Silicon Valley Is Quietly Running on Chinese AI

Opening

Dear reader, back in October last year, Airbnb CEO Brian Chesky said something in a Bloomberg interview. About Alibaba’s Qwen model, he said it was “really good, fast, and cheap.” Coming from someone with a personal friendship with Sam Altman, saying this out loud in public tells you the mood in Silicon Valley has shifted quite a bit.

And it’s not just Alibaba’s Qwen. MiniMax and Kimi, both started as scrappy startups, are in the mix too, along with DeepSeek. Even in the recent OpenClaw craze, the most frequently chosen models turned out to be Chinese models.

NBC News actually conducted anonymous interviews with 15 AI startup founders, ML engineers, and investors, and most of them said they were already deploying Chinese AI models in production—because of cost. But this isn’t simply a story about “using the cheap option.” Today I want to dig into what this phenomenon structurally means.

What the numbers tell us: the truth behind 80%

A figure that Martin Casado, general partner at a16z (Andreessen Horowitz), mentioned in an interview with The Economist shook the industry. “80% of the open-source-based startups pitching us are using Chinese models,” he said.

But there’s context worth flagging here. Casado himself later clarified that this 80% is the share among startups that use open-source models specifically. Since the share of all startups using open source is roughly 20-30%, the real figure works out to about 16-24% of all startups using Chinese open-source models. Still, considering this was close to 0% just a year ago, the speed of this shift is the real story.

Hugging Face download data makes this trend even clearer. Alibaba’s Qwen series overtook Meta’s Llama in cumulative 2025 downloads, and according to MIT research, total downloads of Chinese open-source models have now surpassed American models. In particular, Qwen-based derivative models account for over 40% of new language models on Hugging Face, while Llama-based ones have fallen to around 15%. This means Qwen has effectively become the default base model for global open-source AI.

Why this is happening: price × openness × speed

To understand this phenomenon, you need to look at three axes.

First, the price gap is overwhelming. DeepSeek V3.2’s API price runs about $0.28 per million input tokens. OpenAI’s GPT-4.1 costs $2.00 on the same basis—roughly a 7x difference. Add self-hosting¹​ into the mix, and the cost effectively converges to zero. For an early-stage startup managing burn rate on VC funding, if API costs eat up 20-40% of monthly spend, this is a matter of survival.

Second, openness has built an ecosystem. Major US models (GPT, Claude) are mostly closed—they don’t release weights²​, so there are limits to customization. Chinese models like Qwen, DeepSeek, and Kimi, by contrast, publish their weights, letting anyone download, modify, and redistribute them. According to Antonio Vespoli, co-founder of browser-agent startup Circlemind AI, the models with the richest training guides and technical support materials in the developer community are already the Chinese ones.

Third, release speed is on a different level entirely. MIT researcher Shayne Longpre pointed out that Chinese AI labs release models on a weekly or biweekly cycle. Compared to major US labs updating models every six months to a year, developers effectively see their options expand far more quickly.

Innovation born of constraint: the paradox of chip export controls

Here’s an interesting paradox. The US restricted exports of Nvidia’s latest chips (like the H100) to slow China’s AI progress. But it looks like this constraint ended up accelerating China’s lightweight-model innovation instead.

DeepSeek has said it was developed using older chips (H800) that weren’t subject to export restrictions, and Perplexity CEO Aravind Srinivas described the situation as “necessity is the mother of invention.” According to a UBS China internet analyst, Chinese internet companies’ 2025 capital expenditure was about $56 billion—roughly one-tenth of comparable American firms—yet their AI model performance has reached a “comparable” level.

Lightweighting has progressed to the point where an image-generation model with around 600 million parameters³​ can run on an ordinary laptop—an M2 or later MacBook, or a 12th-gen Intel Windows laptop with 32GB of RAM, is enough. This isn’t the “democratization” of technology—it’s the barrier to entry disappearing altogether.

An inflection point for Western open source: Meta’s choice

Amid all this, Meta’s Llama—the only meaningful open-source model from the Western world—stands at a crossroads of its own.

