Congress is stealing from AI’s builders to shelter the investors who own the boom.

Here are the numbers, and on the builders’ side they are good numbers. Serval, an AI startup filling one hundred engineering roles this fall, is paying a $30,000 signing bonus, $30,000 to employees whose referred candidates get hired, and a $10,000 vacation stacked on top of the offer — seventy thousand dollars to land one engineer, none of it tied to what the engineer does after day one. Carlos Lezama, Serval’s head of talent, gives the reason in one sentence: “Anthropic and OpenAI are taking a lot of talent.”

The premium is real. It reaches the trades. ADP’s payroll data show workers who switch jobs capturing pay-growth premiums of nearly 11 percent in construction and 5.5 percent in manufacturing, most of it on projects tied to data centers and the chips that power them. On data-center build-outs in one stretch of Virginia’s tech alley corridor, IBEW Local 26 journeyman electricians are paid $70 an hour against the local’s $62 base rate — health insurance, overtime, double time in places — a premium this page measured in September at 42 percent for data-center work, alongside the public pushback the campuses draw. The differential is written into the rate schedule. Nobody has to ask what the skill is worth.

Federal data show the split in aggregate: the gap between the workers who stay and the workers who switch has been widening since the summer, and payrolls added 165,000 jobs in October. The total is not the finding. Where it lands is. John Lee, a systems engineer in Denver, took a new role this year that lifted his pay to $300,000, up from $59,000 when he started out in 2008. The market is telling the truth about these people.

Inside companies, the same skill carries a second price. Levels.fyi, which collects offer data from employees of big tech companies and top-tier startups, puts the median base salary for a new software engineer offer 4 percent above current employees at the same level at entry, widening to a 17 percent premium at the most senior levels. Same level, same skill, two prices — and the gap sits in recurring base salary, not a one-time payment for the trouble of moving. The difference between them is what the incumbent’s skill is worth on the open market. The incumbent does not collect it.

Payscale’s survey measures what happens to the people who stay. Fifty-six percent of employees believe they should be paid more for developing AI skills; sixteen percent report an actual increase tied to it. Ruth Thomas, Payscale’s chief compensation strategist, supplies the mechanism — adoption outrunning pay structures — and names the risk the lag creates: employee flight. That is the industry’s own account of the finding. The worker who leaves is the risk. The skill that was never paid for is not a risk; it is a saving. Eighty-four of every hundred employees report no increase tied to AI learning.

The pay structure is not too slow to move. It moves when the company needs a body and stops when the body is already there. Serval’s offer puts $30,000 on naming a candidate. Employers are stealing the raise their workers’ AI skills are worth. Jill Hernstat, who runs a Bay Area recruiting firm, hears it from the other side: companies don’t want to train anyone, candidates who can prove hands-on experience with AI tools are “golden,” and the posture is “we’d rather steal someone else’s person.” That is the hiring side’s own word for what it does.

Now put the capital side of this market on the same ledger. The servers, networking gear and qualifying equipment going into these buildings are written off in full the year they are installed — Section 168(k) of the tax code, the 100 percent first-year expensing provision, made permanent by the reconciliation act of July 2025. Under the definition in the Congressional Budget and Impoundment Control Act of 1974, a revenue loss from a special deduction, credit or preferential rate is a tax expenditure: spending, run through the tax code. The capital cost of the AI buildout is booked the week the switchgear arrives, and the revenue it has to earn is on the record too — analysis published here in October puts the revenue this sector would need by 2032 at 9 percent of U.S. GDP. The public is financing the capital that is bidding up these wages, and the wage is on no ledger at all.

That is the code’s answer to the capital. Here is its answer to the work. Every dollar of the premium above arrives as wages, and the federal schedule for wages runs like this: ordinary rates up to 37 percent on the last dollar earned; 7.65 cents of payroll tax on every dollar — 6.2 percent for Social Security until the annual wage base cuts the tax off, 1.45 percent for Medicare with no ceiling at all — plus a 0.9 percent Medicare surtax above $200,000 of wages for a single filer, and the employer’s matching half priced into the wage before it is ever paid. That is the schedule the code applies to work.

