Washington has spent two years debating how to answer China’s free-to-download AI models, and a new think-tank plan puts antitrust law at the center of the fight. The Manhattan Institute published “An American Answer to China’s Open AI Models: A Safe Harbor for Open Weights,” and the group sums up the core bargain in one line: “A trade: Labs that claim the safe harbor from antitrust would commit to releasing capable open-weight versions of their models on a predictable schedule, once those models pass independent risk evaluations.”
The idea is a safe harbor for open weights that rewards American labs for two things at once. They cooperate on security, and they keep a steady pipeline of downloadable models flowing to developers. Nothing about it is law yet. The cited coverage confirms no enacted bill, no named sponsor and no effective date, so this story is about a proposal and the pressure behind it. It lands in a week when 29 countries signing WAICO while Pax Silica rivals it showed how fast AI is splitting into rival blocs.
This analysis covers what the plan says, what it leaves unanswered, who could qualify, and what it could change for developers and enterprise buyers. It also separates confirmed facts from inference, because several details circulating online remain unconfirmed by the outlets cited here.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the Manhattan Institute Proposed
The headline of the proposal is its own summary. The Manhattan Institute wants Congress to create an antitrust safe harbor for AI laboratories that cooperate on security while committing to release capable open-weight models on a predictable schedule. Each release would follow an independent risk evaluation. The group framed the plan on social media with a short announcement: “We have a plan to ensure America dominates the open-weight tier of Artificial Intelligence.“
Read closely, the proposal is a trade rather than a mandate. Washington would not order any lab to publish its weights. Instead, it would offer a legal benefit (protection from antitrust risk when labs coordinate on security) and ask for a public benefit in return (open models on a schedule developers can plan around). That structure matters because it avoids the hardest legal fight, which is forcing private companies to give away work they paid billions to build.
The people most visible in the early conversation include Brandon Fuller, Arpit Gupta and Sean Speer, all of whom appear in the coverage and social posts tied to the launch. The proposal also names a set of companies as relevant players, including Nvidia, Microsoft and Meta, along with DeepSeek as the foil. Those names show who the authors think matters on each side of the debate.
How the Antitrust Safe Harbor Would Work
An antitrust safe harbor is a legal shelter. Companies that follow a defined set of rules gain protection from certain competition-law claims, even when their cooperation might otherwise look like coordination between rivals. Security work is the classic example. Labs that want to share threat intelligence, red-team findings or evaluation methods can worry that talking to a competitor invites scrutiny. A safe harbor removes that worry, but only for labs that accept the conditions attached.
In this proposal the condition is open weights. A lab claiming protection would commit to publishing capable open-weight versions of its models on a predictable schedule. Release would come only after the model passes independent risk evaluations. The table below separates what the cited coverage confirms from what remains open.
| Element | What the cited coverage says | Status |
|---|---|---|
| Legal tool | Antitrust safe harbor for AI labs | Proposed |
| What labs give | Cooperation on security and capable open-weight releases | Proposed |
| Release timing | Predictable schedule | Proposed, schedule not specified |
| Release gate | Independent risk evaluations before publication | Proposed, evaluator not confirmed in the fact sheet |
| Legislation and sponsors | No specific bill confirmed | Unconfirmed |
| Adoption or enactment | No formal adoption reported | Unconfirmed |
Two gaps stand out. The proposal as reported does not publish a calendar, so “predictable” could mean quarterly, twice a year or tied to each frontier launch. It also leaves open who runs the independent evaluations and what a failing score means. Those details will decide whether the safe harbor works as a growth engine for open models or as a slow-moving gate.
Open Weights Versus Open Source
The coverage defines open-weight models in practical terms. Users can download them, run them on their own machines and modify them. That differs from classic open source, where the training code, the data and the full recipe are typically public. With open weights, a developer gets the finished artifact and the freedom to adapt it, but not necessarily the ingredients.
The distinction shapes the policy debate. Stanford HAI has published an essay arguing that open-weight models are not enough and that science and society need truly open source AI. The Manhattan Institute plan targets the weaker, more achievable tier. Labs can keep their data and training methods private while still shipping weights that others can inspect, fine-tune and deploy. For a safe harbor that needs corporate buy-in, that is the realistic ask.
| Model type | Download and run locally | Modify and retrain | Training data and code public |
|---|---|---|---|
| Closed API model | No | No | No |
| Open-weight model | Yes | Yes | Not required |
| Open-source model | Yes | Yes | Yes |
That table is a framework drawn from the coverage’s definition, not a count of any particular company’s practice. Where a given lab lands depends on its license and what it chooses to publish.
Why DeepSeek Sits at the Center of the Debate
The cited coverage states that DeepSeek published model weights that users can download, run on their own hardware, inspect and retrain for specialized tasks. That choice turned a Chinese lab into an unexpected supplier of AI infrastructure for developers worldwide. A startup in any country can take those weights, run them on its own servers and build a product without sending a single prompt to a foreign cloud.
