The 750,000 AI Jobs Nobody Audited: Labor Data, On-Chain Infrastructure, and the Accounting We Keep Skipping

Over the past nine months, employers attributed 120,136 job cuts to artificial intelligence — roughly 21 percent of every layoff tracked in that window. Over roughly three years, the same narrative insists, AI has created more than 750,000 new positions. The two figures now travel together, in the same articles, as though they belong on the same ledger. They do not. One is a nine-month count of human beings who lost work. The other is a three-year count of job postings, scraped from a platform that also sells AI-skills training, and it was never reconciled against the first. When I read the original report — a piece that ran on a crypto news site and contained, so far as I could find, not one line about crypto — what unsettled me was not the optimism. It was the arithmetic.

I have spent my working life auditing numbers like these. In 2017, as a twenty-one-year-old undergraduate, I audited the governance contracts of early DAO prototypes and found three distinct voting-centralization risks inside a single project, which I documented across forty pages. The lesson was not that code is law. It was that a system's honesty is decided long before the code executes — in how it chooses to count. So when a labor story tells me that AI is simultaneously destroying and creating work, my first question is not whether this is true. It is what exactly is being counted, and what has been quietly left off the ledger.

That question is why this belongs on a chain-adjacent page at all. The AI labor market and the decentralized-infrastructure market are now selling the same thing — verified human and machine labor, priced and settled at scale — and they are repeating the same accounting mistakes. If we intend to build the rails for AI work, we should first be honest about what the existing spreadsheet conceals. A single headline number can carry a dozen incompatible meanings, and the distance between them is where reputations, policies, and tokens are quietly built on sand.

The three jobs hiding inside one word

Break the 750,000 number apart and the story changes shape. Among the disclosed categories, the largest single bucket — 282,000 roles, roughly 56 percent of the total — is data labeling. The second is data-center work, at 117,000, about 23 percent. Only then do we reach AI engineers, at 105,000, or about 21 percent. Read that again: more than half of the so-called new AI jobs sit at the supply-chain end of AI, and fewer than one in four is the high-skill engineering role the phrase conjures.

This is not a minor caveat. It is the whole story. Data labeling is supervised-learning and reinforcement-learning-from-human-feedback data supply — piecework, frequently outsourced, often paid by the task, and geographically concentrated in Kenya, India, and the Philippines. Data-center work is capital-expenditure-driven physical infrastructure: GPU clusters, cooling, power, networking. AI engineering is the generative-model and agent-deployment layer. These are three different technology generations, three different skill stacks, and three different labor economics, folded into a single triumphant number. When a headline says AI created 750,000 jobs, the reader pictures a $180,000 salary. The modal job behind that number may pay closer to a few dollars an hour.

I recognize the move because I have watched crypto make it. In 2020, during DeFi Summer, I spent three weeks reverse-engineering the yield-optimization logic of a farming protocol and found that its headline returns were largely token emissions dressed up as economic utility. The number was real. The meaning was not. The same sleight of hand is at work here: the category is real, the aggregate is real, and the impression it leaves is false.

There is a deeper technical confusion buried in the label. AI in this report is a single black box. It never distinguishes generative models from discriminative ones, or from conventional automation and robotic process automation. That matters, because the two claims at the heart of the piece — that AI is cutting jobs and that AI is creating jobs — actually point at different technology stacks. The layoffs are attributed to deployment of automation and generative tooling; the hiring is split between generative-agent work and, mostly, the labeling supply chain that feeds supervised learning. Lumping them together is like calling a mining rig and a validator the same machine because both are blockchain.

The undefined verb

Then there is the phrase the entire anxiety narrative rests on: uses AI. It is never defined. In the survey underpinning the report, using AI monthly may mean nothing more than opening a chat assistant — a consumer habit, not a professional capability. Deploying an agent, by contrast, is an engineering discipline with its own tooling, failure modes, and audit requirements. The report treats the two as the same skill and then draws a conclusion about layoff risk from the difference. The conclusion is only as meaningful as the definition, and the definition is absent. When the verb at the center of a statistic is undefined, the statistic is a mood, not a measurement.

