Week 4 · Communities & backbones
Who hires America's foreign workers?
Where the hiring is, which occupations travel together, which firms staff the seats, what changes without the biggest firms, and what the filings show beyond the networks, all from the same Department of Labor disclosures. Years are US The US government's year runs October to September: 2025 runs from October 2024 to September 2025..
A shared employer is not a shared labour market.
On the map: two Louvain groups
- New York–Dallastwelve large hubs
- Houston–Philadelphiathe other 28
- Linkskept by the disparity filter at \htmlClass{x-alpha-disparity}{\alpha} = 0.2
537,796certified H-1B filings, 2025
40metro areas with the most filings, 84.5% of the year’s total
Five sections, five findings
Where the hiring is
Cities group by who hires there, not by region, and no single link holds the map together.
Who staffs whom
One certified H-1B filing in five names a client company as the worksite. Yet a client that changes An outsourcing firm that files the H-1B and places the worker at another company, its client. Elsewhere the page also calls it a placing firm. stays inside its group far more often than chance.
Without the biggest firms
Take out the largest filers, Amazon above all, and the metro groups start to follow Census regions; the job clusters shift but hold.
Beyond the three networks
The section 3 groups barely show in lawyers or green cards; the outsourcing firms stand out in the wage levels they file.
Loading place data…
Opening
An H-1B filing is an employer’s request to hire a non-US worker in a A job that needs at least a bachelor’s degree in a specific field, such as engineering or accounting..
It names the job, the worksite, the wage and the company asking to employ the worker. The Department of Labor records the filing even when the worker never arrives, changes employer or is not hired.
Why start with companies? The employer is the thread linking the three networks in this story.
Unless stated otherwise, the figures below use The Department of Labor accepted the filing. Only then can the employer take it forward. H-1B filings in 2025. One filing is a request, not a guaranteed job.
What one filing names
Each section builds its network from one of these fields.
- Employer
- The company asking to hire Links all three networks
- Worksite
- A city, grouped into its metro area 1 · Where
- Occupation
- The job, as an official occupation code 2 · Jobs
- Client
- The company the worker is placed at, when it is not the employer 3 · Staffing
- Wage level
- I (entry) to IV (fully competent) 5 · Beyond
- Law firm
- Who prepared the filing 5 · Beyond
How to read the charts
Background
A job title tells us what work is requested; a company tells us which jobs, places, and clients are connected by the same hiring system. The links mean shared filings. They say nothing about friendships between workers or companies.
The filings leave out workers without H-1B sponsorship, employers that never file, rejected or withdrawn applications, and the wider conditions that shape who gets hired.
Where the hiring is
Cities group by who hires there, and no single link holds the map together.
The cities are the 40 metro areas with the most filings, 84.5% of the year's total. A method that finds groups by moving each node into the neighbouring group that raises modularity most, then merging the groups and repeating until no move helps. It starts from a random order, so two runs can differ. splits them into two groups, and the split is weak but real: How much more of the link weight falls inside the groups than a random network with the same number of links per node would put there. Higher means sharper groups. 0.057 against 0.013 for A random copy of the network in which every node keeps its number of partners, but the partners are dealt out again at random. (How many standard deviations the real value sits from the random baseline’s mean. Beyond about 2 either way is rare by chance. = 36).
How two metros get linkedAn example, not data. A company that files in both metros adds the smaller of its two filing counts to their link.
The opening questions, side by side, before 1A
Which cities hire the most?
San Jose asks for the most positions; New York has the most employers.
Toggle the metric; click a bar or a bubble to inspect one city.
Top cities
Ranked by the active metric. Click a bar to select that city everywhere on this page.
Is it one national job market or several regional ones?
Not regional markets: the hub group crosses regions.
Background
A position is a seat an employer asks to fill, not a worker who arrived.
Louvain separates the twelve largest hubs, which cross regions, from the other 28 metros.
The map shows the partition with the highest modularity Louvain reaches, and every number below is computed on it. The null rewires the company × metro network so each company and each metro keeps its number of partners, deals the filing counts back out at random, and projects it again:
One group holds New York, Dallas, San Jose, San Francisco, Seattle, Atlanta, Chicago, Austin, Boston, Washington, Los Angeles and San Diego. The other 28 metros form the second. Modularity is 0.057 against 0.013 for rewired networks that keep each company's number of metros (\htmlClass{x-z}{z} = 36).
Louvain finds the split shown in 35 of 100 runs; the other 65 find one other split, into three groups, which cuts the hubs into New York–Dallas and San Jose–San Francisco at a lower modularity, 0.049. The page shows the higher one; the maps below show both, and more. 2024 gives the split shown in all 100 runs; across the two years the runs agree at Normalized mutual information: how alike two groupings are, from 0 for unrelated to 1 for the same. 0.64, against 1.00 between two 2025 runs.
NMI with Census regions is 0.04 and with divisions 0.11. A method that finds groups by following a random walk: a group is an area the walker tends to stay inside for a while., which follows a random walk between metros instead of counting links, finds no split at all: one module holds all 40.
Method
A filing counts once in each metro it names, with at most the positions it requests.
More numbers
Census regions and divisions match it no better than shuffled labels (The share of shuffled labellings that match at least as well as the real one. A small value means the match is unlikely to be chance. = 0.46 and 0.32).
In Seattle one company, Amazon, files 31%. Positions reward a few firms asking for many seats; employer counts reward a broad market.
San Jose averages 2.6 positions per filing, New York 1.5. New York's largest filer, EY, has under 4%.
Maps: groups and Census regions
Colour the same map by Louvain's groups, two (the split above) to ten, or by Census region.
On the map
The 48 contiguous states; none of the 40 metros lies outside them. Bubbles are sized by requested positions; colour and opacity follow the active metric. Each metro sits at its first-named city. Click a bubble to select it.
Same cities, two labelings
The same map, coloured by the active labelling. Communities are named after their two largest metros. Click a city to select it.
Do cities group by who hires there instead of by region?
By who hires.
This card uses Louvain's three groups (3 on the map's group control): they follow how much of a city's hiring runs through consulting and IT-services firms.
How well each labelling matches Louvain's three groups
Each bar is the Adjusted mutual information: how alike two groupings are, corrected for chance. 0 means no better than labels dealt at random, 1 means the same grouping. between Louvain's three groups and one labelling. Orange bars describe who hires, grey bars where the city is.
Method
We gave each metro four labels: its Census region, its Census division, the third it falls in by the share of its filings that place a worker at a client, and the third it falls in by the share filed by professional and technical services firms (NAICS 54, the sector of IT consultancies). Thirds, because 35 of the 40 metros have that sector as their largest, so "largest sector" says almost nothing. AMI corrects for the number of labels, so four regions and three thirds compare fairly.