DeepSeek’s founder has framed the choice as strategy rather than charity. In an interview quoted by the Cato Institute, Liang Wenfeng, founder of DeepSeek, said: “We wonât go closed-source. We believe that establishing a robust technology ecosystem matters more.” That reasoning explains why open weights work as a geopolitical tool. A lab that wins developers wins an ecosystem, and ecosystems outlast any single model.
Beijing has also made clear it will defend its firms when Washington pushes back. Mao Ning, spokesperson for China’s Ministry of Foreign Affairs, responded to questions about U.S. restrictions involving DeepSeek by saying that “China will resolutely safeguard the legitimate rights and interests of Chinese companies.” The exchange shows why bans alone look weak. When the product is a file anyone can download, blocking a company does not block the model.
Moonshot and the wider Chinese field
DeepSeek is not alone. Our earlier report on Moonshot’s Kimi K3.1 Pro and its open-weight plans showed how many Chinese labs now treat open weights as a default distribution channel. The Manhattan Institute plan reads as a direct response to that pattern. It asks whether American labs can match the cadence without giving up their commercial edge.
The Security Problem the Safe Harbor Tries to Solve
Security cooperation is the quiet half of the plan. American labs face the same threats, from model theft to misuse by hostile actors, yet each fears that sharing too much with a rival will draw antitrust attention. Arpit Gupta (@arpitrage) put the tension as a question: “How do we best balance security needs for AI labs with broader economic goals of containing Chinese open source models and diffusing AI?”
Gupta’s framing captures the dilemma. Security argues for tighter control and fewer releases. Economic strategy argues for broad diffusion of American-built models so that developers do not default to Chinese ones. The safe harbor attempts to square the two by letting labs coordinate on defense while committing to publish models that clear an independent safety screen.
Independent evaluation is the linchpin. The recent debate over withheld models, including Anthropic’s decision to hold back Mythos after finding 10,000+ flaws, shows labs already judge some systems too risky to ship. A scheduled-release rule only works if an outside party can say when a model is safe enough, and if a lab can defer without losing its legal protection. The fact sheet does not confirm how the plan handles that deferral, so treat any specific mechanism as unverified.
Which American Labs Could Qualify
The cited coverage names Nvidia, Microsoft and Meta among the organizations relevant to the debate. It does not confirm that any of them has asked for, endorsed or agreed to join a safe harbor. Treat that as an open question. A company with a strong closed-model business has less reason to opt in than one whose revenue depends on selling chips or cloud capacity to everyone who runs a model.
Nvidia fits the second category. More open models means more demand for the hardware that runs them. Our analysis of Muse Spark versus Beam and DeepSeek in the local LLM field shows how crowded the open tier already looks, which makes a predictable American cadence more valuable to buyers than another one-off release.
| Organization | Role in the story | Confirmed position on the safe harbor |
|---|---|---|
| Manhattan Institute | Authored the proposal | Proposes it |
| DeepSeek | Chinese lab that published downloadable weights | Not a party |
| Nvidia | Named in the coverage | Unconfirmed |
| Microsoft | Named in the coverage | Unconfirmed |
| Meta | Named in the coverage | Unconfirmed |
The blank spaces in that table are the real news. Until a lab says publicly that it would trade a release schedule for legal cover, the proposal remains a thought experiment with a clear logic and no declared customers.
Market Impact for Developers and Enterprise Buyers
Developers care about predictability more than ideology. A team building on an open model wants to know that updates will keep coming, that licenses will stay stable and that the supplier will not vanish. A promised schedule would turn American open weights from sporadic gifts into a roadmap. That shift could pull enterprise teams toward domestic models, especially in sectors that worry about compliance exposure from foreign AI.
Pricing pressure is the second effect. When capable weights are free to download, hosted API vendors compete with the cost of running the model yourself. Our comparison of DeepSeek and Claude per million tokens already shows how wide the cost gap can get. A steady flow of American open models would narrow the gap on the supply side and push closed vendors to justify their premiums with quality, tooling and support.
Hardware is the third effect. Running capable models locally depends on memory and bandwidth, as the story on AMD’s Ryzen AI Halo running 300B-class models illustrates. More open weights from U.S. labs would raise demand for that class of local hardware, which benefits chipmakers and workstation vendors regardless of which model wins.
Corporate buyers and the ban question
The compliance angle is already visible in the trade press. A TechTimes piece from July argued, by its headline, that open weights make a ban on Chinese AI in corporate America hard to enforce. The National Law Review ran a guide on choosing between U.S. and Chinese AI models in light of export-control risks on both sides. Corporate counsel now read model provenance the way they once read software supply chains.
