{"slug": "who-pays-for-clang", "title": "Who pays for Clang", "summary": "Google and Apple's combined share of Clang commits fell from 84% in 2010 to 16% in 2026, according to a commit-attribution analysis by an unnamed Clang contributor who used Claude Code to collect git data and compute the tables. GPU and accelerator vendors now write more of the compiler: NVIDIA's share reached 8% in 2026, mostly through CIR, and AMD's reached 11%, mostly AMDGPU, HIP, and offloading work, while Red Hat's share is mostly Timm Bäder's bytecode constant expression interpreter. The author cautions the method is unscientific, since commit counts are a crude proxy for work and a few points of difference fall within the error.", "body_md": "Clang's funders have changed since 2010: Apple and Google, which wrote most of it until the mid-2010s, now write a small share, and GPU and accelerator vendors, Red Hat, and new sponsors of C++ conformance work write more.\nThis post counts commits, pull-request approvals, and the kind of code each employer touches.\nThe method is unscientific: commit counts are a crude proxy for work, and the attribution is my best guess, so read the numbers as rough trends, not an authoritative record.\nIt grew out of [ClangIR and the cost of MLIR](https://maskray.me/blog/clangir-and-the-cost-of-mlir), which needed it to explain who wanted an MLIR-based Clang.\n\nLike that post, this one relies on an LLM agent (Claude Code), which collected the git data, looked up affiliations, and computed the tables; I checked and edited the results.\n\nCommits under `clang/` by who paid for the work, usually the author's employer at the time.\n\"Top ten\" is the share of commits by the ten most active authors:\n\n| Year | Commits | Authors | Top ten |  | Apple | Intel | Red Hat | NVIDIA | AMD | Arm | Microsoft | other companies | universities and labs | individuals or unknown | \n|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| 2010 | 8,191 | 72 | 77% | 5% | 79% | 0% | 0% | 0% | 0% | 0% | 0% | 3% | 6% | 7% | \n| 2012 | 7,545 | 135 | 55% | 23% | 54% | 0% | 0% | 0% | 0% | 0% | 0% | 7% | 3% | 12% | \n| 2014 | 6,347 | 200 | 45% | 43% | 18% | 3% | 0% | 0% | 0% | 2% | 0% | 18% | 1% | 15% | \n| 2016 | 5,253 | 305 | 30% | 38% | 21% | 5% | 0% | 0% | 2% | 2% | 0% | 15% | 1% | 17% | \n| 2018 | 4,973 | 351 | 32% | 30% | 19% | 8% | 0% | 0% | 2% | 2% | 0% | 17% | 4% | 18% | \n| 2020 | 5,523 | 551 | 25% | 29% | 10% | 4% | 1% | 0% | 5% | 7% | 0% | 19% | 2% | 22% | \n| 2022 | 6,193 | 677 | 27% | 21% | 7% | 8% | 6% | 0% | 3% | 4% | 2% | 16% | 6% | 27% | \n| 2024 | 7,330 | 882 | 31% | 12% | 6% | 5% | 12% | 4% | 6% | 4% | 3% | 19% | 1% | 28% | \n| 2026 (to October) | 7,160 | 1,024 | 25% | 7% | 9% | 4% | 8% | 11% | 10% | 3% | 3% | 12% | 3% | 30% | \n\nI took the employer from the commit email when it is a company address. For the authors who used personal addresses, which before the 2019 move to GitHub was most of them, I looked up dated employment for about 150 of the most active ones from LinkedIn, Developers' Meeting speaker lists, paper affiliations, and GitHub profiles. Work done in spare time or unrelated to the employer counts as individual, such as Craig Topper's X86 work before Intel paid for it and my own Clang work while I was at Google. \"Individuals or unknown\" is independent contributors plus authors I could not place; people with ten or fewer commits each make up about half of it in 2024 and three-fifths in 2026, and the rest is a few prolific individuals such as Owen Pan (clang-format) and me. Before 2019, the recorded author is the committer, who often landed other people's patches, so the early author counts are low. None of this is rigorous: a commit can be a one-line fix or a month of work, public profiles are incomplete or out of date, employment dates are approximate, and a difference of a few points is within the error. Corrections are welcome.