GitHub at scale · August 2026
Everything gets built.
Nothing gets found.
GitHub is the largest pile of built software in history and it has one way to tell you whether anyone cared. Almost nothing clears it.
Repositories created since this page loaded
0
Modeled at 230 per minute, GitHub's reported 2025 creation rate
The builder
The builder is no longer
a developer.
Thirty-six million people joined GitHub last year, and most of them had an AI assistant open in their first week.
180M
Developers on GitHub
36M
Added in 2025 alone
1/sec
Rate of new developers joining
~80%
Using Copilot in week one
- One new developer joined GitHub every second last year. Fastest absolute growth in its history.
- Roughly 80% of them were using an AI coding assistant inside their first week.
- The gate on whether something gets built at all is gone. Building is no longer the hard part.
The detail+
Correction to a circulating figure
A common secondhand framing splits GitHub into 150 million registered accounts and a separate, looser 180 million developer ecosystem. That distinction does not exist. 150 million is an older snapshot of the same metric.
The caveat that is real: 180 million accounts is not 180 million working humans. Bots, organizations, duplicates and dormant accounts are all inside it.
The output
A fifth of everything
arrived last year.
Repository creation is not growing steadily. It is compounding.
630M
Total repositories hosted
395M
Public repositories
121M
Created in 2025 alone
230
Created every minute
The detail+
Correction to a circulating figure
A common split claims 420 million current repositories and 630 million that have existed over time. That split is not in the source. 630 million is the current hosted count, and 395 million of those are public, which is exactly the 63% figure quoted alongside it.
It matters because every downstream calculation about visibility is built on that base.
The void
Almost nothing
gets a single star.
Every dot below is one 3,950th of every repository on GitHub. Change the threshold and watch the field go dark.
149,465,051
Repositories with at least one star, out of roughly 1.5 billion in the dataset. About 90% have never received a single star from anyone, and 98% have fewer than two.
Where these numbers come from+
This is the one section on the page where the underlying figure is checkable rather than estimated. The counts below were pulled from the public ClickHouse dataset of GitHub events, which anyone can query directly. They surfaced in a Hacker News thread in 2026 and have been reproduced since.
The caveat that matters
This dataset counts every repository ever seen in GitHub's public event stream, including deleted ones and forks, so its denominator is around 1.5 billion rather than the 630 million currently hosted. The shares here are internally consistent but should not be mixed with the Octoverse counts elsewhere on this page.
You will also see the zero-star share quoted as high as 99% in secondhand summaries. The defensible number is closer to 90%, with 98% holding fewer than two stars. Lower than the folklore, and far more checkable.
One consequence worth sitting with: roughly 150 stars puts a project in the top 1% of everything that has ever been starred at all.
The scoreboard
And the one signal
there is, is rigged.
A star is free, takes one click, and at the top of the distribution a meaningful share of them were bought.
6M
Suspected fake stars found
18,617
Repositories involved
16.66%
Of repos with 50+ stars, in a campaign
~0%
Same figure before 2022
The detail+
A peer-reviewed study out of Carnegie Mellon, NC State and Socket, presented at ICSE 2026, scanned 20 terabytes of GitHub metadata covering 6.7 billion events and 326 million stars.
The manipulation is not evenly spread. It concentrates exactly where visibility converts into something, at the top of the distribution. The incentive is not mysterious either: venture firms use star counts as a sourcing signal, so stars became worth buying.
GitHub appears to agree with the findings. Roughly 90% of the flagged repositories had been deleted from the platform within months.
In the wild
People are already building
the missing layer.
Not as a complaint. As shipped code. Every one of these exists because GitHub does not do the job on its own.
Discovery
You cannot find what you need
Skills, agents and harnesses are scattered across thousands of unconnected repos. Search returns repo names, not capabilities. Index projects describe finding the right component for a task as painful, and that is the whole reason they exist.
Onboarding
The page is built for engineers
A link sends a curious person to a source tree. The newcomer wants to know what it does, what it looks like, and whether they want it. The repo answers none of those questions first.
Trust
No way to tell good from dangerous
Skills execute code, so the official guidance is to stick to trusted sources. The only trust signal available is stars, which are free, one-click and demonstrably purchasable. One index resorts to a two-star minimum as a quality filter.
Delivery
Sharing means sending someone to clone
There is no link that just gives a person the thing. Clone, fork and pull are where non-developers stall out, and the stalling is well documented going back a decade.
Payment
The stack is stitched together by hand
The working playbook is a landing page, a Stripe checkout, and a private repo or a zip file. It works, which is the point. People are earning real money through infrastructure they assembled themselves.