Competitive Comparison: Policy Options Against Chinese Open Models
The safe harbor is one of several levers Washington could pull. The table below compares them on the dimensions that matter most. It reflects our analysis of the trade-offs, not an official ranking, and none of the entries carries a confirmed cost estimate.
| Option | Main strength | Main weakness |
|---|---|---|
| Antitrust safe harbor for open weights | Rewards releases without forcing them | Depends on labs choosing to opt in |
| Bans on Chinese models in government and corporate use | Clear rule, easy to explain | Downloaded weights are hard to police |
| Chip and export controls | Targets hardware supply | Does not touch models already released |
| Direct public funding of open models | Gives the government control over goals | Requires appropriations and slow procurement |
| Do nothing and rely on the market | No policy risk | Leaves the open tier to foreign suppliers |
The safe harbor’s advantage is political. It costs the Treasury little and avoids the headline fight over compelled disclosure. Its weakness is behavioral. If the legal shelter is not worth much to a given lab, nobody opts in, and the plan produces nothing.
Historical Context: From Chip Controls to Open Weights
U.S. technology policy toward China has run through hardware for years, with export controls on advanced chips as the main tool. Open weights change the terrain. Once a model is published, its capabilities travel with the file, and no customs rule stops a download. The conversation has therefore moved from keeping technology out of Chinese hands to winning the contest for developer adoption.
Antitrust safe harbors have a long record in American policy. Congress has used research-consortium protections before to let rival firms collaborate on precompetitive work, with the National Cooperative Research Act as the best-known example. The Manhattan Institute proposal borrows that logic and attaches a public-interest condition. Rivals may coordinate on security, but they owe the public a stream of open models.
The Business Times framed the strategic stakes in a recent piece asking whether China can win over the world with open-weight models. A U.S.-China Economic and Security Review Commission paper published this year examines how China’s open AI strategy reinforces its industrial base. Fortune, for its part, has warned corporate leaders that most AI strategies sit in what it calls a death zone as Chinese open models spread. Together, these pieces show the argument has moved from labs to boardrooms and commission hearings.
Risks and Unanswered Questions
Every serious critique starts with safety. Publishing weights is irreversible. Once a model is out, no one can recall it, patch it or restrict how people use it. Independent evaluation helps, but evaluators can miss risks, and a lab that fears a failing score could slow-walk its schedule. The proposal says release follows evaluation, so the quality of that evaluation sets the ceiling on safety.
The second question is competitive. A commitment to release capable open versions on a schedule could hand rivals, including Chinese labs, a free look at American techniques. Weights reveal less than training data and code, but they still reveal something. Labs will weigh that exposure against the value of legal cover.
The third is legal. The proposal needs Congress, and the cited coverage confirms no bill, no sponsor and no hearing date. Antitrust exemptions also draw opposition from consumer advocates who worry that protected cooperation can drift into price or market coordination. Any final text would need guardrails that keep the shelter narrow and tied to security work.
Five Predictions for the Open-Weight Policy Fight
These are our judgments based on the reporting above, not confirmed events.
- The idea gets borrowed before it gets legislated. Expect lawmakers and agency officials to cite the safe-harbor concept in speeches and white papers well before any bill carries its text.
- At least one U.S. lab will publicly test the logic. A company with a hardware or cloud business model has the strongest incentive to say it would trade a release schedule for legal protection.
- The evaluator question becomes the main fight. Who runs independent risk tests, and who can overrule them, will draw more debate than the antitrust language itself.
- Chinese labs keep the release cadence high. Open weights are a core strategy for DeepSeek and its peers, so a slower pace is unlikely in the near term.
- Enterprise buyers add model provenance to procurement. Legal and security teams will ask where weights came from, which license applies and which jurisdiction’s rules bind the supplier.
What to Watch Next
Three signals will show whether the plan has legs. First, a named legislative sponsor would turn an essay into a vehicle. Second, a public statement from a major lab would show real demand for the shelter. Third, a published draft of the evaluation process would reveal whether the gate is workable or a bottleneck.
Readers tracking the broader AI beat can follow related coverage in our AI and machine learning section. For now, the key fact holds. The proposal is public, the logic is coherent, and the legal status remains unconfirmed.
Frequently Asked Questions
What is the safe harbor for open weights?
It is a Manhattan Institute proposal to give AI labs antitrust protection when they cooperate on security. In return, labs would commit to releasing capable open-weight models on a predictable schedule after independent risk evaluations.
Has the proposal become law?
No. The cited coverage confirms no enactment and no formal adoption. It is a policy proposal that would need Congress to act.
What are open-weight AI models?
They are models whose trained weights users can download, run on their own machines and modify. They differ from open-source models, which typically also publish training code and data.
Why does DeepSeek matter to this debate?
DeepSeek published weights that users can download, inspect, run on their own hardware and retrain for specialized work. That made a Chinese lab a global supplier of AI building blocks and put pressure on U.S. policymakers to respond.
Would labs be forced to release their models?
No. The plan is a trade. Labs that want the antitrust shelter would accept the release commitment, and labs that decline would simply not receive the protection.
Which companies support the plan?
The cited coverage names Nvidia, Microsoft and Meta in the surrounding debate, but it does not confirm that any of them has endorsed the safe harbor.
How would independent risk evaluations work?
The proposal says models would pass independent risk evaluations before release. Details on who runs them, the scoring method and the handling of failures are not confirmed in the sources used here.
