Serval spends seventy thousand dollars a head to steal engineers it refuses to train.
The bezzle is John Kenneth Galbraith’s coinage, and it is worth being precise about what it names, because the AI labour market is being read this week as a triumph of scarcity, and the reading is accurate in every detail except the one that matters. The bezzle is the interval between the commission of a fraud and its discovery — the stretch when the embezzler has his gain and the victim does not yet feel poorer. A Wall Street Journal report out this week documents the compensation layer of the interval we are living in: Serval, an AI startup competing to fill a hundred engineering roles this fall, is offering a $30,000 signing bonus, $30,000 to any employee whose referral gets hired, and a $10,000 vacation added on top of the offer. Seventy thousand dollars a head, for a hundred heads. Every dollar in that package is real. The question is what the dollars are buying, and it is not what the headlines say.
It is true that corners of the labour market are genuinely hot, and the numbers are honest ones. Engineers who can put the current generation of large language models to work are genuinely scarce, and hiring beat expectations in October, with 165,000 jobs added. Payscale has the technology industry topping median wage growth in the third quarter — the same industry that has taken the brunt of U.S.-based job cuts this year — and its survey work shows the premiums concentrating in labelled specialties, agentic AI by name. Nela Richardson, ADP’s chief economist, has watched the same heat and located it precisely: “a ramp-up in salaries and employment in the most specialized parts of the tech sector.” Mohit Bhende, who runs Karat, an interview platform for engineers, called the present arrangement “oddly a buyer’s and a seller’s market at the same time,” and allowed that a candidate does not need to be capable of building frontier models to command a premium — “a good software engineer with the fundamentals to know how to use AI efficiently” will do. That is real money, and the scarcity is real too.
The trouble is what the seventy thousand dollars is buying. The bonuses are not bets on revenue arriving. They are wages paid inside the interval — boom money, spent to accelerate the boom — and the arithmetic they rest on ran through these pages in October: for the current capital program to justify itself, the AI sector needs roughly 9 per cent of U.S. GDP in annual revenue by 2032. Trace the money and the beneficiary is not hard to find. These are one-time payments from a startup accelerating hiring ahead of a revenue test the sector has not begun to clear, and a one-time bonus buys the hire without the permanent cost base a raise would create. That is precisely why the buyer prefers it: the bonus purchases an engineer without purchasing the engineer’s permanence.
“We need software engineers. We need deployment strategists,” Carlos Lezama, Serval’s head of talent, told the paper. “Anthropic and OpenAI are taking a lot of talent.” Lezama named the pressure. Jill Hernstat, who runs a Bay Area recruiting firm, put the operating principle behind it more bluntly still: companies do not want to train people, so candidates who can prove hands-on experience are “golden,” and the preference in practice is to “steal someone else’s person.” That is the whole architecture of this labour market stated by the people running it. The training function, which every engineering organization used to absorb as an ordinary cost of doing business, has been converted into a bounty. The firm that develops the skill pays nothing for it; the firm that arrives late pays a premium to take delivery of it already assembled. The worker who stayed funds the difference, because the difference is the raise she did not get. The scarcity the bonuses are chasing is not natural. It is manufactured, in the plainest way available — by employers who would rather pay thirty thousand dollars once than a raise forever, because the person who takes the raise stays. The workers have priced this correctly: they are stuffing AI terms onto their résumés, because the market pays for the credential and cannot inspect the competence.
It is true that pay structures do lag. Payscale’s chief compensation strategist, Ruth Thomas, is right that adoption is moving faster than companies can adjust, and that the lag creates flight risk. The trouble is that a pay structure is not weather. It is a spreadsheet somebody maintains, on a schedule somebody sets, with bands somebody approves, and every cycle that passes without an adjustment is a decision made by a person with the authority to make it. A claim that cannot be checked is just a claim, and the check here produces a forty-point gap: 56 per cent of employees believe they should be paid more for developing AI skills; 16 per cent report any pay increase tied to AI learning.
The inversion shows from the hiring side too. Levels.fyi, the salary platform that collects offer data from engineers at the big firms and the top startups, has new software engineer offers running 4 per cent above current employees at entry level over the past year, widening to a 17 per cent premium at the most senior levels. The most experienced people inside the building are the most underpaid relative to what the company will pay for them outside it — the incumbent who knows where the bodies are buried in the codebase is invisible from outside, and invisible means unpriced. The bonus-versus-raise arithmetic is the subscription model run in reverse: the buyer of labour wants the one-time charge, the seller of labour wants the recurring one. I watched the inside version of that transaction at the Toronto firm where I spent eight years, the one an American holding company bought and spent eighteen months taking apart — above-market offers for the handful of people the deal needed, a hard ceiling for everyone who stayed, and a great deal of talk about how competitive the market was. The playbook did not need updating. It transferred without modification.