Audience
Nothing you build compounds
A repo has watchers and stars. It has no subscribers, no way to notify anyone of the next thing, and no relationship that carries from one project to the next.
The workarounds people have shipped+
A short roster of things built specifically to patch what GitHub does not provide. None of them are the same shape, none of them talk to each other, and all of them are downstream of repos they do not control.
Why this is the strongest signal on the page
A complaint is an opinion. A shipped workaround is a revealed preference, and there are a lot of them. The demand is proven, the willingness to pay is proven, and every current answer is partial.
The pattern is familiar. Before Substack there were newsletter writers duct-taping MailChimp to Stripe to WordPress. The tooling existed. The place did not.
What people actually say+
In 2026 a dashboard claiming most AI-assisted code lands in repos with fewer than two stars hit the Hacker News front page. It drew 337 points and 222 comments, and the thread turned into an unplanned referendum on whether a star means anything. A selection, all from working developers:
Several described stars as a private bookmarking habit rather than a public signal, one noted the first thing they do with a new repo is star it themselves, and a maintainer with 40,000 stars across their projects said most of that code was written as a teenager and is bad, and that the stars came from timing rather than merit.
Another described star-buying as a standard open-core playbook move: publish, buy early stars so the project looks alive, then sell the hosted version. A separate commenter confirmed the demand side from the receiving end, saying investors ask about star counts because of popularity.
The comment that matters most for anyone building an alternative came from a developer who had spent weeks building a personal knowledge system with Claude Code and said of the whole debate: "The framing assumes github repos are supposed to be products."
That is the gap stated by someone who is living in it. The repo is a versioning tool being asked to do a job it was never designed for, and the star is a bookmark being asked to serve as a scoreboard, a quality signal, a funding filter and an audience all at once. Elsewhere in the thread, one commenter recounted meeting someone who had built twenty projects in four weeks. Asked how many people used any of them, the answer was just him.
The byproduct
What happens when
nothing is found.
Money is downstream of visibility, not a separate problem. Here is what the platform's own payment layer has moved in its entire existence.
49K
Developers ever funded
$50M
Cumulative payouts, all time
0.03%
Of developers on the platform
1 in 3,700
Odds of ever being paid
The detail+
The top bar is drawn to scale. It is 0.5% of the one below it. GitHub has roughly twenty-five times the users and has moved about one two-hundredth of the money.
GitHub Sponsors is not a bad product. It has been live since 2019 and personal accounts keep 100% of what comes in. The constraint is upstream of the payment rails: you cannot get paid for something nobody found.
What the numbers say
The problem was
never supply.
- 121 million repositories landed in a single year, and 36 million people showed up to make them.
- Most of those people had an AI assistant open on day one. Building is solved.
- Underneath all of it sits one signal: free, one-click, partly counterfeit, and worth nothing.
- Somewhere between 95% and 99% of public repositories have never received even that.
Created while you read this page
0
On current distribution, almost none of them will ever be found by anyone
Stars are not
an audience.
Sources and method
- GitHub Octoverse 2025. 180M+ developers, 36M added in 2025, 630M total repositories, 395M public (+19% YoY), 121M created in 2025, 230 repos per minute, 63% public, 81.5% of contributions private, ~80% Copilot adoption in week one. Reporting window September 2024 to August 2025.
- GitHub Sponsors. ~49,000 developers funded as of March 2026, $50M+ cumulative payouts across 110+ countries, 0% fee on personal accounts.
- StarScout, ICSE 2026. Carnegie Mellon, NC State and Socket. 20TB of metadata, 6.7B events and 326M stars from 2019 to 2024. ~6M suspected fake stars across 18,617 repositories from ~301,000 accounts. 16.66% of repos with 50+ stars implicated by July 2024. ~90% of flagged repos later deleted.
- Star distribution. Not published by GitHub. Counts pulled from the public ClickHouse dataset of GitHub events (play.clickhouse.com), surfaced and cross-checked in a 2026 Hacker News thread. That dataset's denominator (~1.5 billion repos, including deleted and forked) differs from Octoverse's 630 million currently hosted, so its shares are internally consistent but not interchangeable with the counts above.
- Hacker News, "90% of Claude-linked output going to GitHub repos w <2 stars" (2026, 337 points, 222 comments). Source of the practitioner quotes and the star-count ladder.
- Kajabi. $10B+ in creator revenue enabled, per company reporting.
- Ecosystem evidence. Skill index and installer projects (Clawd, SkillsMP, Skill Retriever), paid marketplaces (Agensi, Agent37), Anthropic's own skills guidance on trusted sources, the AAIF working group notes on skill discovery over MCP, and Andy Lester's talk on GitHub and end users. Practitioner revenue figures are self-reported.