Regulation rarely arrives with a dramatic press conference or a law that explicitly outlaws a new technology. More often, it happens quietly, through the slow accumulation of compliance requirements, safety standards, and oversight rules that sound entirely reasonable on paper.
In the world of artificial intelligence, a ban on open-source models would likely take a similar path. It wouldn't come as a law saying, “Open-source AI is illegal.” The more plausible route is a regulatory regime that requires advanced AI models to meet a set of strict safety, monitoring, and control standards before they can be released. On paper, those rules would apply equally to everyone. In practice, they would treat open and closed models completely differently.
A closed model sits behind an API. The company that made it can monitor usage, change its behaviour, restrict access, push updates or shut it down entirely. An open model can be downloaded and run independently. Once the weights are released, the original developer cannot recall every copy or control what everyone does with it. Apply rules that require those controls to every advanced model and open releases become much harder, or potentially impossible, without anyone ever passing an explicit ban.
This is not entirely hypothetical. The EU AI Act already treats advanced open models differently once they cross its threshold for “systemic risk.” Open models receive exemptions from some general-purpose AI requirements, but those exemptions disappear for models deemed sufficiently capable. Push that logic far enough and the frontier becomes closed by default.
The immediate winners would be obvious: the companies that already own the leading closed models. The losses would be spread across developers, businesses and consumers, which makes them much easier to overlook.
Open models keep closed models honest
There are only a handful of companies with the capital and compute required to train models at the frontier. That already makes AI an unusually concentrated market. Open models provide one of the few ways for everyone else to compete with companies spending tens of billions of dollars training them.
A startup can pay OpenAI, Anthropic or Google for access to a model. It can also download an open model, run it on rented GPUs, modify it and move between infrastructure providers. The startup may still choose the closed API, and plenty do. What matters is that it has somewhere else to go.
That outside option changes the economics of the market. If API prices rise, terms change or a provider restricts a use case, developers can move more of their workload to an open model. A company with sensitive data can run a model inside its own environment. A team can modify a model for a specific task instead of waiting for a provider to ship the feature it needs.
The U.S. government's 2025 AI Action Plan made a similar argument, noting that open-weight models give startups flexibility without making them dependent on a closed provider and allow companies or governments with sensitive data to run AI without sending that data to an outside model company.
Remove that option and the relationship changes. The major model companies stop competing against an alternative that anyone can own. Every startup building on AI becomes a customer of a small group of intelligence providers. Developers can still switch from one closed API to another, but that is a much weaker form of competition. The suppliers have similar economics, they control access, they set the price and they decide what can be built on their systems.
We have seen versions of this before.
IBM gave up control and created a market
When IBM launched its PC in 1981, it made an unusual decision. Instead of designing every component itself, IBM used off-the-shelf parts and published detailed technical information about the machine. That open architecture helped turn the IBM PC into an industry standard, but it also made it possible for other companies to build compatible computers.
Compaq became one of the first major successes, followed by a wave of other manufacturers. By the mid-1980s, inexpensive IBM-compatible machines were pushing PC prices down while expanding the market for software and peripherals. Consumers could buy computers from different manufacturers while still running the same applications.
IBM eventually lost control of the standard it created, but consumers benefited from the competition. The market shifted away from a world where one company had to own every part of the computer and toward one where companies competed on machines, components, software and price.
There is a useful parallel to AI. Open models let thousands of companies start with capable underlying intelligence and compete over what they do with it. Closing the model layer gives the owner of that intelligence much more control over everything built above it.
America tried restricting cryptography too
The encryption fights of the 1990s offer an even closer example. The U.S. government treated strong encryption software as a controlled export, which meant American technology companies often had to offer weaker encryption abroad or maintain separate domestic and international products.
The problem was quite simple: strong encryption already existed outside the United States. Foreign buyers could purchase it elsewhere. American companies argued that international customers would reject deliberately weakened products, leaving U.S. vendors with the choice of maintaining separate versions of their software or weakening security for everyone.