AMI for each labelling of the 40 metros; the p-value is the share of 1,000 shuffles of that labelling that match at least as well.
More numbers
The match is modest: most of what makes two metros alike stays unexplained.
Seven of the eight tech-hub metros sit in the lowest third by placed share: there, companies mostly hire for themselves. In the New York–Dallas group the median metro places 27% of its filings at a client and files 60% through IT-services firms.
Census regions reach 0.06 (\htmlClass{x-p-empirical}{p} = 0.11) and divisions 0.06 (\htmlClass{x-p-empirical}{p} = 0.08), no better than chance.
Where does the backbone break, and whose links hold it?
Nowhere in one place: metros drop off one or two at a time.
We removed the links one by one, least significant first by The disparity filter's threshold. The filter keeps a link only when its weight is larger than chance would give, judged against each metro's own total, and α is the chance it allows. A lower α keeps fewer links. The links that survive form the backbone., and watched the largest piece after each removal.
Metros in the largest piece as the filter tightens
Read right to left: as \htmlClass{x-alpha-disparity}{\alpha}falls, links go and metros drop out of the largest connected piece.
Background
Every pair of the 40 metros shares some employer, so the full network is one hairball of 780 links.
Method
The disparity filter keeps a link when it carries an unusually large share of either metro's total weight ; \htmlClass{x-alpha-disparity}{\alpha} is the test's threshold, and a smaller \htmlClass{x-alpha-disparity}{\alpha} keeps fewer links.
Each step is one or more links removed. The shaded band is the range from \htmlClass{x-alpha-disparity}{\alpha} = 0.1 to 0.05 where the five-value sweep saw the drop.
More numbers
The big outsourcing firms hold no more of the links that cut metros loose than of any others.
The other eleven are led by Amazon (three), Intel, Deloitte, Capital One, JPMorgan Chase, Citigroup, FedEx, Fidelity Investments and the University of Maryland.
The first metro falls off at \htmlClass{x-alpha-disparity}{\alpha} = 0.136. Tried at five values, the largest piece of the map fell from 32 metros at \htmlClass{x-alpha-disparity}{\alpha} = 0.1 to 18 at \htmlClass{x-alpha-disparity}{\alpha} = 0.05, which looks like one snap. The fall from 32 to 18 is 14 separate links, each peeling one metro away.
The five placing firms lead 69 of the 180 links in the whole backbone at \htmlClass{x-alpha-disparity}{\alpha} = 0.2 (38%), so their 5 of the 16 links that cut a metro loose is no more than their share (\htmlClass{x-p-empirical}{p} = 0.37).
No single removal cuts off more than two metros: Dallas–Durham at \htmlClass{x-alpha-disparity}{\alpha} = 0.041 takes Durham and Raleigh with it, and New York–Seattle at 0.023 splits the last five metros three and two.
Table: 14 links that peel metros off
| Link | α | Weight | Leading company | Its share |
|---|
Background
Louvain splits the 40 metros into twelve large hubs led by New York and Dallas, and the other 28 from Houston and Philadelphia down. The rewired networks keep each company's number of metros. The two questions below ask what the groups follow and where the network comes apart.
Method
Two metros are linked when the same company files in both; the link weighs, summed over those companies, the smaller of the company's two filing counts.
Which jobs go together
Employers reveal bundles of work.
Two occupations are linked when the same companies file for both. The clusters across all … occupations are real: A score for how much more a network links inside its clusters than chance would. Higher means cleaner clusters. … against … for Random copies of the network in which every company keeps its number of occupations. They show how much clustering chance alone produces..
Background
Software developers sit in almost every company's mix, so most links run through them. The two questions below test the clusters from two sides: do outsourcing firms and direct employers bundle jobs the same way, and does any job belong to two bundles at once?
Loading job data…
Which jobs are hired together?
Software Developers sit in 9 of the 12 pairs because almost every sponsoring company hires them.
Each bar is one pair of jobs that the same companies hire for. Its length counts the companies that filed for both; the dark bars are pairs with Software Developers.
The 12 most common job pairs
Do outsourcing firms bundle jobs differently from direct employers?
Yes. Their job clusters differ more than those of random groups matched on size.
Each group gets its own occupation network and its own A standard method that splits a network into groups whose members link more to each other than to the rest. clusters. Normalized mutual information: a score for how alike two groupings are, 1 when they match exactly and 0 when they are unrelated. says how alike the two clusterings are.
Do outsourcers cluster jobs like random firms would?
Bars show how alike the two groups' job clusters are. Whiskers span one How far values typically sit from their average..
The largest occupations in each group
Share of each group's filings, for the eight occupations with the most filings overall.
Method
We split the companies in two: the 817 firms that place 20 or more filings at client sites (21% of all filings) and the 58,379 others. Alone, that number means little: splitting companies into a small and a large group changes the clusters even if nobody behaves differently. So the baseline draws 20 random groups that match the outsourcing firms in both respects: the same number of companies of each size, from the one-filing firms to the giants. NMI is measured on the 211 occupations that sit in a cluster of two or more on both sides.
The random groups have the same number of companies and the same share of filings as the outsourcing firms. Whiskers span one standard deviation over 20 random splits. Only the blue baseline matches the outsourcing firms on both number and size of companies, so it is the fair comparison.
More numbers
The half-matched baselines show why the match matters: random groups with only the same number of companies hold 1.3% of filings and agree at 0.42 ± 0.10, which would have hidden the difference. At the top the two mixes look alike: software developers are 28% of the outsourcing firms' filings and 33% of the direct employers'. Below that they part: "computer occupations, all other" is 22% of the outsourcing firms' filings and 5% of the direct employers', and direct employers file for 259 occupations the outsourcing firms never touch.
Does any job belong to two clusters at once?
Not clearly. On a network this dense, link communities pour almost everything into one cluster.
Link communities (Ahn, Bagrow and Lehmann, 2010) group the links instead, so a job belongs to every community its links are in.
Background
A Louvain partition puts every job in exactly one cluster.
Method
The cut is chosen where partition density D, the average of how close each community is to a complete one, peaks. On the 25,431 links between 471 occupations it peaks at D = 0.58 with one community holding 84% of the links; 114 communities have three links or more, counting it.
More numbers
The top is small occupations such as communications equipment operators and electrical power-line installers (5 communities over 11 links each). None of the 3 occupations that section 2's first test flagged as bridges appear in it.
A job's number of communities mostly counts its links (Spearman 0.85), so the table ranks by communities per link, as the course suggests.
Table: 15 jobs in the most communities
| Occupation | Links | Communities | Per link |
|---|
Who staffs whom
One certified H-1B filing in five names a client company as the worksite. Against a random baseline clients group only weakly, yet a client that changes vendor stays inside its group far more often than chance.
