The Developer-Count Source Reference: Accounts, Pushers, or Contributors
For any GitHub-derived developer figure, state its unit, location rule, window, suppression and lag, and the one question it can honestly answer in a deck.
Key takeaways
- Octoverse places developers by self-reported profile location, while the GitHub Innovation Graph places them by IP-derived mode daily location, so the two series should never be differenced as if the gap were growth.
- For the US, India, China and Brazil, the Innovation Graph Q1 2026 developer count is 0.82 to 0.92 of the Octoverse September 2026 figure, which is consistent with a window difference rather than proof that one location rule is wrong.
- US GitHub accounts in Octoverse 2026 (38.3M) are about 22x US Bureau of Labor Statistics employment in software developers (1.718M), so an account total is a ceiling, never a labor-force count.
- The Innovation Graph language series counts anyone with at least one push to a repository containing that language, so mixed repos inflate every language they contain and data formats such as yaml and json must be dropped by hand.
- The Innovation Graph withholds any economy with fewer than 100 developers doing the activity in the quarter, so a missing country in a language file is suppressed, not zero.
- In Refolk's index, Python-skilled engineer profiles outnumber Rust-skilled profiles in Germany 35 to 1 (4,831 to 138), an employment-weighted ratio to test the push-based language series against.
This reference is for strategy, research and talent-intelligence analysts who have to put one defensible number on "how many developers work in language X in country Y" for a hub, expansion or market-sizing deck. Each published figure built from public code-platform activity gets its own entry: the unit it counts, how it assigns geography, its window, suppression rule and lag, whether it survives a definition change, and the narrower question it can honestly answer. Find the entry for the figure you have, read it, and go back to your deck.
Which unit does each published developer figure count?
GitHub publishes at least three different units under the word "developers": registered accounts placed by profile text, accounts placed by IP address within a quarter, and pushers or monthly contributors in the language series. None of them counts employed people.
Two terms first. An economy is GitHub's word for a country or territory in its open data. Mode daily location is the economy that an account's IP address most often places it in, counted day by day across the period.
| Figure | Unit counted | Location rule | Cadence |
|---|---|---|---|
| Octoverse country total | Registered account | Self-reported profile, ISO-coded | Annual |
| Octoverse language ranking | Monthly contributor | None, platform-wide | Annual |
| Innovation Graph developers | Developer account in the quarter | IP mode daily location | Quarterly |
| Innovation Graph languages, licenses, topics | Developer who pushed to a matching repo | IP mode daily location | Quarterly |
| Innovation Graph repositories, organizations | Repo or org, not people | Mode location of members | Quarterly |
Octoverse is GitHub's annual report. The Innovation Graph is GitHub's quarterly open dataset, published as CSV files in the public github/innovationgraph repository along with a datasheet and a methodology page. Most of the disagreement analysts run into comes from comparing figures in different rows of this table.
The entries: unit, geography, window, suppression, lag, and limits
Each entry below lists the same fields so you can compare them line by line. If a field says "not established", the evidence for it is not public, and the entry tells you how to check it yourself.
Octoverse country total
- Unit: registered accounts. More than 225M worldwide in Octoverse 2026; more than 180M in Octoverse 2025.
- Geography: self-reported profile location, standardized to ISO country codes. Profile text can be fictional or missing.
- Window: annual. The 2025 edition ran from 1 September 2024 to 31 August 2025. The 2026 edition is cut in September 2026.
- Suppression: a 100-developer publication threshold, kept in the 2026 edition.
- Lag: one release a year. Read the edition's stated cut date before using any figure.
- Comparable across editions: no. Octoverse 2026 recalculated its 2024 and 2025 figures for a new time frame that accounts for private repositories. That is reported in a secondary source and the restated values are unverified, so take every year from a single edition.
- Inflated by: anything that is an account. The figure cannot separate bots, students or one person's second account, and I found no public ratio that merges one person's multiple identities.
- Answers: how large is the GitHub account base that says it is in country Y?
- Cannot answer: how many people work as developers in Y, or anything by language.
Octoverse language ranking
- Unit: monthly contributors, also described as "monthly participating accounts".
- Geography: none. The ranking covers the whole platform, not individual countries.
- Answers: which languages are gaining or losing share of contributors globally.
- Cannot answer: how many Python developers work in Germany. Do not multiply a global language share by a country account total. The two series count different units, and the product of two mismatched units means nothing.
Innovation Graph developers
- Unit: developer accounts located in an economy. Bots and accounts that GitHub's internal systems flag as "spammy" are excluded.
- Geography: IP-derived mode daily location. A VPN or travel can move an account.