Yann LeCun, Meta’s chief AI scientist and one of the “godfathers of deep learning,” announced his departure in November 2025 after 12 years at the company. He has long publicly argued that “LLMs⁴​ cannot reach human-level AI,” and he’s founded a new startup, AMI Labs, to research “world models”—AI that understands the physical world.

More notable is that Meta’s AI strategy has been shifting since LeCun’s departure. Alexandr Wang, formerly of Scale AI, has taken over as the new head of AI, and the research-focused FAIR team has been scaled back. There’s also talk that a new licensing regime may apply from Llama 4 onward. If the role Llama has played in the open-source camp shrinks, who fills that vacuum? Given the current trajectory, the answer seems clear.

Oz’s Lens

Honestly, I think reading this phenomenon through a “US vs. China” frame is itself a trap.

There’s a pattern I’ve kept seeing while building GTM strategies. What actually changes the game in tech markets isn’t performance superiority—it’s a shift in accessibility. The PC didn’t replace the mainframe because it was more powerful. AWS didn’t transform enterprise IT by building better servers—it eliminated the need to buy servers in the first place.

What Chinese open-source models are doing now is similar. They’re not competing to build “better models”—they’re pursuing a strategy of eliminating the cost and entry barriers to using models altogether. A structure where Xi Jinping’s government directly supports the open-source ecosystem, and companies accept short-term losses to release their models openly, is—like it or not—strategically very coherent.

What genuinely worries me is Korea. Caught between the US and China, in AI models, in the open-source ecosystem, in the chip supply chain—we still haven’t found a seat at the table. We shouldn’t let this end as a “wow, China is impressive” consumption-news moment. The question needs to go further: why can they get this level of performance out of such small models, and how can we make use of this ourselves?

Closing

To sum up:

• A significant share of Silicon Valley startups have already adopted Chinese open-source AI models in production, and this trend is accelerating.
• This isn’t simple cost-cutting—it’s a signal of a structural shift in where value in the AI industry sits, moving from “building models” to “building services on top of models.”
• The paradoxical outcome where chip export controls actually spurred China’s lightweighting innovation deserves attention too.

Next time you’re choosing an AI tool, try using “how efficiently does this service solve my problem” as your criterion, rather than “which company’s model is it.” The market is already moving that way.

References & Further Reading

• NBC News, “More of Silicon Valley is building on free Chinese AI,” 2025.11.30. The feature report that sparked today’s newsletter. The 15 anonymous interviews are the heart of it.
• MIT Technology Review, “What’s next for Chinese open-source AI,” 2026.02.12. The most systematic breakdown of OpenRouter data and Hugging Face download trends.
• Al Jazeera, “China’s AI is quietly making big inroads in Silicon Valley,” 2025.11.13. Includes an interview with Nathan Lambert (ATOM Project). Good picture of how Chinese models are actually being adopted in the US.
• CNBC, “Meta chief AI scientist Yann LeCun is leaving to create his own startup,” 2025.11.19. Helpful for understanding the context of LeCun’s departure and Meta’s AI reorg.
• MIT Technology Review, “Yann LeCun’s new venture is a contrarian bet against LLMs,” 2026.01.22. LeCun explains directly, in his own words, why he broke with the LLM paradigm and where AMI Labs is headed.
• IntuitionLabs, “LLM API Pricing Comparison (2025),” 2025.10.31. A comparison of token pricing across major AI models. Useful if you want to see the cost gap in hard numbers.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. Self-hosting: running an AI model on your own servers instead of calling an external API. There’s an upfront setup cost, but it becomes cheaper than API costs as usage scales up. ↩

2. Weights: the numerical parameters an AI model learns through training. Think of them as “judgment criteria learned from experience,” in human terms. Releasing weights lets anyone copy and modify the model. ↩

3. Parameter: a unit describing an AI model’s size. A 6B (6 billion) model is relatively small; GPT-4-class models are estimated at several hundred billion or more. Fewer parameters mean less computing power needed to run the model. ↩

4. LLM (Large Language Model): an AI model that works by learning from massive amounts of text data to predict the next word. ChatGPT, Claude, and Gemini are all representative LLMs. ↩