American AI Is Losing the Download Race. Is That the Market?

Ben Werdmuller's July 20, 2026 essay, "American AI is locked down and proprietary. It's losing", argues that Chinese labs are turning a compute disadvantage into a distribution advantage by releasing model weights. He thinks American labs are defending centralized services around a thin model-level moat.

I agree with the warning, but not the obituary. Chinese labs have won momentum in the open-model channel because developers can host, adapt, and route around a vendor. That does not mean the whole American AI business has lost. Portability is now a product feature, and a lab that withholds it has to earn the restriction through capability, support, reliability, or integration.

Answer snapshot

Question	My read
What changed?	Chinese model families now lead several measures of open-model downloads, derivatives, and third-party inference use.
What does that prove?	Open weights are a strong distribution strategy. They make local deployment, third-party hosting, and adaptation easier.
What does it not prove?	Downloads are not active production use, and open-model adoption is not the entire AI market.
Who benefits?	Teams that need data residency, customization, provider choice, or the ability to run without sending prompts to the model developer.
My decision rule	Demand model portability, but compare the full operating cost and run your own workload eval before moving production traffic.

The distribution lead is measurable

The strongest evidence for Werdmuller's case comes from the April 2026 ATOM report. It tracked 2.04 billion open-model downloads through March 2026. Models from Chinese builders accounted for 1.15 billion, compared with 723 million for U.S. builders. The gap had widened to 428 million. The report also found that Chinese models reached 70% of new model derivatives by February 2026.

Derivatives matter because people fine-tune or adapt models they consider useful raw material. Hugging Face's spring 2026 review says Chinese models represented 41% of downloads over the previous year, while more than 30% of Fortune 500 companies maintained verified accounts on the platform.

I read that as a distribution win, not a scoreboard for national destiny. Once weights can travel, the original lab no longer needs to serve every request. A local team, cloud provider, or inference marketplace supplies the hardware and puts the model near its users. That reduces data-residency friction and gives developers leverage over pricing and availability.

A download is not a deployment

The ATOM report is candid about its limits. Hugging Face is not the only distribution channel, a production deployment may create just one download, and automated pipelines or repeated pulls can inflate counts. The report says its other usage measures show a similar regional shift, but it does not claim that every download is an active customer.

That limitation also showed up in LocalLLaMA's public discussion. Some developers credited Qwen and DeepSeek's release cadence. Others noted that small infrastructure models and poorly configured automation can generate many pulls. The momentum is real, but the headline number still needs care.

The broader market gives the same mixed picture. Stanford's 2026 AI Index says the performance gap between the top U.S. and Chinese models had narrowed to 2.7% by March 2026. Yet the United States still produced more top-tier models, and U.S. private AI investment reached $285.9 billion in 2025, versus $12.4 billion in China. "Losing" depends on which layer you measure.

Open weights are not the whole source

There is another precision problem in this debate. Portable weights are useful, but they are not automatically open source. The Open Source Initiative's definition also calls for the code and data information needed to study and modify how the system was produced. A downloadable checkpoint without that material gives a team deployment freedom, not full reproducibility.

American AI is not uniformly closed. OpenAI released gpt-oss-120b and gpt-oss-20b under Apache 2.0 in August 2025. That does not erase the Chinese ecosystem's momentum, and OpenAI calls the models open-weight rather than fully open source. It does show that labs can change how much of the stack they release.

Control sends work back to the buyer

Werdmuller is right that enterprise services can become the moat. The mistake would be treating that as superficial lock-in. Contracts, support, identity, audit logs, uptime, regional hosting, and integrations are often the product that an enterprise is buying. A portable model can weaken model-level dependence while leaving plenty of hard operational work untouched.

OpenAI's own gpt-oss support page makes the trade clear. Self-hosted deployments are self-managed, OpenAI does not provide hands-on debugging for them, and the operator pays the compute and storage costs. That can be an excellent bargain for a team with sensitive data, steady utilization, and infrastructure expertise. It can be a bad one for a small team that mainly wants a reliable API.

I would ask a buyer what must remain portable. Prompts and tool schemas should survive a provider change. Evaluations should cover more than one model. Sensitive workloads need a credible local or third-party path, while logs and audit evidence should remain exportable. Those choices create leverage even when the best current model is proprietary.

My takeaway

Chinese labs have made open weights a serious distribution channel, and American labs should treat that as competitive pressure. The 428 million-download gap is a signal that developers value models they can move and adapt. It is not proof that managed services are finished or that the U.S. AI economy is about to collapse.

The decision I would make today is less dramatic. I would pay for a closed service when it earns its margin, but I would design the application so the model is replaceable and keep an open-weight option under evaluation. The winner may not be the lab with the best demo or the most downloads. It may be the one that gives customers a convincing reason to stay without making them unable to leave.