Rocky Linux's Founder Wants an Open AI Data Commons

Open-weights models are not open source, and everyone in the field knows it. You can download the weights, but you cannot inspect what produced them, cannot reproduce the build, and cannot audit whether anything in the corpus was licensed for the use you are about to make of it. OpenWALDO, announced August 11 by Gregory Kurtzer and sponsored by CIQ, is an attempt to fix the layer underneath: a shared, governed, auditable corpus of AI training data, plus the toolchain that turns it into models you can verify.

The acronym is the thesis. Open Weights, Artifacts, Licenses, Data, Origins. Every one of those is a thing the current generation of "open" models does not actually give you.

RelatedShopify swapped Redis for MySQL and hit $5.1M sales a minute

Who is behind it, and why that matters

Kurtzer's track record is the reason to take this seriously rather than file it with the other data-commons proposals. He founded CentOS, then Rocky Linux after Red Hat changed CentOS Stream's role, and he built Warewulf and Apptainer, the container runtime that most HPC clusters use. That is a career spent on a specific problem: how do you keep a shared artifact usable by everyone when the entity funding it has its own commercial interests?

Rocky Linux is the relevant precedent. It exists because a downstream rebuild that thousands of organizations depended on got repositioned out from under them, and the community response was to rebuild it under governance that could not be unilaterally changed. OpenWALDO applies the same instinct one layer up the stack. CIQ sponsors it, exactly as CIQ sponsors Rocky, and the governance is designed to sit outside the sponsor.

What is technically different here?

Three choices distinguish this from previous open-dataset efforts, and all three are lifted from software rather than invented for AI.

Git review governs meaning. Changes to the corpus go through the same public, attributable review process as a kernel patch. Not a dump with a README, an ongoing repository with history you can bisect.

Content-addressed storage carries the bytes. Data is referenced by hash, so a citation to a training example is exact and tamper-evident. If someone claims a model was trained on a given snapshot, that claim is checkable rather than a matter of trust.

DCO sign-off keeps responsibility attached. The Developer Certificate of Origin is the lightweight mechanism the Linux kernel uses so that every contributor asserts they have the right to contribute what they are contributing. Applied to training data, it converts "we scraped it and hope it was fine" into a signed, attributable claim with a name on it.

That last one is the genuinely interesting move, because it changes who carries legal exposure. Scraped corpora concentrate risk at the model trainer, who often cannot prove what went in. A DCO chain distributes and documents it.

How does it compare to what exists?

The comparison flatters OpenWALDO because it is describing intent against incumbents' delivered reality. That asymmetry is the honest caveat here, and it is the one to hold onto.

RelatedCloudflare Open-Sources an OS for AI Agents at Work

What could go wrong?

Corpus scale is the obvious problem. Frontier models train on quantities of text that took years of indiscriminate crawling to accumulate. A corpus where every item needs a licensed, signed, reviewed origin will start orders of magnitude smaller. Kurtzer's framing anticipates this: the shared corpus absorbs "duplicated foundational work" so that competition moves higher up the stack. That is a reasonable long game, but it means the first useful output is a good small corpus, not a GPT-scale one.

The second problem is that governance promises are cheap on launch day. Every sponsor-backed open project says the sponsor will not control it. Rocky Linux is genuine evidence that Kurtzer means it, and it is still evidence about a person rather than a structure. The thing to watch is whether an independent foundation with a real board materializes, or whether governance stays informal while CIQ pays the bills.

Third, contribution incentives are unresolved. Publishers and data holders have spent two years learning that their archives have licensing value. Asking them to contribute into a commons competes directly with selling the same material, and the announcement does not describe what they get back.

Our take

The most useful thing about OpenWALDO is that it names the right problem. The argument over whether Llama or Muse or Gemma counts as "open source" has been going in circles for two years, and it circles because everyone is arguing about the license on the weights while the actual opacity sits in the data. You cannot audit a model you cannot rebuild, and you cannot rebuild a model whose inputs were never recorded.

Applying kernel-development practice to that gap is not a flashy idea, and that is what recommends it. DCO sign-off, content addressing and public review are all boring, proven mechanisms that survived contact with a hostile world at scale. The hard part was never the technology. It is whether enough organizations decide that provenance is worth the friction of contributing under it. Rocky Linux suggests Kurtzer knows how to make that case. Whether the case is as compelling for data as it was for a free RHEL rebuild is the open question, and the honest answer today is that nobody knows yet.

Original analysis by GenZTech, based on the OpenWALDO project materials and the August 11, 2026 launch announcement. Source: GlobeNewswire.