Why Companies Are Quietly Choosing Chinese AI Models

Companies Choosing Chinese Models

In an October 2025 interview, Airbnb’s Brian Chesky said his team relies heavily on Qwen, praising its speed and cost. It’s an American service adopting a model built by a Chinese company.

I think what matters in this shift isn’t just model performance—it’s the terms of adoption. A developer first checks whether a model meets the performance their service needs, then calculates the cost, and also weighs whether they can modify it or run it themselves. When more options open their model weights, as Qwen and DeepSeek do, the basis for comparison changes entirely. Weights are the numerical values that capture what a model has learned. If you can download them, you’re no longer limited to calling an external API—you can also consider running the model on your own servers.

The 80% Martin Casado Was Talking About

In a conversation published on January 20, 2026, a16z general partner Martin Casado talked about the companies that come to him for investment advice. His observation: among those using open-source models, roughly 80% were using Chinese models.

This figure isn’t a market-share survey of all Silicon Valley startups. It comes with a specific condition — it only covers the companies Casado personally met that happened to use open-source models. You also can’t multiply this by open-source adoption rates from other surveys to calculate a Chinese-model usage rate across all startups, since these surveys aren’t sampling the same companies. Still, it’s a meaningful testimony from someone on the ground in venture investing who says he frequently encounters companies that have adopted Chinese models.

Research by Shayne Longpre and others analyzing Hugging Face downloads also confirms the spread of Chinese models. The study reported that when comparing models built between August 2024 and August 2025 specifically, Chinese models had a higher share of downloads than American models. This doesn’t mean Chinese models surpassed all American models in cumulative downloads overall.

Downloads are a data point for gauging developer interest and usage. The same person might download a model multiple times, or downloads might happen automatically — so this figure shouldn’t be read directly as the number of companies using these models or as a share of paid services.

Pricing and operating model can change the choice

For an early-stage startup, how fast API costs pile up really matters. Every additional user means more model calls, and that means the choice of model feeds directly into your service’s pricing and margins.

When DeepSeek announced V3.2-Exp in September 2025, it said it had cut API prices by more than 50%. A price cut like that forces developers to recalculate whether sticking with their current model still makes sense. Of course, you can’t decide based on the price sheet alone. You also need to look at how accurate and fast the answers are for the same task, and how many retries it takes to get there.

There’s also the option of downloading the weights and running the model yourself. In that case, instead of paying an outside company per call, you take on the cost of GPUs, server operations, and incident response. If your usage volume is high and you have the operational chops for it, this can work in your favor — but running it yourself doesn’t magically bring costs down to zero. For teams with lighter usage, the API may still be simpler and cheaper.

The Freedom That Open Weights Provide

Qwen3 released its model weights under the Apache 2.0 license. Within the permitted terms, developers can download and modify the model and apply it to their own services. That’s a meaningful distinction for teams who want to run the model in their own environment or fine-tune it for a specific task.

That said, not all Chinese models are released the same way. Even within a single company, terms can differ from model to model. There’s also a real difference between releasing just the weights and releasing everything—training data and training code included. You need to check the license and scope of what’s actually provided for each model you’re considering adopting, on a case-by-case basis.

Compared to API-centric models like GPT or Claude, open-weight models offer more choices in how you operate them. But that comes with a flip side: the burden of handling installation and operations yourself. Rather than settling the debate with “which side is more open,” I think what matters is whether that freedom of choice actually solves a real deployment problem for the customer.

The frequency of new models and updates also affects developers. DeepSeek released V3.2-Exp in September 2025, then followed up with the official V3.2 in December. Developers end up re-testing cost and performance even against new versions from the same provider. That said, you can’t simply generalize this as “Chinese companies ship every week, American companies ship once every six months.” What matters isn’t the release cadence—it’s whether the new version actually improves your own workflow.

How to Boost Efficiency Under Resource Constraints

There’s an interesting angle when you look at chip export controls alongside the progress of Chinese models. In describing the infrastructure used to build V3, DeepSeek’s researchers said they used 2,048 NVIDIA H800 GPUs. The H800 isn’t some older chip that escapes regulation entirely — according to NVIDIA’s disclosures, the H800 was included when export licensing requirements were expanded in October 2023.

DeepSeek’s researchers described a Mixture-of-Experts (MoE) architecture that activates only a subset of experts rather than the full model for each computation, an MLA scheme that reduces memory usage, and FP8 training that computes with lower-precision numbers. I think it’s plausible that resource constraints heightened the need for this kind of efficiency work. Still, concluding that export controls actually caused this technical progress would require separate evidence.

Meanwhile, more models are emerging that can run on small hardware. The image-generation model Z-Image-Turbo, for instance, uses 6 billion parameters, and its official documentation says it can run on a consumer GPU with just 16GB of VRAM. (VRAM here refers to the memory the GPU itself uses.) That doesn’t mean having 32GB of system memory on a regular laptop gets you the same result. And the sub-one-second generation speed cited in the official materials was also measured on an H800.

As these models appear, more people can test them firsthand before renting a large server. But to actually judge how much the cost of adoption has come down, you need to know exactly what hardware it runs on and at what speed.

Meta Isn’t the Only Alternative

When people talk about openly-weighted US models, Llama comes up constantly. But the US and Europe have other options too, like Gemma and Mistral. The fact that Chinese models are spreading doesn’t mean the entire open-model market has tipped to one side.

Llama also operates under its own separate license. Even if Qwen3 and Llama both get lumped together as “open source,” their actual usage terms aren’t the same. We need to separate two different things: predictions about what models companies will release going forward, and the terms that actually apply to the models you download today.

Companies building open models are competing on performance, but they’re also competing to keep developers using their models going forward. Once a developer builds a service around one model and accumulates operational experience with it, that model becomes the natural first choice for their next product too. This is the GTM angle I find most interesting to watch.

Oswarld’s Lens

I think reading this purely as a US-China national competition misses an important shift.

There’s a pattern I’ve kept seeing while building GTM strategies. It’s never just the highest-performing product that reshapes a market. Products that make adoption easier and cheaper for more people also expand how widely something gets used. The PC let individuals own and use a computer themselves; AWS spread the practice of renting server capacity as needed instead of buying it outright.

I see a similar strategy in China’s open-weight models. Making a model easy to run yourself and pricing the API low lowers the barrier for developers just starting to build something. As more services adopt that model, development resources and operational know-how accumulate around it too. It’s a way of driving adoption that runs independent of raw performance rankings.

What worries me is what role Korea can secure for itself in this race. We already have a foundation built up in semiconductors and model development, but that foundation doesn’t automatically translate into developer and customer choice. Building a good model and getting people to build services on top of that model are two separate problems that each need solving.

I don’t want this to end with “China’s models got better.” We need to look further — which tasks are these models now good enough for, what does adoption and operation actually cost, and how can we use or compete with this.

If I were choosing my next AI tool, I’d run real work through it first. I’d compare the output and total cost, and check whether the data-handling practices and licensing fit our requirements. Rather than choosing based on which country the maker is from, I think it’s more useful for building products to scrutinize these concrete conditions instead.

References

• Bloomberg: Brian Chesky interview, October 21, 2025
• a16z: Martin Casado conversation, January 20, 2026
• Longpre et al.: The ATOM Project — a study of the open-model ecosystem
• DeepSeek: V3.2-Exp and API price cuts
• DeepSeek: V3.2 announcement
• DeepSeek researchers: training infrastructure and efficiency
• NVIDIA: October 2023 disclosure on export controls
• Qwen3 official repository
• Z-Image-Turbo official model card
• Llama 4 license

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.