Counted once per link, the A method that finds groups in a network by moving nodes between groups until the links inside groups are as dense as they can get. It starts from a random order, so two runs can differ. groups beat Random copies of the network in which every firm and client keeps its number of partners, but the partners are dealt out again at random. on How much more of the link weight falls inside the groups than a random network would put there. Higher means cleaner groups. (0.57 against 0.53) and match each client's main vendor (Adjusted mutual information: how closely two ways of grouping the same clients agree, corrected for the agreement random labels would reach by chance. 0.11) a little better than its industry (0.07).
Words in this section
- Vendor
- An outsourcing firm that files the H-1B and places the worker at another company, its client. Elsewhere the page also calls it a placing firm.
- Main vendor
- The vendor that files most of a client's placements in a year.
- Switch
- A year in which a client's main vendor changes, counted for clients with five or more filings in both years.
Background
Here the network links an outsourcing firm to each client company where it places workers, and a link weighs the filings between them.
Method
Louvain runs on the largest connected piece: 21,759 firms and clients, 41,212 links. The rewired networks keep every firm's and client's number of partners. The three questions below ask whether the groups behave like markets.
How many workers sit at a client?
In 2025, 104,732 of the 537,796 certified filings (19.5%) mark a client site.
A filing is a request to employ someone, not a hire. Every year from 2022 on, USCIS denied placing firms about twice the share of first-time petitions it denied direct employers.
Method
The worksites file lists every client a filing names; we leave out the 16% of client entries that name no company, such as "Home Address", and the 1,881 where a firm names itself. Counted that way, 101,763 filings (18.9%) name a client company: the worker is employed by one company and works at another, down from 21.9% in 2022.
More numbers
USCIS denied 2.7% of placing firms' first-time petitions against 1.2% for direct employers in 2022, and 3.4% against 2.0% from October 2025 to June 2026. The lottery shows the same split one step earlier. Each new H-1B worker starts as a registration that USCIS draws at random, and USCIS gave Bloomberg News every registration from the March 2023 draw after a FOIA lawsuit. Every petition that followed names its filing, so we can follow a ticket to its client. Direct employers sent 5.1 registrations per approved petition, placing firms 9.1, and firms with fewer than 20 filings 12.2; those small firms sent 53% of the 758,967 registrations. Most of the gap is drawn tickets nobody used. When USCIS drew a direct employer's registration, a petition followed 76% of the time; a placing firm's, 50%; a small firm's, 35%. That step carries 74% of the gap between placing and direct firms, and the draw itself 24%. Much of it comes from workers registered by several employers: 54% of registrations named one, and when USCIS drew one, a petition followed 23% of the time, against 81% for a worker registered once. 18,307 of the petitions lead to a client company. Citigroup received the most, 342 through 38 firms.
When a client changes its main vendor, does it stay in its group?
Yes, nearly eight times as often as for a random vendor.
A client's main vendor is the firm that files most placements there in a year.
Switches that stay inside the client's group
Share of switches whose new main vendor is in the client's group of the earlier year, for each pair of years. Grey bars are the baseline; whiskers span one standard deviation.
Background
A switch is a year in which that firm changes.
Method
We followed the clients with five or more filings in two consecutive years and found 1,261 switches of main vendor between 2022 and 2025. For each, we asked whether the new vendor sat in the client's Louvain group of the earlier year.
The baseline draws a new vendor at random, a large firm as often as its filings make it likely. The chart's grey bars are the baseline's mean over 1,000 draws.
More numbers
One switch in four stays inside the client's group. Much of that is familiarity: 62% of new main vendors already placed someone at the client the year before.
A stricter baseline that draws only among the firms already at the client narrows the gap to 26.8% against 20.8% ± 0.7% (\htmlClass{x-z}{z} = 9), a lift of 1.29 rather than 8.3.
Which clients change group when filing counts are ignored?
Two in three, but much of that is Louvain's own noise.
We ran Louvain with links weighted by filings and with every link counting one, and called a client a mover when its group changed.
Share of clients that change group
The first two bars compare two seeds of the same kind; the third compares the weighted partition with the unweighted one.
Method
Two runs of the same kind with different The random starting point of a Louvain run. Different seeds can give different groups. set the noise floor.
We matched each weighted group to the unweighted group it overlaps most, and called a client a mover when its matched group changed. All three bars use the same matching of groups.
More numbers
The medians come from ten pairs of seeds each (ranges 26.8% to 39.3% and 49.4% to 55.9%). We expected the movers to be clients with several vendors, since only their filing counts can pull them one way or another. They are, but barely: 31.8% of movers have two or more vendors, against 26.8% of all clients. The largest movers are the largest clients: Citigroup sits with Tata Consultancy Services when filings count and with EY when they do not; Bank of America moves from Infosys' group to IBM's.
Table: 15 largest movers
| Client | Filings | Vendors | Group, weighted | Group, unweighted |
|---|
A group is named after its largest firm.
Which clients sit in two groups at once?
1,823 clients get a fifth or more of their filings from a second group, fewer than in rewired networks.
A Louvain partition gives each client one group, but a client with several vendors can draw on firms from different groups.
Clients split between two groups
The count for the real network against the mean of 100 rewired networks; the whisker spans one standard deviation.
Background
Method
We counted clients whose second group supplies at least 20% of their filings, and did the same on 100 rewired networks in which every firm keeps its number of clients and every client keeps its filing counts, only attached to different firms. Louvain builds the groups from these filing counts, so it already tends to put a client with its heaviest vendors; and rewired networks split into different groups from the real one. Read the count as the groups following clients' main suppliers, not as a separate measure of loyalty.
More numbers
The largest split clients are banks, insurers and manufacturers. USAA gets 44% of its filings from HCL's group and 26% from Tata Consultancy Services'; Stellantis 46% from L&T Technology Services' group and 34% from Tata Consultancy Services'.
Table: 15 largest split clients
| Client | Filings | First group | Second group | Main vendor |
|---|
Without the biggest firms
Take out the largest filers, Amazon above all, and the metro groups start to follow Census regions; the job clusters shift but hold.
The ten largest filers file 19.1% of the filings in the 40 metros, and the five largest Companies that hire foreign workers and send them to work at client companies, such as outsourcing and IT staffing firms. 5.6%.
How much do the groups change?
NMI between the groups after a removal and the full network's groups: 1 means unchanged. Metros on the left, jobs on the right.
Do the metro groups follow Census regions?
AMI between the metro groups and the four Census regions. Orange bars remove named firms, grey bars random firms with the same share of filings (mean of 50, whisker one standard deviation). The dashed line is the full network.
