This article was translated from its original French version with the help of AI. Read the original in French.
In this publication
On April 24, GitHub changed its privacy policy. By default, code from Copilot Free, Pro, and Pro+ users can now be used to train Microsoft’s AI models. There was no major announcement: a technical changelog, two or three articles in the trade press, and the machine moved on. Someone in GitHub’s San Francisco offices decided to switch from opt-in to opt-out and figured it could go through quietly. And it is working.
This update deserves a closer look. Not for what it says about respect for developers’ ownership rights (others have covered that), but for what it reveals about an asset the entire industry uses every day without giving it another thought: GitHub.
Agentic AI, systems that can reason, plan, and act in multiple steps instead of producing a single response, is built on code. An agent’s quality is measured in part by its ability to write and validate that code. From this perspective, control over the code corpus is control over the path toward agents. This makes GitHub structurally different from any other piece of tech infrastructure. And it is what almost never comes up in accounts of the AI war in 2026, where attention centers on models, benchmarks, and partnership announcements.
A misunderstood acquisition
When Microsoft announces its acquisition of GitHub in June 2018, the prevailing view is that it is a tactical PR move. Microsoft is consolidating the shift toward open source that began under Satya Nadella and offering developers a gesture of good faith. The community howls, some migrate to GitLab, the debate lasts two weeks, and then the subject falls off the radar.
Eight years later, that reading deserves another look. In a public appearance in 2024, Nadella says:
“GitHub Copilot is now a bigger business than all of GitHub was when we acquired it.”
In plain terms: in less than three years, a code completion assistant launched in 2021 surpassed the entire annual revenue of the company Microsoft acquired at the time: around $300 million in ARR when the deal was signed, for an acquisition price of $7.5 billion.
What was presented as a goodwill gesture looks, from today’s perspective, like a call option on the software industry, exercised at the dawn of generative AI. When the deal was signed, GitHub claimed around 28 million accounts. In October 2025, its Octoverse report claims 180 million, including 36 million added in a year: more than one new developer per second. The platform now hosts 630 million repositories and receives nearly a billion commits (code submissions) a year.
Microsoft did not buy a version control tool. Microsoft bought the digital identity of the people who write the world’s code. When the deal was signed, it was a PR move that also offered strong synergies with the Azure cloud offering. With GPT and its peers, it became a strategic asset in Big Tech’s AI war.
A corpus no one can recreate
Code is not like other text. To our knowledge, it is the only corpus that can be verified by execution: a program compiles or it does not, its tests pass or they do not. This property makes code models central to the path toward agents. An agent that can write code and validate it itself can, by extension, validate a growing share of its own actions. This is one of the reasons, debated but serious, why progress on SWE-Bench (the benchmark for evaluating coding models, developed at Princeton in 2023) tends to correlate with progress on general-purpose agent benchmarks.
Training these models requires vast amounts of code. The reference corpus is called The Stack (6 TB in v1, 67.5 TB in v2, more than 600 languages, produced by BigCode and hosted on Hugging Face). Its origins are public and documented: GitHub repositories. Most open-source code models have been trained on it, in part or in full.
The court filings in Doe v. GitHub, submitted in November 2022, made public a practice that neither Microsoft nor OpenAI now disputes: Copilot, GPT, and Codex were trained on public GitHub code. The court did not uphold the copyright infringement claims. The technical fact, however, is no longer disputed.
The April 24 terms of service update turns what was already a significant advantage into an asymmetric advantage. All Copilot interactions (prompts entered, suggestions accepted or rejected, context from open files) can now feed into the training of Microsoft’s AI models. This is the default for Free, Pro, and Pro+ accounts; Enterprise contracts remain protected.
OpenAI, Anthropic, and Google have no equivalent. They can reuse frozen datasets from earlier scraping (The Stack v2, RedPajama-code, and their derivatives), but GitHub has progressively tightened access to its API since 2023 (rate limits, restrictions on public calls, pricing terms renewed in March 2024), and collecting fresh data at scale is becoming increasingly difficult. Anthropic does have its own usage tracking through Claude Code, but in a different format and at a different order of magnitude: a few hundred thousand paying users (mostly independent developers) compared with the 26 million claimed by GitHub Copilot (many of them entire development teams).
At the scale needed to strengthen tomorrow’s agents for industrial code production, Microsoft is, in practice, the only player able to observe and capitalize on what happens at a developer’s keyboard.
A distribution channel no one is really challenging
At this point, an objection is inevitable. For eighteen months, the trade press has been repeating that Microsoft is losing the AI war on every front: models, IDEs, coding assistants, hardware. This hasty diagnosis misses what is happening out of sight.
The development tools gaining momentum in 2026 (Cursor, Claude Code, Devin) all rely on GitHub in one way or another. They authenticate through GitHub OAuth (the authentication mechanism that lets a third-party service connect to a GitHub account), read and write to GitHub repositories, trigger their continuous integration through GitHub Actions, and can run in Codespaces. Alternatives exist: Atlassian’s Bitbucket, GitLab at several major European enterprises, and self-hosted solutions such as Forgejo among advocates of digital sovereignty. But together, they account for only a marginal share of developers. For the vast majority of organizations using an AI coding assistant, no significant action runs without touching Microsoft infrastructure at some point.