I have made a version of this argument before, in a different register. In 2024, when the first spot Bitcoin ETFs were approved, I spent three months analyzing the custody arrangements of the major providers. The lesson was not that institutional adoption was good or bad. It was that the wrapper decides the risk, and the wrapper is never in the headline. A custodial ETF and a self-custodied node are both Bitcoin exposure. They are not the same thing. An AI-engineer salary and a data-labeling wage are both AI jobs. They are not the same thing either.

The median that was never weighted

The report's most-quoted statistic is the pay gap: a median AI salary near $180,000 against roughly $80,000 for the workforce at large — a premium of about $100,000, or 2.25 times. It is a clean, seductive number. It is also almost certainly a survivor's number.

A median is only as honest as the population it is drawn from. If the median is computed over the roles that make the most noise — AI engineers and architects — then it describes a small, elite subset while implying a market. If it is computed over all disclosed roles, then it should have been dragged down hard by the 282,000 labeling positions, which are overwhelmingly low-wage, gig-structured, and often offshore. The report never says which. The honest figure — the one weighted by actual headcount — is very likely a fraction of $180,000, and the distance between the two numbers is the size of the narrative being sold.

There is a second problem, and it is structural rather than statistical. The salary data originates from a professional network whose own commercial interest lies in selling AI-skills courses and certifications. That does not make the figure false. It makes the figure interested. When the entity publishing that AI jobs are booming also sells the course that promises to get you one, the burden of proof shifts — and the report never carries it. This is not cynicism. It is a basic auditing rule: the party who benefits from a number is the party who must show the work.

The narrative engineering of one percent versus three times

Watch how the two emotional statistics in the piece are chosen. On one side, a figure that only about one percent of laid-off workers listed AI as the primary reason for their dismissal — a number that softens the sense of direct technological displacement. On the other, the claim that low-frequency AI users face three times the layoff risk — a number that sharpens the sense of personal culpability. Both point in the same direction: downplay the technology, elevate the individual. That is not balance. That is a single thesis dressed as two findings.

Good reporting uses its statistics to complicate the story, not to converge on a predetermined moral. When every number in a piece happens to reinforce the same conclusion, the reader should reach for the ledger, not the conclusion. I have learned this the hard way. My own dissenting report during DeFi Summer was initially ignored by my team precisely because it cut against the number everyone wanted to believe. Being right did not make me popular. It only made me careful about which figures I let do the talking.

The timeline that does not line up

Now the accounting error that almost no one will say out loud. The layoff figure covers nine months. The job-creation figure covers about three years. Annualize the first and you get roughly 160,000 AI-attributed cuts per year. Annualize the second — even taking the 750,000 at face value and spreading it across three years — and you get about 250,000 postings per year. Set against each other on the same basis, the net creation shrinks from a triumphant 750,000 to something close to a rounding error. And that comparison still flatters the creation side, because a posting is not a hire. The report never once compares like with like, and that omission is the entire argument.

There is a sharper tell. The report states that mentions of AI in layoff announcements this year were more than double the whole of 2025 — a sentence that cannot be true, because the piece was published in November 2025, when the year was not yet over. Either the figure refers to 2024 and was transcribed incorrectly, or the reporting is unreliable at the level of its own sentences. I flag this not to score a point but because it is exactly the kind of error that survives when a story is assembled for traffic rather than for truth. The publication — a crypto outlet running a Web3-free AI-labor piece with a promotional footer — is a content-farm pattern I have learned to read from the byline down.

Meanwhile the denominator is nearly flat. September nonfarm payrolls added just 29,000 jobs, with unemployment at 4.2 percent. In a labor market that is essentially not growing, a sudden surge of AI jobs is far more plausibly a relabeling of existing roles than a wave of new ones. When total employment is flat and one category is booming, the most likely explanation is not creation. It is reclassification. Companies repackage existing technical roles as AI roles to attract talent and financing, and no job-board classifier can screen out that noise. The same reflex exists in crypto, where a wallet with a chatbot bolted onto it becomes an AI agent, and a staking wrapper becomes an infrastructure protocol.

