Lynceus

Case study · Public records

The highest H-1B salary in the government's data is $586,919,840.
An assistant professor.

It is, of course, wrong. Getting to the real number took three passes through 437,496 Department of Labor filings, and the fix every data textbook recommends was the one that failed. The cleaned file and the scripts are at the bottom of this page.

By · Data through 2026-06-30 · Last reviewed: September 2026

The question

What does H-1B practice look like? Public records only.

A client wanted to know how H-1B work is actually distributed across immigration attorneys and firms, how concentrated any one practice tends to be, and how a given practice compares with the market. No interviews, no firm websites, nothing self-reported. Only what the government publishes.

The first thing the records tell you is which agency to ignore. USCIS decides every petition and publishes nothing with an attorney's name on it. The Department of Labor certifies the Labor Condition Application that precedes every petition, and its quarterly disclosure files carry the attorney's name, firm, email, city and bar state on every row. For FY2026 through June that is 437,496 cases, 7,694 distinct attorneys, 4,189 firms, the largest of which filed 60,804 cases in nine months.

Names are free text. One practice we followed through the files appeared under three spellings. The email address was identical on all of them, so we keyed identity on email rather than name, and the count rose by a tenth. To benchmark a practice against the market we needed the real wage distribution. That is where the $586 million came from.

Cleaning the wage column

Four passes. The obvious one fails.

01.

max(). $586,919,840.

An assistant professor. The employer typed an annual salary and left the pay unit on "Hour": 282,173 dollars an hour, times 2,080. The next four rows were the same error. Anyone who reports the maximum of this column is publishing a typo.

02.

Ratio to prevailing wage. Textbook. Fails.

DOL sets a prevailing wage for every filing and the offer must meet it, so an offer at 2,000 times prevailing is obviously wrong. Except that when an employer mislabels the unit they mislabel both fields identically. A $250,000-an-hour physician sat against a $312,927,680 prevailing wage and passed at 1.7x. The ratio test cannot see a unit error because the error cancels out.

03.

Absolute ceilings per unit. Better. Still wrong at the top.

Nobody earns more than $500 an hour on an LCA, so a higher hourly figure is an annual salary wearing the wrong label. That repaired 381 rows. The new maximum was a "Vice President" at $20,800,000 against a prevailing wage of $18,865,800. A clean times-100 error on both fields. Ratio 1.1x. Waved through again.

04.

The prevailing wage has a ceiling. Measure it.

The prevailing wage is not employer-typed free text. It comes from a government occupational survey, so it has a real upper bound. We measured it across all 437,496 rows: the 99.99th percentile is $431,000, and only 16 rows exceed $500,000. Every one of the 16 is corrupt. That ceiling, plus the unit ceilings, plus a 12x ratio test, rejected 51 rows out of 326,385 and left a defensible answer.

Verified distribution, 326,334 attorney-represented cases
p50
$130,728
p75
$170,068
p90
$217,000
p99
$360,000
p99.9
$650,000
max
$2,480,000

The real maximum is a law firm partner. The raw figure was off by 236x. 29 cases exceed $1,000,000; 897 exceed $500,000. Median: $130,728.

What the clean table shows

The visa nobody talks about is winning the top of it.

E-3 share of all filings
1.8%
7,671 of 437,496
E-3 share of the 50 highest wages
32%
16 of 50, including the single highest

An 18x overrepresentation, and it is structural. E-3 is open only to Australian nationals, has its own quota that never fills, and no lottery. A firm moving a senior person from Sydney to New York files and it happens. Moving an equivalent person from anywhere else means a March lottery with roughly a one-in-four chance. The very top of the "H-1B" wage table is partly selected by which visa the highest earners can actually get.

That is not a wage story. It is a queue story wearing a wage story's clothes.

And AI?

Seven of the top 50. But one lab has five.

Top 50 credible wages, by occupation
Chief Executives 14
Software Developers 7
Lawyers 6
Computer and Information Research Scientists 4
Cardiologists 4

Only seven of the fifty carry any AI signal, by job title or by employer. The top of this distribution is partners, chief executives and surgeons, people whose pay is profit share or practice income rather than a wage.

The "but": one AI lab is the single most represented employer in the top 50, with five entries, more than any bank, law firm or big-tech company: "Member of technical staff" at $1,380,000, $1,380,000, $1,120,000, $1,120,000 and $1,120,000. Across all 253 of that lab's filings the median is $188,000 against a national median of $130,728. AI does not own the top 50. It owns the top of the salaried tier.

Back to the practice

Seven years of one practice, in three numbers.

FY2019, full year
~190
filings, most of them for one client
FY2023, the trough
~10
filings in the comparable quarter
FY2026, through June
~80
filings, over a dozen clients, largest under a quarter

The single client that carried the practice is absent from every year after FY2022. Volume halved, then rebuilt on a wider base with a median offered wage near $108,000, roughly the 25th percentile nationally, which fits an IT-staffing book rather than a big-tech one. None of that is on a website. All of it is in a 240-megabyte spreadsheet the government posts every quarter.

The part that took longest

Public is not the same as available.

  • —The Department of Labor blocks curl, wget and Python outright. Every request returns 403 regardless of headers. The files download only from a real browser.
  • —The FY2026 spreadsheet is 1,032,735 rows. 595,239 of them are blank padding. Real cases: 437,496.
  • —FY2019 has 260 columns and no attorney name field. FY2020 onward has 98 columns and a different header for the firm name. Three schema conventions in eight files.
  • —The FY2020 through FY2025 "Q4" files cover July through September only, despite a page stating every dataset is cumulative. Summed naively they undercount by roughly 4x. The FY2026 file is cumulative.
  • —USCIS publishes nothing with an attorney's name on it. Its bulk CSV exports stop at FY2023. Newer years exist only inside a dashboard whose export returns a fixed top-100 view and ignores every filter.

We spent more time discovering the format than analyzing the content. That ratio is normal for government data, and it is why the scripts below exist: next quarter's file will have a new surprise in it.

Check it yourself

The cleaned file and the scripts.

All 437,496 cases with offered wage, prevailing wage, wage level and visa class, gzipped CSV, 14 MB. The attorney and law-firm columns are removed. Plus the four Python scripts that produced every number on this page. We ask for a working email and send a six-digit code; the link is good for 24 hours.

No disposable addresses. We add you to the Lynceus list; one email a month at most, unsubscribe in one click.