By the mid-1990s, members of Congress were pointing to hundreds of foreign encryption products already available and warning that export controls were hurting American software companies without preventing access to strong cryptography. The restrictions were eventually relaxed because limiting American companies did little to stop the technology from spreading.
That history matters because banning open AI in the West does not make open AI disappear. Chinese labs can still release models. Developers elsewhere can keep improving them. Researchers can still download them. Compute can move to jurisdictions where the rules are different. The restriction changes who is allowed to participate more than it changes whether the technology exists.
The same logic applies to capital. If companies can access cheaper or better AI outside the United States or Europe, more investment will eventually follow that access. A Canadian or American startup that cannot legally run a model at home has an incentive to put that workload somewhere it can. A multinational can place more engineering resources in that jurisdiction. A new company deciding where to incorporate has another reason to choose somewhere else.
You do not need to believe recursive self-improvement is around the corner for that economic argument to hold. Restricting a technology that remains available elsewhere can shift economic activity without removing the technology itself.
Consumers eventually pay for concentration
Most consumers will never download model weights, but they still benefit from the fact that someone can.
Consider what happens inside an AI application today. A developer can use a large closed model for difficult requests and a smaller open model for simpler ones. Certain workloads can run locally. Models can be modified for specific tasks. Workloads can move between cloud providers as model quality and prices change.
That competition lowers the cost of producing the application. Take away the open option and more of that application sits on top of someone else's meter. Every token has a supplier margin attached to it, every product is exposed to API price changes and every developer has to work within another company's rules. Those costs eventually reach the customer.
The Microsoft browser case is a useful reminder of how control over an important technical layer can shape competition. Microsoft did not need to make competing browsers illegal to weaken Netscape. Control over Windows gave it enormous influence over how users reached the internet, and U.S. courts later found that Microsoft's actions reduced consumer choice and harmed competition.
AI could develop a similar dependency at a deeper layer. If five companies become the only legal suppliers of sufficiently capable models, they become gatekeepers for thousands of products that have yet to be built.
Price is only one part of that cost. A law firm may prefer running a model within its own infrastructure. A bank may want complete control over where customer data goes. A startup may want to modify a model in a way its API provider does not support. Researchers may want direct access to weights so they can understand how a model behaves. Open models give each of them more control. A closed-only market leaves those decisions with the supplier.
The model is becoming a component
There is another reason restricting open models would be particularly strange right now. The model itself is becoming a smaller part of what makes an AI product useful.
We have written about this before at Venture's Edge. Take two developers using the exact same underlying model. One gives it a prompt and returns the answer. The other gives it memory, tools, proprietary context, permissions, software it can interact with, a way to check its own work and a record of what has worked before. The underlying intelligence may be identical, but the products can perform very differently.
This is becoming obvious in coding. Put the same model inside a basic chat interface and inside a well-designed coding agent and you can get dramatically different results. The improvement comes from the surrounding system: how context is assembled, how tasks are broken down, when tools are called, how results are checked and how the system decides what to do next.
As that surrounding layer gets better, benchmark leadership at the base-model level matters less to the end user. A slightly weaker model inside a much better system can beat a stronger model that is poorly integrated into the task. Cost matters too. A company may be able to use a cheaper open model for most requests, reserve a frontier closed model for the hardest ones and route between them based on the job.
That changes where companies can create durable value. Models can be swapped. The surrounding system accumulates the workflow, company knowledge, user context, tools and feedback that make the product useful. Open models accelerate that shift because developers have more freedom to choose the right model for each job.
A ban moves in the opposite direction. It reduces the number of models developers can choose from, increases their dependence on a handful of providers and raises the cost of experimenting at the application layer. It also protects the part of the AI stack that may be becoming less differentiated.
The West could end up protecting a small group of model companies while making life harder for the thousands of companies trying to build on top of them. The model will continue to matter, but more and more of the value will come from the system that knows what to do with it.
