labor supply it has never audited.
The underlying shift is real, and it is not about compute. Post-training โ RLHF, preference ranking, evaluation sets, red-teaming โ has become the binding constraint on model quality. Public web text is close to exhausted; the marginal gain now comes from credentialed judgment. This is why Scale AI's Outlier, Surge AI, and Mercor can bill laboratories $20 to $150 an hour for expert labor while Appen's BPO model keeps contracting. Meta's 2025 purchase of a 49% stake in Scale, roughly $14.3 billion for a minority position, was not a software bet. It was a bet on the supply of human expertise.
Crypto arrived with the familiar pitch: cut out the intermediary, tokenize the contribution, let workers own the network. Elegant on a whitepaper. In practice most of these networks have rebuilt the intermediary and deleted the accountability. The platform still sets the price. The lab still owns the output. The worker receives a token whose only terminal value is the next buyer's conviction โ which is to say, no terminal value at all. AI-data tokens are being underwritten as infrastructure. They are, structurally, labor marketplaces with a governance wrapper.
Three observations, ordered by how much they should change your positioning.
First, the expert premium is a scarcity price, not a wage. It exists because the data wall arrived before synthetic reasoning traces became reliable. That is a timing arbitrage, and timing arbitrages decay. My working estimate is an 18-to-24-month compression window โ faster in code and mathematics, slower in clinical and legal domains where liability still demands a human signature.
Second, the take rate is the real story, and almost no network discloses it. In the centralized model the spread between what a lab pays and what a worker receives can exceed 50%. Decentralized platforms rarely beat that; they frequently bury it inside emissions. Which brings me back to something I learned the hard way. In the summer of 2020 I spent forty hours tracing more than $50 million of inflows into early Compound deployments, expecting to find organic demand. I found printed incentives wearing demand's clothes. The current cohort of data networks runs the same architecture: token subsidies manufacture supply, inflated supply flatters the dashboard, and the dashboard recruits the next depositor. Liquidity is a narrative, not a metric โ and in labor markets, it is a narrative about wages.
Third, verifying expertise is a trust assumption, and crypto has relabeled it rather than solved it. When a network attests that a contributor holds a relevant doctorate, it relies on an issuer, an oracle, and a challenge mechanism โ three parties, two of them permissioned. That is the same architecture I criticized in LayerZero's verification stack: the decentralization lives in the diagram, not the dependency graph. What looks like noise is often pattern, and the pattern here is that every "trustless" credentialing layer resolves into a small set of humans who decide who counts as an expert.
Last year, modeling agent-driven flow, I watched bots react to a macro print in under 400 milliseconds and pull $500 million of volume out of a single DEX pool. The lesson was not that machines are fast. It was that the humans behind the machines had already decided what the number meant. Expertise is that decision. It cannot be tokenized; it can only be attested.
The consensus contrarian take is that synthetic data will collapse this market. I think that is the wrong variable, and the more uncomfortable point is different.
The expert gig worker is a gravedigger. Every reasoning trace she writes improves the model that will eventually price her out of the task. The industry has built a labor market whose product is its own obsolescence, and it has described the arrangement as opportunity while doing so. That is not a technology story. It is a story about who absorbs the transition cost โ and the honest answer is the people with the least bargaining power, in the jurisdictions with the weakest labor law.
Crypto had a genuine opening here: portable reputation, worker-owned provenance, payment rails that clear across borders in seconds rather than weeks. Instead the sector mostly issued governance tokens, which are non-dividend equity in a marketplace the holder does not control. The only exit is a later buyer. Structure survives where sentiment fades, and there is no structure in a token that confers no claim.
The durable asset is not the data. It is the evaluation harness and the provenance record โ the layer that can prove who produced what, under which terms, for which model version.
Watch three signals over the next two quarters: whether any data network publishes a take-rate disclosure without being compelled; whether the EU Platform Work Directive forces reclassification of expert annotators; and whether frontier labs internalize evaluation, which would strand the entire intermediary layer at once.
If the premium is temporary, the equity sits in the audit layer. If it is structural, the equity sits in the worker's bargaining position. Either way, the question is not how much these networks pay. It is who holds the ledger when the rate collapses โ and whether anyone outside the platform can read it.