\n\nGoogle and Apple wrote 84% of Clang commits in 2010 and 61% in 2014, but 16% in 2026.\nApple built Clang for Objective-C and its platforms' toolchain; Google's priorities were large C++ code bases, compile time, and tooling.\nRed Hat's share is mostly Timm Bäder's bytecode constant expression interpreter.\nNVIDIA's is mostly CIR; outside CIR it wrote 3% of commits.\nIts Clang people joined from elsewhere: Erich Keane from Intel in August 2023, Andy Kaylor from Intel in January 2025, and Reid Kleckner from Google in December 2025.\nAMD's is mostly AMDGPU, HIP, and offloading work, and since 2026 also CIR (14% of its commits).\nIntel's share includes Aaron Ballman, Clang's lead maintainer; Intel's ICX compiler is built on Clang.\nIntel's commits are increasingly specific to a language mode or target (SYCL, OpenMP, CIR, OpenCL, X86): 22% in 2024, 58% in 2026.\nLanguage conformance work comes from new sponsors.\nBloomberg has sponsored Corentin Jabot's Clang work since 2022, and in 2024 also paid Vlad Serebrennikov, who \"[used to contract for Bloomberg](https://discourse.llvm.org/t/88927/9)\", and Matheus Izvekov, who [wrote papers for Bloomberg](https://github.com/cppalliance/wg21-sage/blob/master/inputs/matheus-izvekov.md); the three wrote 4.4% of 2024's commits, about the same as NVIDIA's 3.7%.\nThe C++ Alliance now employs Krystian Stasiowski and Matheus Izvekov.\n\nThe same count over `llvm/` shows that the shift is not specific to Clang.\nThe employment lookups covered Clang contributors, so more `llvm/` authors are unknown:\n\n| `llvm/` | Commits | Authors |  | Apple | Intel | Red Hat | NVIDIA | AMD | Arm | Microsoft | other companies | universities and labs | individuals or unknown | \n|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| 2010 | 13,245 | 97 | 5% | 67% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 2% | 26% | \n| 2014 | 12,423 | 273 | 26% | 19% | 1% | 0% | 0% | 7% | 3% | 0% | 20% | 2% | 22% | \n| 2018 | 14,737 | 494 | 18% | 7% | 11% | 1% | 0% | 4% | 5% | 0% | 22% | 0% | 31% | \n| 2022 | 18,538 | 946 | 15% | 6% | 2% | 7% | 0% | 10% | 6% | 1% | 21% | 2% | 31% | \n| 2026 (to October) | 18,539 | 1,561 | 10% | 7% | 2% | 3% | 7% | 20% | 7% | 1% | 15% | 2% | 28% | \n\nGoogle and Apple went from 72% of `llvm/` commits in 2010 to 17% in 2026, close to their 84% to 16% in `clang/`.\nWhat is particular to Clang is Red Hat (8% of `clang/` commits in 2026, 3% of `llvm/`) and the CIR part of NVIDIA's rise: NVIDIA grew in both, to 11% of `clang/` and 7% of `llvm/`.\nAMD is the largest funder of `llvm/`, at 20%.\n\nCommit counts undercount review, and cleanup commits inflate some authors: Kazu Hirata's mechanical cleanups were 158 of Google's `clang/` commits in 2024 and 326 in 2025 (24 in 2026).\nFor 2026 I also counted approvals on the 7,005 merged pull requests that touch `clang/`, each approver once per pull request (8,333 approvals).\nReviewers are attributed the same way, with the GitHub profile's current company as a fallback.