80% of Startups Already Use Chinese Open AI Models, a16z Says

⏱️ Lectura: 11 min

80% of startups using generative AI already run at least one Chinese model somewhere in their stack, according to an estimate from investor Martin Casado of a16z, cited in The Economist. The figure explains why the debate around Chinese open models stopped being a geopolitical curiosity and became an architecture decision that every engineering team makes today, whether they notice it or not.

📑 En este artículo

1. TL;DR
2. What happened: Chinese open models are accelerating
3. Context and history behind the shift
4. Technical details and performance
5. How to try it today
6. Impact and analysis
7. What’s next
8. Frequently Asked Questions
   1. What’s the difference between an open-weight model and an open source one?
   2. Is it legal to use a Chinese AI model at a company outside China?
   3. Are Chinese open models cheaper per token than GPT or Claude?
   4. Does migrating from a closed model to an open one break my application?
   5. What’s the real risk of depending on a model trained in China?
   6. Can the United States respond by open-sourcing its own frontier models?
9. References

Moonshot AI and Alibaba released models that, according to their own benchmarks, match OpenAI and Anthropic with a fraction of the training compute. This is happening while Washington keeps betting on GPU export controls as the main brake on China’s AI progress. The result, according to analyst Robert Hart at The Verge, is that the United States’ edge at the AI frontier keeps getting tighter.

TL;DR

• Martin Casado (a16z) told The Economist that 80% of startups use at least one Chinese AI model in production.
• Moonshot AI and Alibaba launched models that claim to match OpenAI and Anthropic at a fraction of the training compute cost.
• US GPU export controls didn’t stop Chinese training, but they did limit centralized deployment at global scale.
• China releases its models as open weights: they can be downloaded, self-hosted, and modified without relying on a closed API.
• The real moat of an LLM isn’t the model itself, it’s the contracts, enterprise integration, and support, according to the werd.io analysis.
• Switching LLM providers costs little in code: most commercial APIs are compatible with the OpenAI format.
• Open weights aren’t the same as open source: they’re portable and permissively licensed, but don’t always include the dataset or the training pipeline.

What happened: Chinese open models are accelerating

Writer Ben Werdmuller summed up the phenomenon in an essay published on werd.io: American labs keep their frontier models behind a closed API, while Chinese labs release them as open weights. That difference in strategy, not in raw technical capability, is what’s moving the needle.

The sequence of releases was fast: Moonshot AI updated its Kimi family and Alibaba expanded Qwen with models that, according to their internal benchmarks, compete on equal footing with OpenAI and Anthropic’s closed products, trained with a fraction of the compute that Western frontier models require. Robert Hart described it in The Verge as a double blow: capability closing in on the frontier and cost moving away from it.

Context and history behind the shift

For several years now, the US government has restricted the export of cutting-edge GPUs to China, on top of limiting what data can be shared with Chinese servers. The practical consequence wasn’t stopping large model training in China, which keeps happening, but making it harder for Chinese companies to offer the same kind of centralized, global service that OpenAI or Anthropic run from their own cloud today.

That’s where the play appears: if you can’t compete by selling centralized access at global scale, you can compete by giving away the model’s weights. An open model can be hosted wherever the user wants, on their own server, on a different cloud provider, even without a permanent internet connection. It turns a compute disadvantage into a distribution advantage, exactly the central argument of the werd.io essay that sparked this piece.

The historical parallel is with internet infrastructure: open technologies, HTTP, TCP/IP, Linux, won adoption precisely because anyone could use them without asking permission. A closed model, on the other hand, depends on the customer trusting a single provider long-term, something almost nobody actually does with an LLM: you pick the best available model and switch providers if a better one shows up.

Technical details and performance

The underlying technical reason is that a language model, as a product in itself, has little moat beyond brand loyalty and the surface-level cost of switching providers. Nearly every commercial LLM API today replicates the request and response format that OpenAI popularized, which means migrating from one model to another is usually a matter of changing a base URL and an API key, not rewriting the application’s logic.

Option	When to use it	Advantage	Limitation
Closed model (GPT, Claude)	When you prioritize enterprise support, SLAs, and the best capability available at a given moment	Ready-made integration, constant updates, no infrastructure to run	Total dependence on one provider and its pricing
Open weights via third-party API (Kimi, Qwen, GLM)	When you want low cost without running your own hardware	Much lower price per token, no infrastructure lock-in	Fewer contractual guarantees, variable availability
Self-hosted open weights	When data can’t leave your network or you need full control over latency	Zero external dependency, customizable with fine-tuning	Requires your own GPUs and an MLOps team

The point that gets overlooked most often is that open weights doesn’t equal open source. A model with open weights can be downloaded and run, but it typically doesn’t include the full training dataset or the exact pipeline used to produce it. It’s portable and permissively licensed, not necessarily auditable end to end.