- Window: one calendar quarter, with history back to January 2020.
- Suppression: any economy with fewer than 100 unique developers doing the activity in the quarter is withheld. This file covers more than 215 economies.
- Lag: nominally one quarter. In practice the Q1 2026 data came out on 5 August 2026, about four months after the quarter closed.
- Scope: public activity only.
- Comparable across quarters: within the series, yes. I found no public dates for changes to the location method or the bot filter, so download the full history again at each release instead of appending new quarters to an old extract.
- Answers: how many non-bot accounts with public activity were placed in Y during this quarter?
- Cannot answer: employment, or work that happens only in private or enterprise repositories.
Innovation Graph languages, licenses and topics
- Unit: unique developers in an economy who pushed code to a repository with a given language, license or topic.
- Trigger: at least one git push to a repo that contains the language. The developer does not need to have written any of that language.
- Classifier: Linguist, GitHub's language classifier, leaves out vendored, generated and documentation files when it computes a repo's language statistics. I found no source confirming that the Innovation Graph applies the same exclusions. Check it yourself by comparing a known repo's language bar with your own reading of the repo.
- Inflated by: mixed repos. One push counts the developer under every language in the repo. How much this inflates counts compared with primary-language repos is not publicly established. Researchers drop yaml, json, text, svg, Markdown and xml by hand.
- Coverage: most of these files cover about 110 to 180 economies, so suppression removes many more economies here than in the developers file. One study built on this data covered 163 economies and 150 languages from 2020 to 2023.
- Answers: how many accounts in Y pushed to a repo containing X this quarter?
- Cannot answer: how many people write X for a living.
How one push becomes a language count
- PushA developer makes at least one git push to a public repository
- ClassifyThe repository's languages are detected, including data and markup formats
- CountThe developer is counted once under every language present in that repository
- PlaceThe developer is assigned an economy by IP-derived mode daily location
- SuppressEconomies with fewer than 100 developers for that language are withheld
Innovation Graph repositories and organizations
- Unit: repositories and organizations, not people.
- Geography: a repository is placed by the mode location of all members with triage access or above. An organization is placed by the mode location of all its members.
- Answers: where code ownership sits.
- Cannot answer: how many developers there are. Never add these counts to a headcount.
How far apart are the Octoverse and Innovation Graph country counts?
For the four countries where both figures are public, the Innovation Graph Q1 2026 count is between 0.82 and 0.92 of the Octoverse September 2026 count, and above the Octoverse 2025 count. Germany's Innovation Graph figure could not be retrieved, so it is not in the table.
| Country | Octoverse 2025 (M) | Innovation Graph Q1 2026 (M, share of Oct 2026) | Octoverse 2026 (M) |
|---|---|---|---|
| US | 28.0 | 32.41 (0.85) | 38.3 |
| India | 21.9 | 26.73 (0.84) | 32.0 |
| China | 10.7 | 11.35 (0.92) | 12.4 |
| Brazil | 6.89 | 8.32 (0.82) | 10.2 |
The Innovation Graph column comes from a third-party processing of the public files, and the share column is derived. Three things separate the series:
- Window. Q1 2026 closes six months before the Octoverse 2026 cut, so a growing account base will put the quarterly figure lower.
- Location rule. Octoverse uses profile text; the Innovation Graph uses IP address.
- Restatement. Octoverse 2026 recalculated earlier years, so the 2025 column may not match what the 2026 edition now holds for that period.
In every row the quarterly figure falls between the two annual snapshots, which is what you would expect if the window does most of the work. That means the gap does not show that one location method is wrong. It also means it is not established how much of the gap the location rule explains. To isolate it, you would need both series for the same window, and that is not public.
How do account counts compare with employment and survey estimates?
US account counts run 16x to 22x official US employment in software developers, and global account counts are about 3.8x a survey-based estimate of all developers worldwide. Account totals and labor-force numbers measure different things, and the slide has to say so.
| Numerator | Denominator | Ratio |
|---|---|---|
| Octoverse 2025: 28.0M | BLS 15-1252: 1.718M | 16.3x |
| Innovation Graph Q1 2026: 32.41M | BLS 15-1252: 1.718M | 18.9x |
| Octoverse 2026: 38.3M | BLS 15-1252: 1.718M | 22.3x |
| Octoverse 2026: 38.3M | BLS computer occupations: 5.036M | 7.6x |
| Global Octoverse 2025: 180M | Survey estimate: 47.2M | 3.8x |
BLS is the US Bureau of Labor Statistics. SOC 15-1252 is the Standard Occupational Classification code for Software Developers, with 1,717,800 employed in 2025. Computer occupations as a whole come to 5,036,200. As a second official reading, the BLS occupational employment and wage survey (OEWS) counted 1,656,880 software developers in May 2023. The survey estimate is an independent count of 47.2M developers in early 2025, of whom 36.5M were professional. All ratios above are derived.