There is also a geographic question the report never asks. If a meaningful share of the 282,000 labeling roles are performed offshore or as cross-border gig work, they are not straightforwardly American jobs at all, and the headline conflates a global labor supply chain with a national employment story. We have lived through this confusion in the crypto industry, where value is created across dozens of jurisdictions and then reported as though it belonged to one. Counting is never neutral. It decides who gets credit, who gets taxed, and who gets counted at all.

Where the chain actually touches this

Here is why I write this as someone who spends her days inside decentralized infrastructure rather than as a labor economist. Every category in that spreadsheet already has a crypto analog, and the analog is being built right now.

Data labeling maps directly onto decentralized labeling networks — token-incentivized protocols that pay contributors to tag, rank, and align training data. Compute maps onto DePIN GPU marketplaces that rent distributed hardware for training and inference. Agent orchestration maps onto on-chain agent frameworks that let autonomous software hold keys, transact, and coordinate. The 1,721 percent surge in agent orchestration postings — a figure that, yes, starts from a tiny base and should be read with suspicion — is the same demand that on-chain agent infrastructure is racing to serve.

So the AI labor market is not a distant topic. It is the market our protocols claim to be the settlement layer for. That is precisely why the accounting matters. If the underlying labor data is inflated, relabeled, and category-confused, then the token economies built on top of it inherit the same distortions — only now they are on a ledger, immutable and priced. A decentralized network that pays for low-quality labeling does not fix the labor problem. It industrializes it, and then it prints a token to celebrate.

This is where the Uniswap V4 lesson applies. When hooks turned the DEX into programmable Lego, the complexity ceiling rose faster than the developer base could follow; I have argued for two years that the spike would scare off most builders, and I still believe it. Agent orchestration is heading for the same cliff. The composability is real and the tooling is powerful, but the gap between you can build it and you can audit it is where users get hurt. Every new primitive multiplies the surface area for error, and every error in an AI-adjacent protocol is a labor question, because the thing being settled is someone's work.

There is a valuation layer the report ignores entirely. It never prices a single AI company or product; its only investor is Mark Cuban, whose optimistic line — that AI will not take your job, but someone using AI will — is itself a position with commercial interest. Cuban is an active technology investor and a media figure, and the reassurance that AI is more opportunity than threat aligns neatly with the narrative interests of a technology portfolio. That does not make him wrong. It makes him partial. And the report presents his partial view as the emotional resolution of the whole piece, without ever disclosing the stake. The only investment the data actually supports is a human-capital one — a $100,000 annual premium for AI skills — and even that carries a time-lag risk: the individual who invests today in a skill may find the skill automated before the investment pays back.

That last risk deserves its own question, the one the report never asks. Will AI automate the AI engineers? Every year the tooling gets better at writing, testing, and deploying code, and the same agent-orchestration demand that is booming today could be served by a smaller, more senior team tomorrow. If that happens, the premium that justifies the entire skills narrative compresses within a single hiring cycle. The spreadsheet treats today's wage as a durable fact. It is more likely a snapshot of a moment before the market reprices the skill.

And the compute layer carries externalities the spreadsheet never prices. Those 117,000 data-center jobs correspond to real power draw, real water consumption, and real friction with the communities that host the hardware. I have watched DePIN projects market themselves as frictionless infrastructure while quietly exporting the same environmental costs that centralized cloud providers are criticized for. A distributed network does not automatically make an externality disappear. It sometimes just distributes the blame. If we are going to claim the compute layer, we should be willing to account for what it costs the places it touches.

The individual-responsibility story is the oldest trick in the book

Now the contrarian turn, and it is the part that genuinely worries me.

The emotional center of the report — and the sentence that earns the headline — is, in spirit: it will not be AI that takes your job, but a person who uses AI better than you. It is memorable, shareable, and a masterclass in offloading structural risk onto individuals. The supporting statistic, drawn from a Gallup survey, is that tech workers who use AI less than once a month are three times more likely to be laid off. Three times. That is a striking figure, and it is almost certainly backwards.

