Tech leaders back open AI, but deeper openness needed—data access key

On Friday, Jensen Huang published a short letter on Twitter calling for American investment in "a strong, open ecosystem" around open-source artificial intelligence. Huang argued that open source had "created a shared foundation of knowledge on which generations of American engineers and entrepreneurs built their institutional sovereignty," and predicted similar benefits would flow from embracing open-source AI. The letter carried implicit political weight: the Trump administration had been signaling it would ban open-source software as a protectionist move to suppress Chinese language models, a policy that would have benefited Anthropic and OpenAI by eliminating open competitors. Recognizing this, CEOs from across the tech industry—all major and smaller firms except Anthropic—quickly signed on to Huang's letter on Friday.

The author of this post celebrates the development but argues Huang left out the most difficult and important piece. Building a competitive language model requires three inputs: people, compute, and data. On people, the author notes that although the promise of wealth has lured talent away from open-source work, industry backing should now stabilize recruitment. On compute, the author challenges the hyperscaler narrative that massive datacenters are an insurmountable moat: Chinese companies DeepSeek and Moonshot trained competitive models for reportedly under ten million dollars, and open-source development could be supported through compute-sharing agreements among tech firms, which would represent only a minuscule fraction of each firm's budget. The author predicts that once models are built openly, efficiency will follow naturally: "machine learning is a field that innovates through frictionless reproducibility," and openly available research, code, and data—evaluated through competitive benchmarking—rapidly improve systems. The author cites the example of high-quality ImageNet models, which went from trainable only at Google to buildable on a desktop in less than a year.

The hard part, the author argues, is data. Proprietary AI companies have trained large language models on vast quantities of copyrighted material: pirated libraries of books and academic papers, copyrighted imagery, collaborative knowledge bases like Wikipedia, volunteer forums like Reddit, all public code on GitHub, and transcripts of every video posted to YouTube. These companies have been found liable in court and have admitted this practice in their published papers, claiming it falls under fair use because training is transformative. The author accepts this legal reasoning but argues it cuts both ways: if closed-model training on copyrighted material is fair use, then open-source distillation of proprietary model outputs should also be fair use, and open-source models should have equal fair-use access to the same training corpora. This "open corpus" would require a fundamental rethinking of intellectual property law and copyright, but the author frames it as essential for competitive parity. The author concludes that while the challenges are large—recruiting talent, securing compute collaboration, and overhauling copyright—they are surmountable, and that the tech industry's support, combined with the brief history of open-source dominance before the proprietary 2020s, makes this the moment to "open things up again."

Jensen Huang's Friday letter represents a major shift: the Trump administration's protectionist stance against open-source software unexpectedly unified nearly the entire tech industry (with Anthropic as the sole holdout) in support of open AI—a reversal that signals both geopolitical anxiety about Chinese AI competitiveness and recognition that openness itself drives innovation. The author endorses this momentum but argues it is incomplete. The letter addresses open weights and open source, but sidesteps the data problem: proprietary AI labs have trained models on vast quantities of copyrighted material—books, academic papers, code repositories, videos—often without explicit permission, citing transformative fair use as legal cover. Open-source developers, by contrast, lack comparable access to these datasets, creating an asymmetry that prevents fair competition. The author frames this as a moral and strategic question: if closed models can legally use pirated and copyrighted material for training, why should open models not have the same fair-use right, especially in service of national technological sovereignty against China? The author acknowledges three hard challenges—recruiting talent away from private AI companies, securing compute infrastructure through industrial collaboration (citing Chinese examples like DeepSeek and Moonshot that trained models for under ten million dollars), and overhauling copyright law—but argues all are surmountable given industry consensus and the precedent of open-source innovation in earlier decades.

No comments yet. Be the first to share your thoughts!