The multiple rose from about 16x to 22x across the two Octoverse editions. Account sign-ups grow much faster than employment, so expect the multiple to keep moving. Part of the reported 10.3M one-year US increase may come from the 2026 restatement rather than from new accounts, and the edition-consistent figure has not been verified. I found no comparable official-occupation ratios for countries other than the US. The global survey ratio is the only non-US benchmark in this reference.
From US accounts to US software developers
- 38.3MOctoverse 2026 accounts
Profile location, annual
- 32.41MInnovation Graph Q1 2026
IP location, public activity, one quarter
- 5.04MBLS computer occupations
Employed, all computer roles
- 1.72MBLS software developers
Employed, SOC 15-1252
Octoverse 2025 also projected 57.5M developers in India and 54.7M in the US by 2030. Those are projections of accounts. Do not use them as a workforce forecast.
An account total is the ceiling of a market, never its size.
What does a bottom-up profile count add?
A named-profile count gives employment-weighted ratios between slices and a list of real people. It is far smaller than any account total, and it does not replace one. In Refolk's index of professional profiles, the counts below cover profiles that are both skill-tagged and title-filtered to engineer or developer roles.
| Slice | Refolk count | Octoverse 2026 accounts | Derived |
|---|---|---|---|
| Python, Germany | 4,831 | 6.30M | 0.077% |
| Python, UK | 5,874 | 6.47M | 0.091% |
| Rust, Germany | 138 | 6.30M | Python:Rust 35:1 |
The percentage column is not a coverage rate. It only shows how far a skill-and-title filter on professional profiles sits from a raw account base. The ratios are what to use:
- Country to country. UK:Germany is 1.22x for Python in Refolk's index against 1.03x in Octoverse account totals. Account counts make the two markets look almost equal. Professional profiles give the UK a clearer lead.
- Language to language. Python:Rust in Germany is 35:1. Compare that with the same pair in the Innovation Graph languages file for Germany. If the push-based ratio is much closer, mixed repositories are probably inflating the smaller language.
- Where inside the country. For Python in Germany the largest region is Berlin. For Python in the UK, London dominates the sample, with 12 of 25.
- Who employs them. The top employers of Rust-skilled profiles in Germany are SAP, Helsing and Wolt.
The hard part of this job is getting from "32 million accounts" to "these are the people who would actually take the role". Public files cannot do that step. Refolk handles it by taking a plain-English request and returning named profiles across GitHub, LinkedIn and the open web, so the bottom-up check takes an hour instead of a week of boolean strings.
How developer counts mislead: failure modes and false positives
Most bad developer-population slides fail in one of seven ways. Each one produces a confident number that answers a different question from the one the deck asked. For each failure below you get the false positive it produces and the check that catches it.
1. Account totals placed next to labor statistics
- False positive: "38.3M US developers" on the same slide as 1.7M from BLS, presented as two measures of one thing.
- Check: compute the multiple from the employment table above and print it beside the figure. If the slide cannot carry a 22x footnote, it should not carry the account total.
2. Mixing location rules
- False positive: a country "grew 15%" because the analyst subtracted a profile-located Octoverse figure from an IP-located Innovation Graph figure.
- Why it lies: a VPN or travel moves IP location, and profile text can be fictional or missing. Octoverse 2026 itself adjusts pushes using profile-country-versus-IP ratios extracted by a language model, which shows GitHub accepts that VPN use biases IP location. Changing the location rule can change country rankings.
- Check: never difference across series. Growth comes from one series.
3. Dormant account bases
- False positive: a large account count read as a large, active talent pool.
- Example: in Q1 2026 China had 11.35M developer accounts and 2.92M git pushes, or 0.26 pushes per account, against a top-20 median of 1.65. That is about a sixth of the median, so China's share of accounts overstates its active share.
- Check: pushes per developer against the top-20 median for every economy in the deck.
Reading an account count against push intensity
4. Language inflation
- False positive: a language looks like it has a large community in a country because it appears in many mixed repositories.
- Why it lies: one push to a mixed repo counts the developer under every language that repo contains, and data and markup formats appear in nearly everything.
- Check: drop yaml, json, text, svg, Markdown and xml, then compare each language count with developers.csv for the same economy and quarter. A language count close to the economy's whole developer count is a sign of inflation, not of dominance.
