The US Census Bureau's County Business Patterns counted 68,051 custom computer programming establishments (NAICS 541511) in 2023, employing 1,012,250 people. Seventy-two percent of those establishments had fewer than five employees. Only 23% had between 5 and 49, which is the size of the agency you probably run, and just 5% had 50 or more. The peers who could absorb your overflow are rarer than the headline count suggests.
What does the Census count, and what doesn't it?
County Business Patterns (CBP) is the Census Bureau's annual tally of establishments with paid employees, built from federal administrative records and the Economic Census. The 2023 release is the latest the Bureau has published. NAICS 541511, Custom Computer Programming Services, is the code for writing software to a customer's specification. That is the bucket most dev shops, product studios and fractional-CTO practices file under, but the code also holds contract programmers placed with a single client and in-house teams of consulting firms.
Two limits matter before you quote any of this to a client. CBP counts establishments, meaning locations, not companies, so a firm with two offices counts twice. And it excludes anyone with no employees on payroll, so a solo freelancer who pays only themselves is invisible. The real population of people who sell development is larger than 68,051. The population who have a payroll, and so some chance of a bench, is the one this data describes.
How are those 68,051 shops split by size?
Here is the Census Bureau's own breakdown of 2023 establishments and employment for NAICS 541511, by employees per establishment.
| Employees per establishment | Establishments | Share of establishments | Employees | Share of employees |
|---|---|---|---|---|
| Under 5 | 49,009 | 72.0% | 66,294 | 6.5% |
| 5 to 9 | 6,850 | 10.1% | 45,460 | 4.5% |
| 10 to 19 | 4,795 | 7.0% | 65,304 | 6.5% |
| 20 to 49 | 4,044 | 5.9% | 125,991 | 12.4% |
| 50 or more | 3,353 | 4.9% | 709,201 | 70.1% |
Read the last column against the first. Establishments of 5 to 49 people are 23.1% of the count and employ 236,755 people, 23.4% of the industry. Seventy percent of the people who write custom software for a living in this code work in the 3,353 establishments with 50 or more. At the bottom, 49,009 shops share 66,294 employees, roughly one and a third people each.
Why does that make overflow hard to find among peers?
Take the owner of a 20-person agency who signs a project that needs two more senior engineers by the first of the month. The instinct is to call another agency. Run the CBP numbers on who picks up.
Seven in ten establishments are under five people. A four-person shop might take your work, but it has no bench: each person it lends you is a person its own clients lose. The mid-sized peers in the 5-to-49 band, about 15,700 establishments nationwide, are staffed against their own backlog, which is the same problem you have. That leaves the 3,353 establishments with 50 or more, which hold most of the capacity. Whether a firm that size would rent you two people rather than sell you a project is a question only a phone call answers.
I can't tell you from this data that none of them will help. CBP doesn't measure utilisation. What it does show is that the shape of the supply is lopsided: a lot of very small sellers and a thin layer at the top. If your plan for a capacity spike is to find a similarly sized agency with two idle engineers, the odds are not in the data's favour. That's an opinion, and the useful test is cheap: email five peers this week and count how many have anyone free in 30 days.
Has the middle moved since 2018?
Compare the 2018 release of the same series. Establishments in NAICS 541511 grew from 65,733 to 68,051 between 2018 and 2023, up 3.5%. Employment grew faster, from 886,213 to 1,012,250, up 14.2%. The 5-to-49 band barely moved: 15,557 establishments in 2018 against 15,689 in 2023, up 0.8%. The 50-or-more group grew from 3,003 establishments to 3,353, up 11.7%.
So employment growth over those five years went mostly to the large end, and the agency-sized middle held its count. Pay moved too. Census reports annual payroll per establishment size class, and dividing by employment, the 20-to-49 band paid $117,942 per employee in 2023 against $100,671 in 2018, up 17.2%. That figure covers everyone on payroll, project managers and sales included, not only engineers, and it excludes benefits and payroll taxes. It still tells you the direction your fixed cost is going. For an owner pricing a fixed-bid project, the cost of the bench you already carry rose about 17% in five years while the number of peers your size did not change.
What does this change about how you staff a spike?
Picture a 24-person agency that wins a fixed-bid project needing two extra engineers for five months. It has three realistic routes: hire, borrow from a peer, or add a subcontracted pod. Hiring two people at the 20-to-49 band's average payroll of roughly $118,000 each commits about $236,000 a year in base pay before benefits for a five-month need. Borrowing from a peer runs into the supply problem above. A subcontracted pod avoids the fixed cost but carries its own risks: overlap hours, IP assignment and contract terms, which are the reason the economics of white-label development and the overlap-hours table exist.
None of these is right every time. Hiring is right when the work will outlast the project, because a hire you keep for three years amortises. A peer is right when the work needs one narrow specialist for a week. A subcontracted pod fits when the need is real but bounded, and you want to judge the engineers on a small task before you commit. The Bureau of Labor Statistics found that computer systems design shed 29,000 jobs in the year to July 2026. Whether or not you read that as a reason to keep headcount fixed, it is the pressure that makes renting capacity for a bounded project worth pricing.
When is outside capacity the wrong answer?
It's wrong when the work needs a US person on the project. Defence and some federal work can require US-person-only handling, and ivector's partner page says so plainly on the first call. It's wrong when you need twenty engineers on Monday, since a pod-at-a-time model is built to add two or three. And it's wrong when your real problem is that you under-priced the project; more capacity at the same price loses money faster.
If a bounded spike is your situation, the cheapest way to find out whether a subcontracted engineer fits your standard is a paid pilot on one real backlog item, in your own repo, with the price fixed in writing first. Details of how ivector runs one are on the agency partners page, and you can also describe your spike through the contact form if you would like a scoped answer before deciding.