Which firm hides the regions?
Amazon.
We removed each of the ten largest filers alone, then the top 1, 2, 3 … 20 filers in turn.
The largest filers removed in turn
Orange: the metro groups’ match with Census regions as the largest filers leave. Grey band: random cuts of the same size. Hover a point for the firm.
One firm out at a time
Orange: the match with one firm removed. Grey: random cuts of the same size. Dashed line: the full network.
Method
Each removal sits beside random cuts of companies that remove the same share of filings.
50 random cuts for a single firm, 20 for each step of the sweep.
More numbers
Amazon's 0.14 sits 3.5 standard deviations above its random cuts. No other single firm pushes the match up beyond its random cuts: removing EY, Meta, Deloitte or Apple alone tips Louvain into a two-group split that ignores regions (AMI −0.005). Removed in rank order, the largest filers keep the match above random cuts at every step from one to twenty, but not smoothly: it dips to about 0.07 without the top 17 to 19, where several partitions compete, and peaks at 0.20 without the top 20. 2024 tells the same story more strongly. Its full network shows no regional match (AMI −0.005); without its ten largest filers the match is 0.22 (\htmlClass{x-p-empirical}{p} = 0.001), against −0.01 ± 0.01 for random cuts.
Background
A company that files everywhere links every pair of metros and jobs, so its footprint could be all the structure there is.
The five largest placing firms are Tata Consultancy Services, Cognizant, Infosys, HCL and Compunnel. The ten largest filers are Amazon, Cognizant, Google, Microsoft, EY, Meta, Deloitte, Apple, Tata Consultancy Services and Infosys.
Method
We removed each set, reran 100 A method that finds groups in a network by moving nodes between groups until the links inside groups are as dense as they can get. runs on the metro network and on the job network (weighted here by filings, since a count of companies barely moves when ten of 59,196 leave), and compared the groups with the full network's. Removing less data changes the groups too, so each removal sits beside 50 random cuts of companies that remove the same share of filings.
More numbers
The job clusters hold (Normalised mutual information: how much two groupings of the same items agree, from 0 (unrelated) to 1 (identical). 0.90 and 0.81) but shift more than random cuts of the same volume do.
Removing the five placing firms changes little: the groups stay close to the full network's (NMI 0.92, random cuts 0.85 ± 0.14), and the regional match rises only to 0.09, inside the range of random cuts (0.04 ± 0.05). Without the ten largest filers the clusters also sharpen, modularity rising from 0.27 to 0.30. Every version still beats its own rewired networks by a wide margin (\htmlClass{x-z}{z} = 25 or more). Without the ten largest filers the regional match (AMI 0.13, \htmlClass{x-p-empirical}{p} = 0.013) sits 4.6 standard deviations above random cuts (0.01 ± 0.03). Random cuts leave the job clusters closer to the full network's (0.96 and 0.88, 2.9 and 4.3 standard deviations away), so the biggest firms do shape which jobs cluster together.
Beyond the three networks
The section 3 groups barely show in lawyers or green cards; the outsourcing firms stand out in the wage levels they file.
The same filings name the law firm that prepared them, and the Department of Labor's green-card files name the employers that sponsor permanent residence. Three short questions follow the section 3 groups into those records.
- 5ALaw firms: barely follow the section 3 groups.
- 5BGreen cards: outsourcing firms sponsor fewer per H-1B filing, though the intervals overlap.
- 5CWage levels: a placed filing has 3.6 times the odds of level I or II.
Do immigration law firms split companies the way vendors do?
Barely. Law firms group the companies along the section 3 lines only slightly more than chance.
Agreement with the section 3 groups
AMI between the law-firm groups and the staffing groups, against 50 rewired law-firm networks in which every company and law firm keeps its number of partners.
Method
75.6% of certified filings name a law firm; after name cleaning there are 4,479. We linked each firm that places workers in section 3's network to its law firms, weighted by filings, ran Louvain, and compared the groups with the section 3 groups for the 4,576 firms in both. Most companies use one law firm: 64% of filings come from such companies, which makes the network a set of stars.
More numbers
The law-firm network scores 0.90, below the 0.93 of its rewired copies, because stars split into near-perfect groups whatever the wiring. Fragomen files the most (14,087 filings for 131 of these companies); EY Law files 10,274 for eight. The network of law firms themselves, across every employer, is in Who files the paperwork? in the deep dive.
Do outsourcing firms sponsor fewer green cards?
Fewer per H-1B filing, but the staffing groups do not differ beyond chance.
Here the outsourcing firms are the 817 companies that place 20 or more filings at client sites.
Green-card filings per H-1B filing
The six on the right are the largest section 3 groups, named after their largest firm.
Background
A green card through work starts with a The Department of Labor's permanent labor certification: the employer shows that no qualified US worker is available for the job. Most green cards through work start there. filing.
Method
For each company with 20 or more H-1B filings we divided its 2025 PERM filings by its H-1B filings, matching companies by name and tax number.
Pooled over the companies in each bar, with a 95% A range for a result found by recomputing it on many resamples of the data, drawn at random with repeats; 95% of the resampled results fall inside it. for the first two.
More numbers
The gap shrinks to 0.13 against 0.17 when a firm counts as outsourcing only if most of its filings go to clients. Single companies swing these rates more than any group does: counting every case status, filings in the names of Amazon and Google fell from 3,638 and 1,618 in 2024 to 15 and 3 in 2025. Green cards per company and the strongest employer ties are in the deep dive.
Outsourcing firms file 0.11 green cards per H-1B filing (95% interval 0.08 to 0.15), direct employers 0.18 (0.14 to 0.21). Across the six largest staffing groups the rate runs from 0.03 in Cognizant's group, where Cognizant itself filed almost none, to 0.15, a spread that We dealt the companies out to the groups again at random, many times, to see how big a spread chance alone produces. match 28% of the time (\htmlClass{x-p-empirical}{p} = 0.28).
Do outsourcing firms file at lower wage levels for the same job?
Yes. For the same occupation, a placed filing has 3.6 times the odds of wage level I or II.
Every filing states a A level set by the experience and skills the job asks for. Each level carries a wage floor. from I (entry) to IV (fully competent).
Share of filings at level I or II
The five occupations with the most filings. Orange bars are filings that place the worker at a client, blue bars filings for the employer's own site.
Background
92% of filings give one.
A level describes the job as filed, not the worker, so this shows cheaper job descriptions, not lower pay for the same person.
Method
We compared placed and direct filings within each occupation, so software developers are compared with software developers.
More numbers
Overall, 79% of A placed filing sends the worker to a client company's site; a direct filing is for the employer's own site. filings sit at level I or II against 58% of direct ones.
