David Sacks vs Anthropic: The Fight Over Open-Source AI Labels
David Sacks accuses Anthropic of using IP theft claims to ban Chinese open-source AI models and harm US developers. Here's what the debate means for AI builders.
David Sacks, co-chair of the President's Council of Advisors on Science and Technology, went on the All-In Podcast to accuse Anthropic of using intellectual property theft claims as cover to restrict open-source AI competition. Sacks warned that branding Chinese-origin open-source models as tainted would "put a dagger through the heart of the entire American open source ecosystem." His comments land as more than 20 companies, including Meta, Microsoft, and Hugging Face, are already lobbying Washington to keep open-weight models accessible.
What happened
| Detail | Fact |
|---|---|
| Date | July 27, 2026 |
| Forum | All-In Podcast |
| Sacks role | Co-chair, President’s Council of Advisors on Science and Technology |
| Target of criticism | Anthropic, maker of Claude |
| Open-source letter signatories | More than 20 companies and organizations |
| Model cited as example | Kimi K2.5 (Chinese open-source) |
Speaking on the All-In Podcast, Sacks argued that Anthropic’s push to restrict Chinese-origin open-source AI models is less about national security and more about killing off competition. His direct quote: the restrictions would “put a dagger through the heart of the entire American open source ecosystem,” and Anthropic “doesn’t want to have the competition.”
Sacks pointed to two concrete counterexamples. Mira Murati’s startup Thinking Machines and AI coding tool Cursor both used the Chinese open-source model Kimi K2.5, then trained it further with their own proprietary data to build distinct products. Under Anthropic’s proposed framing, that workflow would be suspect.
Why the “country of origin” argument falls apart
Sacks made a technical point that often gets lost in the policy debate. Once model weights are released publicly, any developer anywhere can download, modify, and retrain them on their own hardware. The weights carry no live connection to their original creator, and their country of origin has no technical bearing on what a US company builds with them.
He also drew a line between two practices that policymakers frequently mix up:
- Weight theft: Copying proprietary, non-public model weights without permission. Sacks agreed this is straightforward theft.
- Distillation: Using a model’s outputs to train another model. Sacks described this as standard industry practice, noting that both OpenAI and Anthropic have defended training on publicly available content in legal proceedings, including OpenAI’s defense in the New York Times copyright lawsuit.
Who else is pushing back?
Sacks is not alone. Nvidia CEO Jensen Huang recently endorsed a letter signed by more than 20 organizations urging Washington to support open-weight AI. Signatories include Meta, Microsoft, IBM, Palantir, Hugging Face, Mistral AI, and Perplexity. The letter argues that broad restrictions would slow innovation, weaken cybersecurity defenses, and cut off access for startups, researchers, and public institutions.
The coalition’s position: address IP concerns through targeted legal frameworks, not blanket bans. Sacks backed the same view, calling on companies to strengthen their own safeguards and enforce terms of service rather than lobby for sweeping rules. For more context on how Chinese AI models became a political flashpoint, see our earlier coverage on Kimi, DeepSeek, and the US debate over Chinese AI.
Our take
Anthropic’s position is convenient. The company benefits commercially if the open-source models that compete with Claude are restricted or stigmatized. That does not automatically make the IP concerns wrong, but it is a reason to read the argument carefully rather than take it at face value.
Sacks is right that “Chinese model” has become a fuzzy label. Kimi K2.5 being used as a base by US startups, then retrained with US data on US infrastructure, is not obviously a national security problem. The more honest debate is about distillation and whether model outputs should carry IP obligations, which is a legitimate open question the courts and Congress have not resolved.
For businesses building on open-source AI today, including teams exploring AI integration for their own products, the practical risk is regulatory uncertainty rather than an imminent ban. Watch how the letter-signing coalition’s lobbying plays out before making long-term bets on any single model lineage.
What to do about it
- Audit which open-source models your products or workflows depend on and note their origin.
- Review the license and terms of service for each model, particularly clauses about commercial use and distillation.
- Follow the open-weight letter coalition’s progress in Washington; new restrictions, if they come, will likely apply to commercial deployment first.
- Keep proprietary fine-tuning data documented so you can demonstrate the added value your team contributed, separate from the base weights.
The open-source AI policy fight is moving fast. Check the Lumien news feed for updates as lobbying and legislation develop.
Frequently asked questions
What is David Sacks' argument against restricting Chinese open-source AI models?
Sacks argues that once model weights are released publicly, they are no longer tied to their country of origin. US developers can legally download, modify, and retrain them on their own infrastructure, so labeling them as tainted would harm the American open-source AI ecosystem without a clear security benefit.
What is the difference between AI model weight theft and distillation?
Weight theft means copying proprietary, non-public model weights without permission. Distillation means using a model's outputs to train a separate model, which Sacks describes as standard industry practice that both OpenAI and Anthropic have themselves defended in legal contexts.
Which companies signed the open-weight AI letter endorsed by Jensen Huang?
More than 20 companies and organizations signed the letter, including Meta, Microsoft, IBM, Palantir, Hugging Face, Mistral AI, and Perplexity. The letter urges Washington to support open-source AI models rather than restrict them.
Did any US startups build products on Chinese open-source AI models?
Yes. According to David Sacks, both Mira Murati's startup Thinking Machines and the AI coding tool Cursor used the Chinese open-source model Kimi K2.5 as a base before further training it with their own proprietary data.