\nAuthorship matches the 2026 commit shares above; \"CIR\" means the 1,285 pull requests whose title is tagged CIR:\n\n| 2026 pull requests |  | Apple | Intel | Red Hat | NVIDIA | AMD | Arm | Microsoft | other companies | universities and labs | individuals or unknown | \n|---|---|---|---|---|---|---|---|---|---|---|---|\n| approved | 8% | 13% | 7% | 2% | 17% | 10% | 3% | 6% | 21% | 3% | 10% | \n| authored, excluding CIR | 8% | 11% | 5% | 10% | 3% | 11% | 2% | 4% | 12% | 3% | 30% | \n| approved, excluding CIR | 10% | 17% | 8% | 2% | 7% | 11% | 3% | 8% | 22% | 3% | 10% | \n\nIndividuals write 30% of the pull requests but give 10% of the approvals; review is concentrated in fewer, mostly company-paid people. The ten most active approvers give 36% of the approvals, and 24 people give half, about as concentrated as commits were in 2016–2018. Outside CIR, the most active approvers are Corentin Jabot (Bloomberg), Aaron Ballman (Intel), Balázs Benics (Apple, the Static Analyzer), Gábor Horváth (Apple, mostly lifetime safety), Joseph Huber (AMD), Eli Friedman (Qualcomm), and Erich Keane (NVIDIA). Inside CIR, NVIDIA gives 57% of the approvals, mostly Andy Kaylor and Erich Keane.\n\nNot all of this work is shared. I split the 2026 commits by the code they touch: \"specific\" is one language mode (HLSL, OpenCL, CUDA/HIP, SYCL, OpenMP, OpenACC), CIR, one target, or one OS's driver; \"tools\" is the Static Analyzer, clang-format, Tooling, libclang, and clang-repl; \"core\" is the rest, such as Sema, AST, the parser, modules, and target-independent CodeGen. The majority of a commit's files decides, and its title tag breaks ties. The rule matters: other reasonable rules move the vendor cells by up to 8 points (Intel's ranges from 50% to 65%).\n\n| 2026 |  | Apple | Intel | Red Hat | NVIDIA | AMD | Arm | Microsoft | other companies | universities and labs | individuals or unknown | \n|---|---|---|---|---|---|---|---|---|---|---|---|\n| specific, of the employer's commits | 17% | 5% | 58% | 3% | 86% | 68% | 79% | 86% | 36% | 44% | 33% | \n| share of core commits | 11% | 13% | 3% | 17% | 3% | 6% | 1% | 1% | 11% | 3% | 32% | \n| share of approvals on core pull requests | 12% | 20% | 9% | 3% | 8% | 7% | 1% | 1% | 25% | 3% | 10% | \n\nMeasured by the code touched, NVIDIA's, AMD's, Arm's, and Microsoft's commits are 68–86% specific to a language mode or target. The shared core is written mostly by individuals, Google, Apple, and Red Hat, and six people give a third of its approvals.\n\nIn 2024, 23% of all `clang/` commits were specific; in 2026, 41%.\nCIR, which went from 14 specific commits to 1,336, accounts for more than the whole increase; other specific work fell slightly.\nVendor-specific does not mean niche: AArch64 and x86 work reaches most users, and AMDGPU and HLSL matter to their own users.\nThe difference is who reads the code and keeps it working.\nSpecific code is mostly maintained by the people who need it, and its cost to everyone else is building it and occasionally fixing a test; [CIR](https://maskray.me/blog/clangir-and-the-cost-of-mlir) is the case where that cost became large.", "url": "https://wpnews.pro/news/who-pays-for-clang", "canonical_source": "https://maskray.me/blog/who-pays-for-clang", "published_at": "2026-10-12 00:00:00+00:00", "updated_at": "2026-10-15 05:15:25.478559+00:00", "lang": "en", "topics": ["developer-tools", "ai-tools"], "entities": ["Clang", "Google", "Apple", "NVIDIA", "AMD", "Intel", "Red Hat", "Claude Code"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/who-pays-for-clang", "markdown": "https://wpnews.pro/news/who-pays-for-clang.md", "text": "https://wpnews.pro/news/who-pays-for-clang.txt", "jsonld": "https://wpnews.pro/news/who-pays-for-clang.jsonld"}}