A few large firms file most placements, so the interval comes from resampling whole employers: 2.87 to 5.04. It is not one firm's doing: without the 5, 10 or 20 largest placing firms the ratio rises to 3.85, 4.28 and 4.94. Software developers: 87% against 52%. Offered wages follow: placed filings offer a median 1.00 times the prevailing wage in the five largest occupations, direct ones 1.00 to 1.08 times.
Closing
The groups in these networks are weak, but they are not noise.
What surprised us
The staffing groups looked weak in our first round, following industry about as much as vendor,
yet 26.5% of vendor switches stay inside them, against 3.2% for a random vendor.
And the backbone that seemed to snap between \htmlClass{x-alpha-disparity}{\alpha} = 0.1 and 0.05
never snaps: no single link cuts off more than two metros.
Sources and the full analysis choices are recorded in Data and methods.
AI use and how we checked it
AI coding assistants helped structure the page, wrote analysis and page code, drafted and revised text, and tested the visual presentation. The numbers come from the public sources listed in Data and methods.
Each analysis step writes the numbers its section quotes to a JSON file the page reads. A schema check tests each file against the fields the page uses, a name-matching check tests the company-name rules against tax numbers, and the site tests fail when the numbers in a card or in this closing drift from the analysis output. We checked generated tables, comparisons, source scope, and the page behaviour against the local data before including a claim.
Background
Cities group by who hires there, not by region. Outsourcing firms bundle jobs differently from direct employers of the same size. And when a client drops its main vendor, the new one comes from the same Louvain group more than eight times as often as a random vendor would, though mostly because clients return to firms they already use. Take out the ten largest filers, most of them national tech and consulting employers, and the metro groups start to follow Census regions; Amazon alone does all of that. Where outsourcing shows most is outside the networks: a filing that places a worker at a client has 3.6 times the odds of a lower wage level for the same occupation. Lawyers and green cards barely follow the staffing groups.
Deep dive
Earlier questions, more networks, and the data behind every number.
Each answer still holds; the questions above replaced them because they test the communities against something that could have come out the other way. Pick any question to open it.
- Which jobs belong to two clusters?→
- Do the clusters follow official job groups?→
- Do occupations the same companies hire together also need similar skills?→
- Does that agreement hold for whole hiring clusters, not just direct ties?→
- How do two occupations' day-to-day skills actually compare?→
- Change the damping factor: does the ranking move?→
- Stepped one round at a time, how fast does the ranking settle?→
Earlier questions now open sections 1 to 3.
Which ones
Where the hiring is
Once the small links go, what's left of the map?
Once the small links go, what's left of the map?
Keep only links heavy for one of their metros and the map comes apart.
The control sets the A filter that keeps a link only when it carries an unusually large share of either metro’s total weight. \htmlClass{x-alpha-disparity}{\alpha} from Week 4.
Backbone at this α
Each line is a link the filter keeps at the \htmlClass{x-alpha-disparity}{\alpha} set above, thicker when more filings share it.
Giant component vs α
How many metros stay in the largest connected piece as the filter tightens. The dashed line marks where it snaps.
Background
With every link the map is one blob. Cities are linked when they share an employer.
Metros in grey have fallen out of the largest connected piece. Colours are section 1's three metro groups. Click a metro to select it.
Method
Two metros are linked when a company files in both; the weight adds up, over those companies, the smaller of its two filing counts. One weight threshold would keep the links among the big hubs and cut a mid-size metro's strongest tie, which is light next to New York and Dallas. The disparity filter keeps a link when it carries an unusually large share of either endpoint's weight at level \htmlClass{x-alpha-disparity}{\alpha}, the method the course used for the philosophers backbone.
| α | Edges kept | Giant component |
|---|
Do the same employers tie distant cities together?
Do the same employers tie distant cities together?
Mostly not: a single company rarely carries a long link.
The shortlist is the five firms that The firm files for the worker, then sends them to work at a client company’s site. the most filings at client sites.
Distance vs weight
Each point is a link in the backbone at \htmlClass{x-alpha-disparity}{\alpha} = 0.2, the default above; size grows with its weight.
One employer’s map
The backbone links (\htmlClass{x-alpha-disparity}{\alpha} = 0.2) this company leads: of all the companies filing in both cities, it adds the most to the link's weight.
Background
The shortlist holds the five largest placing firms: Tata Consultancy Services, Cognizant, Infosys, HCL and Compunnel. The staffing section follows them to their clients.
Method
In the distance chart, a label names a company on its heaviest link over 1,500 km, for the five heaviest; hover any point for its own. The legend separates links led by the shortlist from links led by any other company; click a point to select a city.
In the employer map, the list holds the six companies that lead the most links. Click a city to select it.
More numbers
Big direct employers lead the long links, Amazon above all. Distant metros are tied by many companies filing in both.
Amazon leads the most long links (30), then Cognizant (17), EY (8) and Deloitte (7). The leading company carries a median 14% of a long link's weight, and only one long link, San Jose to Fayetteville (Walmart), has a company with half of it.
Where is the hiring densest? Filings per 1,000 jobs
Where is the hiring densest? Filings per 1,000 jobs
San Jose, at 42.9 filings per 1,000 jobs, against New York's 6.9.
Section 1 counts filings.
Method
Divide each metro's 2025 filings by its jobs in the Bureau of Labor Statistics' May 2025 employment survey (OEWS) and the map shifts: nationally it is 4.5 filings per 1,000 jobs.
A filing is a request, not a hire, so a rate can run high.
More numbers
Among the 203 metros with 100,000 jobs or more, count and density rank alike (A rank correlation: how closely two orderings agree. Higher means closer. 0.90), yet only 5 of the 10 largest by count stay in the top 10 by density: Dallas, San Jose, San Francisco, Seattle and Austin. For software developers alone the national rate is 139 filings per 1,000 jobs, and Fayetteville, Arkansas, the metro around Bentonville, reaches 896, 6.4 times the national share. New York files the most, 65,935; Trenton files 17.2 and Seattle 16.8 per 1,000 jobs.
The course's four community methods, step by step on the 40 metros
The course's week 4 community methods, run on section 1's 40 metros. Each tab steps through one method, so you can watch where its groups come from and set them against A method that finds groups by moving each metro to the neighbouring group that raises modularity most, then merging the groups and repeating..
Loading the community explorables…
Explore · after the course's week 4
Girvan–Newman, one cut at a time
Thick line: the next link to go, the one carrying the most shortest paths. Dashed: links already cut. Hollow dots: metros cut off on their own.
Background
link to every other metro, so each split strands a single metro, starting with .
Explore · after the course's week 4
Move a metro, watch modularity
Grey band: rewired networks in which each company keeps its number of metros, Louvain's best on each, mean and one standard deviation.