5. Suppressed cells read as zero
- False positive: "no Rust activity in country Z" because Z is missing from the languages file.
- Check: confirm whether the economy is absent from that file, or appears with a value. Absent usually means fewer than 100 developers for that activity in that quarter.
6. Restated history
- False positive: a year-on-year growth rate built from the 2025 figure in the 2025 edition and the 2026 figure in the 2026 edition.
- Why it lies: Octoverse 2026 recalculated 2024 and 2025 for a new time frame that accounts for private repositories.
- Check: take every year in a comparison from one edition. If the restated values are not published in a form you can verify, say the growth rate is not established.
7. Index counts read as a census
- False positive: "there are 138 Rust developers in Germany".
- Why it lies: a skill-tag and title filter undercounts by design. It finds people who say what they do, in the words the filter expects.
- Check: use Refolk's figures for ratios between slices and for finding named people, never as the total size of a market.
Step-by-step: from a deck question to a defensible developer count
The procedure takes about four analyst-days from question to triangulated figure. Most of that time goes into pulling the files and triangulating.
Developer-count procedure
- Fix the questionWrite one sentence naming the unit: registered accounts, active pushers, a language community, or employed workers. About 30 minutes.
- Pick the seriesChoose Octoverse (annual, profile location) or the Innovation Graph (quarterly, IP location, public only) to match that unit, and fill in the source row. About an hour.
- Pull the raw filesClone github/innovationgraph and filter developers.csv and languages.csv. Done when quarters and economies are listed and suppressed cells are flagged, in one analyst-day.
- Align windowsMap each quarter to the Octoverse year it sits inside and write the period beside every comparison. About half a day.
- Clean the language seriesDrop yaml, json, text, svg, Markdown and xml, then check each language count against developers.csv. About half a day.
- Check push intensityCompute pushes per developer and flag economies far below the top-20 median of 1.65 as dormant bases. About an hour.
- TriangulateDivide the account figure by an official occupation count or a survey estimate and put the multiple on the slide. About a day.
- Run a bottom-up checkCount named, skill-tagged profiles for the hiring-relevant slice and use them for ratios and names, not totals. About an hour.
A few steps need more detail:
- Step 2, which series. Sources disagree here. One practitioner write-up tells readers to prefer the Innovation Graph for anything at country level. I agree for comparisons across countries, quarters and languages, because it is one public file with one location rule and history back to Q1 2020. Use Octoverse when your audience has already seen its headline, and footnote its profile-based location rule so nobody puts the two side by side.
- Step 3, suppression flags. Add an explicit status column to your extract (reported or suppressed) before doing anything else. Once a blank cell turns into a zero in a pivot table, nobody notices.
- Step 7, which denominator. For the US, use BLS 15-1252 for the "software developer" question and all computer occupations for the broader one, and state which one you used. For other countries I have no official-occupation ratio to give you. Use the global survey ratio of 3.8x only as an order-of-magnitude check, and say that is what it is.
- Step 8, the slice. Run the query for the role you are actually hiring for, not the whole language. In Refolk, a plain request such as Python data engineers in a named city with a minimum experience level returns the slice directly, ready to count by city and employer.
Record every figure with the same fields before it goes into the deck:
Figure and value: Publisher and edition or quarter: Unit counted: registered account | account in quarter | pusher to matching repo | monthly contributor | employed worker Location rule: self-reported profile | IP mode daily location | survey | establishment survey Window (start and end dates): Suppression rule and whether this cell was suppressed: Release date and lag from window end: Public activity only, or includes private: Definition changes or restatements since last pull: Pushes per developer vs top-20 median of 1.65: Ratio to employment or survey benchmark: Question this figure answers: Question this figure cannot answer:
Fill one row per figure on the slide; leave a field as "not established" rather than guessing.
Keeping the number current
A developer count goes out of date in three ways: a new Innovation Graph quarter arrives, a new Octoverse edition restates history, or a method change happens without notice. Re-check each one by mechanism rather than relying on the value you used last time.
- New quarter. The Innovation Graph is nominally one quarter behind, but Q1 2026 arrived on 5 August 2026. Look for the release on the methodology page or the repository's commit history, and rebuild your extract from the full history each time.
- New Octoverse edition. Read the edition's stated window and any note on recalculated prior years before reusing a single number. If earlier years were restated, replace the whole series, not just the latest point.
- Silent method changes. No dates are documented for changes to the Innovation Graph's location method or bot filter. Compare the previous quarter's values in your old extract with the same quarter in the new download. Any difference means something changed upstream.
