Tiezhen Wang, a former Hugging Face executive, is the kind of person who has watched AI development from the inside, across both “model building” and the messy ecosystem around it. In Rest of World’s piece, she frames a pivotal question in the US-China AI competition: can open-source beat closed AI labs like OpenAI and Anthropic, simply by moving faster and spreading capability more widely?

The core claim is not that open-source automatically wins, but that China’s open-source strategy changes the tempo of the race. If more teams can see, adapt, and build on the same foundational work, progress becomes less dependent on a single lab’s willingness to ship or disclose. In a closed model world, one organization can control access to models and training methods. In an open model world, many organizations can contribute improvements, fine-tune for specific tasks, and reduce experimentation time. That speed difference is the stake. If capability dissemination accelerates enough, the advantage can shift from “who invented the first version” to “who operationalized the ecosystem first.”

To understand why this matters, you have to zoom out from any one model. AI has become a platform industry. Compute access, data pipelines, toolchains, and deployment patterns often matter as much as raw model architecture. Closed-source labs can create a moat by limiting what others can learn from, but they also create coordination bottlenecks. Every new capability involves licensing decisions, product timelines, and compatibility hurdles. Open-source communities typically trade some central control for parallelism: more contributors, more experiments, and more adoption paths. For boards and executives, that means the risk is not just technical. It is organizational. A partner ecosystem that grows faster than your internal R&D can effectively outflank you even if your core team is strong.

This is where the US-China split in engineering philosophy becomes more than a slogan. American pioneers such as OpenAI and Anthropic are often associated with a closed-source approach, at least in the way their models and surrounding systems are delivered. Meanwhile, China’s open-source push, as Wang describes it, suggests a different playbook: reduce friction for building on top of existing work, encourage forks and adaptations, and let the broader developer base help scale experimentation. When more actors can iterate, improvements can accumulate quickly. That has second-order effects for the whole market, because it changes what buyers expect. Enterprises typically want reliability, integrations, and customization. If open ecosystems support those needs faster, they can pull demand away from slower-moving closed platforms.

There is also a regulatory dimension to this story, even when regulators are not explicitly mentioned in a single headline. In the AI space, oversight can touch model release policies, data provenance, and how systems are used in high-stakes settings. Closed-source strategies can sometimes be framed as a way to keep tighter control over safety and compliance. Open-source strategies can sometimes be framed as transparency and broad review. The practical consequence is that regulation can shape which approach is easier to deploy at scale across different jurisdictions. If open ecosystems fit regulatory and compliance processes better in certain markets, that further reinforces adoption, which then reinforces the ecosystem itself.

And then there is the capital and talent flywheel, the part executives feel in budgets and hiring plans. Open-source tends to lower the barrier to entry for new entrants and accelerates skill accumulation. Developers can study, reproduce, and improve work without waiting for a proprietary release cycle. That can expand the pipeline of future engineers and researchers who can contribute value quickly. For leadership teams, that means competitive threats can arrive from places that would not have mattered in a purely closed environment. A startup that would have needed access to black-box model capabilities might now be able to build on a community base and still compete.

So can open-source beat closed AI labs like OpenAI? Wang’s framing points toward a “yes, potentially, if the ecosystem advantage is big enough.” In other words, the win condition may not be who can produce the first headline-grabbing model. It may be who can turn model progress into widespread, operational capability faster. For decision-makers in adjacent AI roles, the strategic stakes are clear. If your competitors are designing for parallel adoption, your differentiation cannot only be “we have great models.” It has to be “we have great integration, great iteration speed, and an ecosystem that keeps partners close.”