Explore · after the course's week 4
Louvain, move by move
Hollow dots are alone in their community. A community of two or more takes the colour of the metro group most of its members end in.
Method
The run uses seed and all weighted links. The moves lift The usual symbol for modularity. from to ; the second level finds nothing to merge.
Explore · after the course's week 4
Overlap: k-cliques and link communities
Dark links and filled dots: the community shown. Rings: metros that belong to two or more communities at this setting.
Method
Link communities group the links, and a metro joins every community its links are in.
Jobs and skills
Which jobs belong to two clusters?
Which jobs belong to two clusters?
None clearly: only … of … pass, fewer than the … chance alone passes.
A ringed node would mark an occupation with more employer ties to a second cluster than any A random copy of the network that keeps each company's number of occupations and each occupation's number of companies. It shows what chance alone produces. gives it; click a node to inspect it.
Occupation network
Background
Colours are A standard method that splits a network into groups whose members link more to each other than to the rest. clusters in the full co-hiring network, named after their largest occupation.
Do the clusters follow official job groups?
Do the clusters follow official job groups?
Only in part: Normalized mutual information: a score for how alike two groupings are, 1 when they match exactly and 0 when they are unrelated. … against … for shuffled labels.
We compare the official groups with the clusters found from hiring patterns.
What each hiring cluster holds, by official group
Every occupation in the four largest clusters, 406 in all, split by its official major group: the three largest named, the rest grey.
How closely the clusters match the official groups
Orange: the hiring clusters. Ring: Infomap's clusters. Grey: the official labels shuffled 100 times over the same clusters, up to their highest score.
Background
The government groups occupations by their first two The Standard Occupational Classification, the federal list of occupation codes. Its first two digits name a major group, such as 15 for computer and mathematical occupations. digits.
Normalised mutual information runs from 0 (unrelated) to 1 (the same groups).
Method
We keep certified H-1B filings and identify companies by tax number, as in the other sections. A link counts the companies that filed for both occupations. Filings still on 2010 codes (…) move to their 2018 successors through O*NET's 2010-to-2019 crosswalk. Louvain runs 100 times on the full projection and the best modularity run is kept; the runs agree at a median NMI of …. The null rewires the company × occupation network … times, keeping each company's number of occupations and each occupation's number of companies, and projects it again; real and rewired networks are scored on their largest connected piece (z = …). An occupation's second cluster is the one its employer ties exceed most over the expectation modularity uses (its strength times the cluster's, over twice the total weight). A ratio above 1, our first rule, marks … occupations, but the rewired networks mark … on average with the same cluster labels. So an occupation now counts only when its ratio beats its own ratio in every rewired network. The disparity filter at \htmlClass{x-alpha-disparity}{\alpha} = …, as in the place section, keeps … of … links and … occupations. Louvain on that backbone finds … clusters, which match the full network's at NMI …, against … between two runs on the full network: the clusters only partly survive the filter. NMI and AMI leave out occupations alone in a cluster and are compared with 100 shuffles of the major-group labels. The same method on 2024 gives clusters that match 2025 at NMI … on the … occupations both years share. Infomap, the random-walk method, finds … clusters of two or more occupations; they agree with Louvain's at NMI … and match the official groups at ….
The shuffled bars keep the clusters fixed and scramble only the official labels. AMI, a version of NMI adjusted for chance agreement, is …. Both scores cover the … occupations in clusters of two or more.
More numbers
The other two clusters of two or more hold two occupations each.
Skills behind the jobs, from O*NET
Loading the O*NET comparison…
PageRank on the jobs network, step by step
Loading the PageRank explorable…
Outsourcing firms and their clients
Do clients group by industry or by the firm that staffs them?
Do clients group by industry or by the firm that staffs them?
By both, weakly, and slightly more by vendor.
More numbers
With every firm–client link counted once, Louvain splits the network into about 65 groups, and the split beats rewired networks that keep everyone's number of partners (modularity 0.57 against 0.53). Among the 1,209 clients with a known industry and two or more firms, the groups match the main vendor at an adjusted mutual information (AMI) of 0.11 and the industry at 0.07; both beat shuffled labels.
AMI corrects NMI for chance, which matters here: these clients have 478 main vendors but only 17 industries. Counting filings pulls clients to their main vendor (AMI 0.48 against 0.07), but that split scores below rewired networks with the same filing counts (0.60 against 0.74), and the vendor's head start is built in: the vendor is a node in the same network, and 86% of these clients land in its group.
Who relies on a single vendor?
Who relies on a single vendor?
Small clients.
14,678 of the 18,900 clients use one firm, but they hold 22% of placed filings.
More numbers
Of the 629 clients with 20 or more filings, 69 get over 90% from one firm, and the median one gets 37% from its largest.
Citigroup, the largest client, uses 114 firms, and Tata Consultancy Services supplies a quarter. The eight largest firms supply only 27% of what the 20 largest clients receive.
The client network, year by year
The client network, year by year
The 25 largest clients
| Client | Sector | Filings | Vendors | Largest vendor | Its share |
|---|
Who supplies the largest clients
The eight firms that place the most filings, and the 20 clients that receive the most, in the year chosen above. Band width is the number of placed filings from a firm to a client; the grey source gathers every other firm. Each client sits next to the named firm that supplies it most. Hover a firm or a client to follow its bands.
Background
Only three sectors get a colour: finance and insurance, manufacturing and health care. Other known sectors are light grey, and the palest dots are clients with no sector on record.
Band width is the number of placed filings from a firm to a client; the grey source gathers every other firm. Each client sits next to the named firm that supplies it most.
With filing counts or without?
With filing counts or without?
Filing counts change the grouping: the two partitions share an Normalised mutual information: how much two groupings of the same clients agree, from unrelated to identical. Unlike AMI, it is not corrected for chance. of …, less than two seeds of either kind (… weighted, … unweighted).
Filing counts pull clients toward vendors: AMI with each client's main vendor rises from … to … when filings count, while AMI with industry stays near ….
Against rewired networks, without weights the real network wins (… against …); with weights it loses (… against …).
| Weighted | Unweighted |
|---|
Method
We ran A method that finds groups in a network by moving nodes between groups until the links inside groups are as dense as they can get. It starts from a random order, so two runs can differ. on the 2025 firm–client network 100 times each way: with links weighted by filings, and with every link counting one.
The A random version of the network that keeps some of its features, here each node's number of partners, to show what chance alone would give. rewires the network so every firm and client keeps its number of partners.
Infomap agrees with Louvain at NMI …. Finer partitions raise every NMI; AMI corrects for that, so it is the number to compare across methods.