Since 2024, GitHub Models has closed the loop on the economics. The marketplace offers models billed through GitHub, with Azure OpenAI providing the plumbing. Not every call will go through this chain: advanced agents will continue to call Anthropic’s or OpenAI’s APIs directly when the use case justifies it. But for a developer who regularly switches models, GitHub Models is the path of least resistance. As third-party tools connect to it, a growing share of calls ends up in a chain where Microsoft takes its margin.
Microsoft does not need to win the battle to be “the provider of the best model.” It just needs to hold on to the developers’ operating system. This strategic maneuver recalls the 1990s: Microsoft could lose the visible battles over applications as long as Windows remained the default layer.
A convenient silence
What stands out in the public debate about AI in 2026 is the almost complete absence of any mention of GitHub. There are two likely reasons.
The first has to do with communications. To function, GitHub must preserve its aura as an open, neutral platform. The day developers see GitHub as the strategic weapon it is, they will migrate to alternatives. By this logic, saying little about the appropriation is a prerequisite for the appropriation itself. The low-key announcements from GitHub in recent years (including the April 24 update) are no accident.
The second has to do with how analysts assess businesses. GitHub generates about two billion dollars in annual revenue (a figure Satya Nadella confirmed on the July 2024 earnings call), with no breakdown between Copilot and the rest of the platform. So that is $2 billion in a company that generates more than $245 billion: invisible on the income statement, but critical to the strategy. Wall Street does not know how to value plumbing. Competition authorities, for their part, think in terms of clearly defined markets: cloud, office applications, models. And GitHub fits none of those categories. The most powerful competitive advantage is often the one that eludes external analytical frameworks.
A blind spot for leadership teams
In 2025-2026, the European conversation about AI sovereignty has focused on models and data centers: Mistral, the announcements around “AI gigafactories,” the funding. These things matter. They are not enough, because none addresses the developer layer where the code that runs everything else is written, versioned, and distributed. This infrastructure will be all the more crucial in a world where anyone, and any agent, can become a developer in the course of a conversation.
For corporate leadership teams, the question is more immediate. Most organizations rigorously audit their infrastructure’s exposure. Very few audit their code’s exposure. The scale of that dependence is now comparable, and substitution may be harder: cloud migrations are routine, with well-developed tools and practices. GitHub migrations much less so.
Three concrete questions for an executive committee that wants to start seeing the situation more clearly:
- What share of our production code currently lives in repositories on Microsoft-owned GitHub, as opposed to self-hosted repositories or GitLab?
- What is our annualized Copilot cost, and how many internal users have become so dependent on it that going back would mean a measurable loss of productivity?
- If Microsoft decided tomorrow to double the price of Copilot Enterprise, or change the terms for data exports as it has just done for individual accounts, what is our six-month exit plan?
None of these three questions can be resolved in a week. All deserve to be asked before circumstances force us to ask them.
The real war is not the one we hear about
As long as we describe the AI war as a fight to build the “best model,” we will not see who is winning it. Models are becoming commodities, their inference prices are falling at a tremendous rate each year, and no lab’s technical lead lasts more than three months.
The only lasting assets seem to be those AI cannot reproduce itself: the distribution channel, the user contract, the authenticated identity, the usage data, the hardware. GitHub checks many of these boxes. An acquisition that looked like a tactical PR move in 2018 could, in 2030, look like the best strategic acquisition Microsoft has ever made.
This observation actually extends beyond Microsoft. While media attention centers on OpenAI, Anthropic, DeepSeek, and others, the tech giants are quietly securing positions that generate economic rents. Google with its 9 billion Search queries a day, Android (3 billion active devices), and vertical integration through TPUs that makes its inference costs unmatched. Amazon with AWS (60% of the enterprise cloud market) and Bedrock as a distribution layer for multiple models. Apple with 1.4 billion active iPhones and its unstoppable advance in inference with its unified-memory chips. Nvidia with its near-monopoly on current training hardware (more than 80% market share in frontier AI GPUs) and the CUDA ecosystem, for which no credible alternative has emerged in fifteen years. None of these assets can be bought. None can be reproduced in less than five years. They all predate generative AI and form the backbone on which it and its uses are built.
We are told a story of a war in which the insurgents are expected to take it all. In reality, it is more a war in which the established empires are repositioning themselves to tighten their grip on an economy ever more tied to technology. The real question over the next eighteen months is not who wins between OpenAI and Anthropic. Instead, it is: what does the economy of the AI era look like, divided among Microsoft, Google, Amazon, Apple, Meta, and Nvidia?
Sources and further reading
- GitHub Octoverse 2025
- GitHub acquisition announcement (Microsoft, June 2018)
- Microsoft Q2 FY26 earnings (January 2026)
- The Stack v1 and v2 (BigCode / Hugging Face)
- Doe v. GitHub (Joseph Saveri Law Firm)
- GitHub Privacy Statement update (March 25, 2026 changelog)
- The Register, InfoQ, Help Net Security on the terms of service changes
- Stratechery (Ben Thompson, August 2025 and February 2026)
- TechCrunch (July 2025)
- The Next Web (April 2026)
- SWE-Bench (Princeton NLP, 2023)
- Copilot Workspace documentation (GitHub, 2024)