- Benchmarks. BLS publishes employment projections and OEWS on its own schedule. Update the denominator when it changes, and update the multiple on the slide at the same time.
- Bottom-up slice. Rerun the profile query when the role definition changes. Slice ratios are more stable than totals, so a large change in Python:Rust or UK:Germany deserves a second look.
Before the slide ships, put a footnote like this under the figure:
Source: GitHub Innovation Graph, developers series, Q1 2026 (released 5 August 2026). Counts developer accounts placed in an economy by IP-derived mode daily location, public activity only, bots and spam-flagged accounts excluded; economies under 100 developers are withheld. This is not a count of employed developers: the US figure of 32.41M is about 19x US BLS employment in Software Developers (SOC 15-1252, 1.718M).
Swap in your figure, quarter, release date and benchmark; keep every clause.
Before the developer count goes in the deck
- The slide names the unit counted: registered accounts, accounts in a quarter, pushers, monthly contributors, or employed workers.
- The location rule (self-reported profile or IP mode daily location) is stated in the footnote.
- Every figure shows its window, and no comparison mixes an Octoverse figure with an Innovation Graph figure.
- All years in any growth rate come from a single Octoverse edition or a single Innovation Graph download.
- Every suppressed cell is reported as fewer than 100, not as zero.
- Language counts exclude yaml, json, text, svg, Markdown and xml, and none exceeds its economy's developer count.
- Pushes per developer are computed and any dormant base is flagged against the 1.65 median.
- The ratio to an official occupation count or survey estimate is printed beside the account figure.
- Any profile-index count is presented as a ratio or a named slice, not as a market total.
Questions practitioners ask
How many developers are there in a country according to GitHub data?
It depends on which GitHub figure you use, and none of them counts employed developers. Octoverse 2026 reports 38.3M US accounts placed by profile location; the Innovation Graph reports 32,410,529 US developer accounts for Q1 2026 placed by IP. Both are account counts. Against US Bureau of Labor Statistics employment of 1.718M software developers, they run roughly 19x to 22x. Quote the figure with its unit, location rule and window.
Why do Octoverse and Innovation Graph country figures disagree?
They disagree because they use different windows and different location rules. Octoverse places accounts by self-reported profile location over an annual window; the Innovation Graph places them by IP-derived mode daily location for one quarter and covers public activity only. For four countries checked, the Q1 2026 Innovation Graph figure sat between the Octoverse 2025 and 2026 figures, which points to the window as the main driver. Never difference one against the other as growth.
How do I find the number of Python developers by country?
Use the Innovation Graph languages file, read as accounts that pushed to a repository containing Python in that economy during the quarter. It is not a count of professional Python developers. Mixed repositories count a developer under every language present, and data formats like yaml and json must be dropped by hand. For an employment-weighted check, compare slice ratios from a named-profile index: Refolk's index has 4,831 Python-skilled engineer profiles in Germany and 5,874 in the UK.
Can I put a GitHub developer count next to labor force statistics?
Only if you state the multiple between them. US accounts in Octoverse 2026 are 22.3x Bureau of Labor Statistics software developer employment and 7.6x all computer occupations. Those numbers measure different things: accounts versus employed people. Put the ratio on the slide, label the account figure as an upper bound on reachable platform users, and never present it as the size of the hireable workforce.
Does a missing country in the Innovation Graph mean zero developers?
No. The Innovation Graph withholds figures for any economy with fewer than 100 unique developers doing the relevant activity in the quarter. The developers and repositories files cover more than 215 economies, but most other files cover only about 110 to 180. A blank or absent row usually means the value is under the threshold. Mark it as suppressed, and if you need a value, report it as fewer than 100.
Which source should I use for a country-level developer talent pool?
Use the Innovation Graph for country-level and language-level comparisons across quarters, because it is quarterly, has history back to Q1 2020, and uses one consistent public file. Use Octoverse only when the audience already knows its headline, and footnote its profile-based location rule. Then triangulate either one against an official occupation count, and use a named-profile index for ratios and actual people.
Try it on the search you came here for
Stop building boolean strings. Just describe the person.
Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.
01Describe them
One plain sentence. Role, city, stack, stage, whatever matters to you.
02I read the web live
GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.
03You read the shortlist
Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.
- Staff backend engineers in NYC who shipped Rust in production
- Series A fintechs in SF under 50 people, growing headcount this year
- Maintainers of fast-growing Rust web frameworks on GitHub
- No boolean, no filters, no seat to buy. One box.
- Read at search time, so a profile updated yesterday counts today.
- Every step visible as it runs, every name with its reason.
500 free credits on sign-up. No card, no demo call. See real searches.