More numbers
Infomap, which follows a random walk instead of counting links, splits the same network into … small modules, most of them a firm with its clients. Like weighted Louvain, it follows the vendor far more than the industry (AMI … against …).
The null also deals the filing counts back out at random. Rewiring breaks the network into a median of … pieces, each a free community, so we score each rewired network on its largest piece, as we do the real one. The real network also loses when only the filing counts are shuffled on the real links (…). The real counts leave … of filings on links between groups, against … with shuffled counts: clients that use several firms hold … of the links but … of the filings, and only their links can cross, since a client with one firm sits in that firm's group.
Strong ties, weak ties and pay
Strong ties, weak ties and pay
Here the pattern runs the other way, faintly.
Among friends, the strongest ties sit inside tight groups where your close friends also know each other, and weak ties bridge the groups (Granovetter 1973; Onnela and colleagues confirmed it on millions of phone users in 2007).
Background
A link's overlap measures the tightness: of the firm's other clients and the client's other firms, the share that are linked to each other.
More numbers
Over the 28,104 links where overlap is defined, filings and overlap correlate at Spearman -0.04; with the filing counts shuffled over the same links the correlation is 0.00 ± 0.01 (\htmlClass{x-z}{z} = -6.3; 2024 gives \htmlClass{x-z}{z} = -3.7). Links with one filing have a mean overlap of 0.065, links with 21 or more 0.034. Heavy links mostly belong to the largest firms, whose many clients rarely share other firms, so part of this is size. It agrees with the result above: the real filing counts put weight on links between groups. Each filing states one of four wage levels, from entry (I) to fully competent (IV). Averaged per client over the filings that reach it, the groups explain 13% of the variance in wage level; averaged per firm over all its filings, 3%. None of 1,000 shuffles of the group labels reached either. A client's filings come from the vendors that also decide its group, so part of the 13% is built in. Outsourcing firms file 66% of their applications at level II and 5% at level IV; direct employers file 35% and 22%. From January to June 2026, level IV rose to 17.7% of all filings from 13.6% a year earlier, and level I fell to 18.0% from 21.8%.
Strength against degree: where do the heavy links go?
Strength against degree: where do the heavy links go?
Three of the top five are therapy and rehab clinics.
The course compares a node's degree (how many partners) with its strength (how many filings over all its links), and finds the exceptions tell the story.
Background
More numbers
In the 2025 firm–client network the two rank firms almost alike (Spearman 0.91) and clients less so (0.75). The clients with the most filings from a single firm are Ultimate Therapy, 133 filings from one firm; Sigma Rehab, 95; Post Rehab Services, 61; and Grady Memorial Hospital, 58.
Every worker and company, grouped by what they do
Do workers group by job and pay, or by who files for them?
Loading the filings…
What the groups follow
NMI between the groups and each label, weighted by workers. Hollow dots were never part of the network. Bands: the same label shuffled, the level a label with that many values reaches by chance. Dashed: two Louvain seeds against each other.
The network against random ones
Dots: the real network. Bands: rewired networks that keep every degree, or shuffled kinds.
How many links?
Share of occupations, metros, levels and sectors with at least a given number of links or workers, log-log.
The largest groups
| # | Group | Workers | H-1B | PERM | Main occupation | Largest employers |
|---|
Paperwork, the lottery and green cards
Who files the paperwork?
Who files the paperwork?
Three filings in four name an outside law firm, and five firms file 40% of those.
Background
Two law firms share filings when the same employer uses both: for each such employer, the smaller of its filings through either. That network has one giant hub, the case the A way to thin a network: it keeps a link only when its weight is larger than chance would give, judged against each firm's own total. α sets how strict that test is. was made for. A weight threshold of four shared filings keeps 771 links and spends 26% of them on the five largest firms. The disparity filter at \htmlClass{x-alpha-disparity}{\alpha} = 0.2 keeps 697 and spends 19%. It also keeps 115 small firms the threshold drops, such as one law office whose link to BBI Law Group is 3 of its 5 shared filings and 3 of BBI's 342. The threshold keeps the larger connected core, 395 firms against the filter's 351. Another 144 firms stay only because they form a pair linked to nobody else, where neither end can judge the link.
A method that finds groups in a network by moving nodes between groups until the links inside groups are as dense as they can get. finds 34 groups at How much more of the link weight falls inside the groups than a random network would put there. Higher means cleaner groups. 0.67, against 0.55 for rewired networks that keep each employer's and each law firm's number of partners (\htmlClass{x-z}{z} = 12). The groups are barely regional (Normalised mutual information: how much two groupings of the same items agree, from 0 (unrelated) to 1 (identical). 0.05 with Census regions, though above every shuffle), and the largest gather around shared employers. Google, Apple and Meta share Fragomen, Ogletree Deakins and Berry Appleman & Leiden; Tata Consultancy, LTIMindtree and Salesforce share Usilaw, Goel & Anderson and Chugh; Vialto, once PwC's law firm, serves Doordash and Databricks. The groups move from year to year: 2024 and 2025 agree at NMI 0.30 on the 1,041 law firms in both, against 0.88 between two runs of 2025.
More numbers
Fragomen alone files 78,531 for 3,129 employers. Outsourcing firms mostly do without: they file 52% of their applications with no outside firm and send 3% to the five largest, while direct employers send those five 48%. Per employer the averages are 2% and 30%, and none of 1,000 shuffles of which employer is which produced a gap that wide.
Do the firms that register the same workers staff the same clients?
Do the firms that register the same workers staff the same clients?
No. Firms that register the same workers are spread across the staffing groups.
We took the 2,942 firms in the 2023 staffing network that sent 20 or more registrations to the March 2023 draw, and split them at the median share of workers another employer had also registered (78%).
More numbers
If the high firms clustered, a high firm's Louvain group would be mostly high firms. It is 51.2% high, against 50.0% when the labels are shuffled (\htmlClass{x-p-empirical}{p} = 0.001 over 1,000 shuffles), and Adjusted mutual information: how closely two ways of grouping the same firms agree. 0 is what chance gives, 1 is a perfect match. with the groups is 0.004 over 100 runs. The March 2022 draw against the 2022 network gives 51.9% against 50.0%. The lottery data is USCIS's, obtained by Bloomberg News.
The lottery a year apart, and who receives the winners
The lottery a year apart, and who receives the winners
One approved petition took 5.2 registrations in 2022 and 8.5 in 2023.
The ratio divides each year's registrations by its approved petitions, using USCIS's lottery files, obtained by Bloomberg News.
Method
From March 2024 USCIS drew by person, not by registration.
The every-draw chart divides eligible by selected registrations from the historical table on USCIS's H-1B Electronic Registration Process page; its selections include later rounds, so its ratio is lower than the one per approved petition.