In practice, for a development team the difference shows up in the code, not in the application’s architecture:

import openai

client = openai.OpenAI(
    base_url="https://api.openai.com/v1",
    api_key="sk-your-openai-key"
)

respuesta = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Explain what an open-weight model is"}]
)
print(respuesta.choices[0].message.content)

That same client, with two lines changed, points to a provider serving a Chinese open-weight model:

import openai

client = openai.OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="sk-your-openrouter-key"
)

respuesta = client.chat.completions.create(
    model="moonshotai/kimi-k2",
    messages=[{"role": "user", "content": "Explain what an open-weight model is"}]
)
print(respuesta.choices[0].message.content)

The rest of the code, the prompt, the error handling, the response parsing, doesn’t change. That’s exactly Casado’s argument: if switching providers costs two lines, the moat of a closed model is much more fragile than it appears in commercial messaging.

How to try it today

To compare a closed model against an open-weight one without committing to your own infrastructure, the fastest route is an API aggregator like OpenRouter, which exposes Chinese and Western models under the same request format.

Windows (PowerShell):

python -m venv venv
.\venv\Scripts\Activate.ps1
pip install openai
$env:OPENROUTER_API_KEY = "sk-your-openrouter-key"

macOS:

python3 -m venv venv
source venv/bin/activate
pip install openai
export OPENROUTER_API_KEY="sk-your-openrouter-key"

Linux:

python3 -m venv venv
source venv/bin/activate
pip install openai
export OPENROUTER_API_KEY="sk-your-openrouter-key"

If you’d rather self-host the model instead of consuming it through an API, the usual route is downloading the weights from Hugging Face with the huggingface_hub library and serving them with a local inference runtime. To confirm the call is actually using the expected model and not a silent fallback, check the model field in the JSON response: it should exactly match the identifier you requested, for example moonshotai/kimi-k2.

💡 Tip: before migrating a production flow, run the same prompt against the closed model and the open one and compare latency and quality with your own data, not someone else’s benchmarks.

Impact and analysis

The underlying reasoning in the werd.io essay is economic, not technical: releasing open models commoditizes the layer where American companies make money, but in exchange it creates a much bigger global ecosystem, with more people building on that foundation. For China, that means gains in manufacturing, scientific research, and any sector that can connect an open model to its own infrastructure without negotiating a contract with a foreign provider.

flowchart TD
    A["GPU export controls"] --> B["China trains with fewer cutting-edge GPUs"]
    B --> C["Releases open weights: Kimi, Qwen, GLM"]
    C --> D["Global adoption without a license"]
    D --> E["The model layer gets commoditized"]
    E --> F["The moat shifts to integration and support"]

There’s an honest limit to this argument: an open model still reflects the worldview of whoever trained it. Chinese models systematically avoid certain topics sensitive to the Chinese government, something any team adopting them in an end-user product needs to evaluate before integrating them, not after.

⚠️ Careful: open weights doesn’t mean editorial neutrality. Before putting a model like this in front of users, test how it responds to topics that are politically sensitive for its country of origin.

The other point the werd.io analysis raises is that a large share of recent US economic growth is tied to Big Tech’s AI spending. If the closed-model business loses margin because the open alternative is just as good and much cheaper, the impact doesn’t stop at startups: it reaches the valuations of the labs that currently sustain much of the sector’s capital spending.

What’s next

The original essay’s author himself proposes a way out: betting on public AI infrastructure, federated services, and open research, with real backing instead of export controls as the sole strategy. None of these initiatives currently has the funding or scale of the closed labs, but they’re the direct counterpart to China’s strategy of giving away its weights.

For a development team in Latin America, the practical implication is more immediate: the price gap between a closed frontier model and a competitive open one is already small enough to justify, at minimum, testing both before committing to a single provider for a new product.

📖 Summary on Telegram: View summary

Try it yourself: set up the second code block from this article against your own production prompt and compare the result line by line against your current provider.

Frequently Asked Questions

What’s the difference between an open-weight model and an open source one?

An open-weight model can be downloaded and run freely, but it typically doesn’t include the full training dataset or the exact pipeline. Open source, in the strict sense, means the entire process is reproducible and auditable.

Is it legal to use a Chinese AI model at a company outside China?

In most countries, yes, as long as you respect each model’s published usage license and local data protection regulations, which usually apply the same way regardless of the model’s origin.

Are Chinese open models cheaper per token than GPT or Claude?

Yes, consistently, because they’re served through multiple providers that compete with each other on price, something that doesn’t happen with a closed model available through a single API.

Does migrating from a closed model to an open one break my application?

If your integration uses the OpenAI-compatible API format, the change is usually limited to the base URL, the key, and the model name, without touching the rest of the code.

What’s the real risk of depending on a model trained in China?

The main one is bias on topics sensitive to the Chinese government and regulatory uncertainty around data handling, not a technical performance risk.

Can the United States respond by open-sourcing its own frontier models?

Technically yes, but the current commercial incentive pushes American labs to keep their models closed to protect direct revenue, according to the werd.io analysis.

References

• werd.io: the original essay arguing that closed American AI is losing ground to open Chinese AI.
• The Verge: Robert Hart’s coverage of Moonshot AI and Alibaba’s releases and their impact on the AI race.
• a16z: venture capital firm whose partner Martin Casado estimated the percentage of startups using Chinese models.
• Hugging Face: repository where the weights of most Chinese open models are published.
• OpenRouter: API aggregator that exposes closed and open-weight models under a common format.

📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.

Imagen destacada: Foto de luca romano en Unsplash