Consider the reverse causality. Who is least likely to use AI at work? People in roles where AI is not relevant, not permitted, or not provisioned — which is to say, people already at the margin. The finding may not be that not using AI gets you fired. It may be that people who were already vulnerable were also the people who were never given AI tools. The report packages a correlation as a cause, and the cause conveniently assigns blame to the worker rather than the employer. It also generalizes from a sample of tech workers to the whole labor force, ignoring everyone whose job will never involve a chat window.

I have seen this narrative before, in crypto. When the bear market of 2022 arrived and my firm laid off 40 percent of its staff — including my mentors — the industry's reflex was to tell the survivors to learn to code, to upskill, to stay bullish. The advice was not wrong, exactly. It was just a way of converting a collective failure into a private anxiety. I spent that winter writing a newsletter, The Quiet Chain, producing twenty-four deep dives on Layer 2 scaling for about five thousand subscribers who valued consistency over hype. What I learned is that resilience is real, but it is not a substitute for structural honesty. Telling individuals to adapt is not the same as admitting that the system failed to distribute the transition.

The same evasion now surrounds data labeling. The 282,000 labeling roles are counted as AI's gift to the economy, but the report never mentions their wages, their employment form, or their geography. If a meaningful share of those roles are gig work, offshore, or paid by the task, then 750,000 new AI jobs is describing a labor-arbitrage structure and calling it opportunity. That is not a side note. It is the moral core of the dataset, and it is the one thing the report declines to examine.

There is a technical reason this matters beyond fairness. Labeling quality — including the human feedback used to align models — directly shapes how safe and unbiased those models become. Underpaid, high-turnover, culturally mismatched labelers do not merely get a raw deal. They degrade the alignment layer of the systems everyone else will depend on. The conscience of the supply chain is not separable from the capability of the product.

And here is the trap for our own house. The KYC that most projects bolt onto their labeling and contributor networks is, in my experience, largely theater — a few wallet checks and a form that honest users pay for while bad actors route around it. The compliance cost lands on the people who were already playing straight. If decentralized AI networks reproduce that pattern, they will not be more ethical than the platforms they replace. They will be the same ethics with a token attached. The hashrate analogy is exact: after the fourth Bitcoin halving, miner revenue collapsed, and hash power drifted toward a handful of pools until decentralized consensus became a phrase with quotation marks around it. Concentration is not prevented by architecture. It is prevented by continuous, uncomfortable governance — and governance is the first thing a token sale sacrifices.

Build not for the peak, but for the plain

So what should you actually watch, and what should you actually build?

Watch three signals. First, the next BLS employment report: if AI-attributed hiring keeps rising while overall employment stays soft, the structural reallocation thesis gains ground and the net creation story loses it. Second, the monthly Challenger layoff data: track whether AI's share of cited reasons keeps climbing, or whether it was a convenient label for ordinary cost-cutting dressed for the capital markets. Third, the capital-expenditure guidance of the major cloud providers: data-center jobs are the most cyclical category in the whole spreadsheet and the first to shrink if AI infrastructure spending cools. Those 117,000 roles are not a foundation. They are a bet on a spending cycle.

Build differently. If you are constructing decentralized labeling or compute networks, the durable moat is not the token — it is the quality and verifiability of the work and the dignity of the people doing it. Pay by outcome, publish the wage distribution, and make the labeling supply chain auditable end to end, because that is the only version of this that survives contact with regulators and with users. If you are building agent orchestration, treat auditability as the product, not the afterthought. The complexity will arrive whether you plan for it or not; the only question is whether anyone can check the result.

I keep returning to a line I wrote years ago, after spending two months interviewing fifty women digital artists who had been systematically excluded from the crypto boom, and turning their stories into a series that reached ten thousand readers and seeded a women-focused grant fund. The technology was never the hard part. The hard part was deciding whose work counts. We audit the code, but who audits the conscience?

The 750,000 number is not a lie. It is a category error with a marketing budget. And the reason it should trouble anyone building on-chain is that we are about to make the same mistake at settlement speed. The ledger will record every wage and every task, faithfully — which is exactly why we cannot afford to be careless about what we ask it to record. Decentralization is not a product you ship; it is a practice you keep. Build not for the peak, but for the plain.

The jobs are coming. The question no dashboard will answer is who gets to define them.