More numbers
Between the March 2022 and March 2023 draws, Entries in the H-1B lottery. Each spring employers register the workers they want to sponsor, and USCIS draws at random from the entries. for a worker whom another employer had also registered rose from 35% to 54% of the total, and the share of drawn registrations that became a petition fell from 74% to 49%. The order held both years: direct employers needed the fewest (4.3, then 5.1), placing firms more (6.3, then 9.1). Of the March 2023 petitions, 20% lead to a client company, 66% of those through placing firms. Citigroup received 342, 94% through placing firms; Microsoft 150, 95%, with LTIMindtree supplying 47%; AT&T 153, with Tech Mahindra supplying 41%. USCIS also denied 2.0% of the placing firms' lottery petitions against 1.1% of direct employers'.
USCIS denials, year by year
USCIS denials, year by year
Share of first-time H-1B petitions USCIS denied, for employers with 20 or more certified filings that year.
From the USCIS's H-1B Employer Data Hub, which publishes approved and denied petitions per employer.'s Tableau export; 2026 runs October to June. Placing firms put half or more of their filings at a client.
| Year | Placing firms | Direct employers | Placing firms counted | Direct employers counted |
|---|---|---|---|---|
| 2022 | 2.71% | 1.22% | 1,057 | 2,288 |
| 2023 | 3.57% | 1.48% | 873 | 2,080 |
| 2024 | 2.65% | 1.28% | 884 | 2,403 |
| 2025 | 3.20% | 1.56% | 933 | 2,466 |
| 2026, Oct–Jun | 3.35% | 1.98% | 601 | 1,863 |
Who keeps them? Green cards as the strong tie
Who keeps them? Green cards as the strong tie
An H-1B filing is a weak tie between employer and worker; a green-card filing (The Department of Labor's permanent labor certification, the first step of most green cards through work. The employer sponsors the worker to stay.) is a strong one.
Only 65% of certified green cards come from an employer we can match to an H-1B filer.
Method
More numbers
In 2025 the median employer with 20 or more H-1B filings filed 13.2 green cards per 100 of them, and the two counts rank employers only loosely alike (A rank correlation: 1 means two counts put employers in the same order, 0 means no relation. 0.50). As with degree and strength in the course, the exceptions carry the story: Oracle filed 95 green cards per 100 H-1B filings, Uber 64 and Salesforce 45, while Amazon (22,509 H-1B filings), Cognizant (11,085) and Google (8,657) filed almost none. Whether outsourcing firms sponsor fewer is section 5B. Clients sponsor their own staff too: Wells Fargo receives 1,547 H-1B filings from vendors, files 624 of its own and 167 green cards. Firms with more clients sponsor slightly more green cards, not fewer (Spearman 0.13).
Where are they from? A network of countries
Where are they from? A network of countries
Green-card hiring does not sort countries into regional blocs.
We link two countries when the same employers file green cards for citizens of both, then ask whether those links form regional groups. The groups match world regions only weakly (AMI 0.10).
Method
A worker's citizenship is personal, so we read it only in memory and keep counts per country and employer, dropping every count under 10. That drops 96% of the cells and 43% of 2023's certified green-card filings, and no row about a single person is ever saved.
More numbers
India holds 52% of those filings and China 12%; among lottery registrations, India holds 77% in the March 2022 draw and 81% in March 2023. Link two countries by the green cards their citizens receive at the same employers and 55 countries remain. Louvain splits them into three groups. Counted once each, the links group a little more than rewired copies (modularity 0.10 against 0.06); weighted by shared green cards, they group less (0.26 against 0.40). India and China share one group with Canada, Belarus and Costa Rica, and India keeps 79% of its weight inside it. Among the ten largest sponsors, Google's green cards are the most varied, 4.7 The number of equally sized countries that would give the same spread. More means more varied. with India at 30%, against Amazon's 3.0 at 67%. The third group gathers the Philippines, Kenya, Ghana, Zimbabwe, Ethiopia, Cameroon and Jamaica. Wayne Farms, a poultry company, filed 832 green cards in the counted cells, none for Indian citizens.
The groups match world regions (AMI 0.10) and Week 3's migration communities (0.10) only weakly: green-card hiring does not sort countries into regional blocs.
Five years
Five years of filings, 2022 to 2026
Loading five years of filings…
Who filed for which roles, 2022 to 2026
Certified filings, stacked by detailed occupation, SOC major group or employer, across the same five fiscal years as the box above. Switch the split, the scale or the window; 2026 covers October 2025 to June 2026 only.
Who filed, and for which roles?
Loading…
Each band is one series, largest at the bottom. Hover a band for its numbers; click a legend entry to hide it.
Does it hold from year to year?
Does it hold from year to year?
Only in part.
Consecutive years agree less than two runs of the same year, so the groups carry over only in part.
Background
We compared vendor and industry on 2025 only.
Method
We compare each year's client groups with the next year's, on the clients both years share, and with a second run on the same year as the ceiling.
More numbers
On the clients present in both years, consecutive years agree at NMI 0.21 to 0.27, about half the 0.48 to 0.50 between two runs of the same year on the same clients.
2026 breaks the pattern at the top. We compare January to June of each year, because October 2025, the month of the federal shutdown, holds 1,306 certified filings against 35,258 a year earlier. From January to June, certified filings fell 5.8% after rising 9.5% the year before, and filings that name a client company fell 16.5% after holding flat (-0.1%). Tata Consultancy Services filed 2,079, down from 5,256. Of the 349 clients it supplied most from January to June 2025, 223 still appear, and 111 of those now get most of their workers from another firm, most often Infosys. Across clients with five or more filings in both years, 47% changed their main vendor, against 44% a year earlier. These are applications, so they show what employers asked for, not why.
Data and methods
The numbers on this page come from these public sources.
- US Department of Labor, Office of Foreign Labor Certification ↗
- Labor Condition Application: the form an employer files with the Department of Labor before it can sponsor an H-1B worker. disclosures and worksites file, public domain
- USCIS H-1B Employer Data Hub ↗
- October 2021 to June 2026
- H-1B lottery registrations ↗
- the March 2021 to March 2023 draws, USCIS data obtained by Bloomberg News under The Freedom of Information Act, which lets anyone request records from US federal agencies.
- Census CBSA delineations ↗
- counties to metro areas
- Bureau of Labor Statistics OEWS ↗
- jobs per metro and occupation, May 2025
- SEC EDGAR ↗
- industry codes of listed clients
- Wikidata ↗
- industries of clients the SEC does not list, CC0
- O*NET database ↗
- skill, knowledge and work-activity ratings per occupation, US Department of Labor, CC BY 4.0