Begin with the owners’ first instrument. The managers of the investment funds financing the boom take their pay as a share of the funds’ profits, and the code lets them book that share — their carried interest — as a long-term capital gain once a three-year holding period clears. The schedule for that dollar is a 20 percent top rate plus a 3.8 percent surtax on high earners’ investment income, and no payroll tax whatsoever. Set the two schedules side by side: the last dollar of wages above the Social Security wage base meets the ordinary top rate, the uncapped Medicare tax and the surtax; the carried dollar meets 20 percent plus 3.8 percent, and nothing else. It is compensation by any honest name. The schedule treats it as ownership income.

The public-market side compounds under the convention called realization: the code taxes a gain when the owner sells, and not one day before. An unsold position compounds untaxed. Cash goes to shareholders through repurchases rather than dividends, because a repurchase pays in appreciation and a dividend pays in taxable income; the 2022 law’s entire gesture at this channel was a 1 percent excise on repurchases, Section 4501, which the Joint Committee on Taxation scored at $74 billion over ten years. One percent on the shareholder channel. Thirty-seven and the stack on wages.

For founders and early holders the sequence extends. Never sell. Borrow against the appreciated stock, because a loan is not income. At death, the asset’s tax basis resets to market value — the step-up — and the accumulated gain leaves the income tax permanently; the estate tax, assessed on that same value, carries an exemption of roughly $14 million per person, indexed for inflation, and almost never collects. Law professors call the sequence buy, borrow, die. The code has a word for each step of it: legal.

Name the ledger properly. Anyone who has read the Joint Committee’s annual tax-expenditure tables knows the shape of the list: roughly 165 provisions, about $1.8 trillion a year in aggregate, with the preference for capital gains and qualified dividends sitting among the two largest items, beside the exclusion for employer-provided health insurance. The builders’ premiums are booked as income. The owners’ preference is booked as expenditure. Same boom.

The asymmetry reaches even the corporate tax the boom’s companies pay. Incidence — who actually bears a tax once prices and behavior adjust — is a convention choice, and the conventions disagree. The Joint Committee books 75 percent of the corporate tax on capital and 25 percent on labor; Treasury’s Office of Tax Analysis, 82 and 18; the Tax Foundation, 30 and 70 — an allocation the empirical literature does not support for an economy of this size, where the measured range for labor’s share runs roughly 20 to 35 percent. The distributional tables I helped build at CBO ran on these conventions, and as conventions they are defensible; I name them anyway, because a quarter of every corporate dollar in the official tables is assigned to labor — borne, in practice, out of the same wages the schedule above is already taxing at the door. The builder is taxed on arrival and again by assignment.

Now watch the before-and-after, because it is the cleanest symmetry test in the code. The 2017 Congress moved the holding period on carried interest — Section 1061 of the Internal Revenue Code — to three years and left the preference itself standing. The 2022 climate-and-tax bill was drafted with a full repeal; the Joint Committee scored it at about fourteen billion dollars over ten years; the provision was dropped from the final bill. Two congresses, two directions, one outcome: the preference stands. And the stakes scale with the boom: revenue at 9 percent of GDP, if it arrives, compounds under the same preference, and the preference’s cost compounds with it.

The repair is base broadening, and it is boring, which is the point: tax the fund managers’ profit shares as the compensation they are, tax gains at death on the schedule wages meet every paycheck, and make the preferential rate defend itself in the open — the expenditure tables already itemize it every year; what is missing is that the itemization carries a vote, the way an appropriation does. The write-off is spending, and spending can carry a condition: any firm claiming it on data-center property can be required to publish what it pays an incoming engineer against what it pays an incumbent at the same level. The price is already written on their own offer letters. It is simply not recorded where it would do any work. The 1986 reform ran on exactly that discipline — broaden the base, then cut the rates — and it was scored honestly. It remains the standard the current code has abandoned.

The market has done its part: $30,000 to sign at Serval, $70 an hour on the switchgear, a 17 percent premium at the senior desk. The Congress wrote a full repeal into its own bill, and then took it out. That is not an oversight. It is a deliberate choice.