None of this makes the workers who switch villains. Getting paid is the correct response to a market that announces its preferences in dollars, and there is nothing irrational about any single transaction. ADP’s data has job switchers in construction and manufacturing capturing pay-growth premiums of nearly 11 per cent and 5.5 per cent respectively, much of it data-centre and chip work, and the federal gap between job huggers and job switchers has been widening since the summer. Valentino Stoll, a 42-year-old AI software engineer in New Jersey, moved in July from a job paying about $260,000 to a smaller startup, got a raise and a promotion, and deserves both. John Lee, a 41-year-old systems engineer in Denver who has passed through Lockheed Martin and Northrop Grumman on his way up, has hopped from $59,000 in 2008 to $300,000 this year, and reports that “money talks sometimes.” The record shows people responding rationally to the prices in front of them. The interesting question is who set the prices.
To be fair — and here the phrase is doing honest work — the electricians have earned every cent of the seventy dollars. The work is real infrastructure: hot, heavy, skilled, load-bearing for everything else in this story, overtime and double time included. In one corridor of Virginia’s tech alley, data-centre build-outs are paying journeymen electricians $70 an hour against the local’s $62 base rate, with health insurance, overtime and in some cases double time, as Adam Boles of IBEW Local 26, the electricians’ local for the D.C. region, told the Journal. But the story being told about who set that rate is wrong. Seventy dollars an hour is not the market’s verdict on journeymen electricians; the market’s verdict was available all along and sat closer to the commercial rate. Sixty-two dollars an hour exists because Local 26 negotiated it, contract by contract, across decades in which no reporter was following electricians anywhere. The boom is not paying a premium for skill. It is bidding past a floor somebody else built — the same story, read from both sides of the fence, as the 42 per cent premium on data-centre construction arriving alongside the neighbours’ pushback on where the buildings go.
Note what carries each premium. The electrician’s rate is bargained, posted, and owed to every member of the local for the length of the agreement. The engineer’s premium is a signing bonus — paid once, resets at the next hop, and attached to no obligation on either side. One is a wage. The other is a lure, and a lure is only as good as the next offer. My father spent thirty years as a journeyman millwright on a bar mill, and when Gerdau bought the place in 1995 he came home saying the rate for the incoming operator and the rate for the man already on the line had stopped being the same number. The floor had been negotiated in an earlier decade; the boom was bidding over it. First the floor, then the boom. Never the reverse.
The distance between the electrician’s seventy dollars and the software engineer’s stalled pay structure is not talent, and it is not scarcity — both trades are short-handed against the same buildout. It is organization. Brian Merchant’s Luddites did not hate the looms; they objected to who owned them and to what the owning let the owners do to the people who worked them, and the premium structure in front of us makes the same point with a spreadsheet instead of a frame-breaking riot. What the fifty-six per cent who developed AI skills without a raise need is not a bonus or a bootcamp. It is a floor — the journeyman rate for software: negotiated bands, sectoral standards, or, at the unglamorous minimum, internal pay structures published and audited against the market on a schedule, so that loyalty has a price and the price is visible. Publish the band with the job ad and ban the salary-history question outright, so the offer anchors to the role rather than to whatever the last employer got away with. Interoperability for pay is not a metaphor. It is the same principle the trades have applied for a century, and it is the only version of this market that does not require each worker to renegotiate her worth every eighteen months. The payroll data now being collected is, read correctly, a recruitment document. It states, in numbers, exactly what a floor is worth.
Cory Doctorow’s account of what constrains this kind of behaviour — competition, regulation, self-help, labour — is useful precisely because it is a checklist rather than a mood. On the labour line, the constraint is not absent; it is being paid out as individual signing bonuses instead of accruing as shared, durable power. Scarcity leverage evaporates the moment the layoffs arrive. A rate in a contract does not.
Seventy thousand dollars, a hundred roles, one autumn. The company will pay you to arrive, pay you to recommend the person who arrives after you, and pay you ten thousand dollars to take a vacation about it. The people who stayed get a flight-risk label and a compensation cycle. The company has told you exactly what it thinks your loyalty is worth. It is in the recruiting budget, and it is not zero — it is what they pay to avoid paying you.
Everything in this boom is priced by the day: the bonus, the referral, the vacation, this quarter’s wiring rate. The floor was priced across decades, in years when nobody was writing about electricians at all. Bonuses don’t compound. Floors do.