Latest / China Economy with Fexingo: Beijing's Policies, Trade, and Asia-Pacific Markets / How Chinese AI Models Are Winning US Business Clients
Transcript
- Lucas: So there was a headline this morning I think a lot of people in tech circles are chewing on: Chinese AI models are starting to land real contracts with U.S. companies. This isn't just a lab experiment anymore. Luna: Yeah, I saw that. The piece mentioned OpenAI and Anthropic costs are surging, and suddenly the price gap is opening a door for Chinese players. But what's the actual scale here? Lucas: Let me give you a number that caught my eye. Chinese ai related exports — software, models, API access — grew about 25 percent year-over-year in the first half of 2026. That's according to data from China's General Administration of Customs. And a chunk of that is going to U.S. firms looking to cut their AI bills. Luna: Twenty-five percent is real. And you know, if today's deep dive is making the economics of this shift click, it's because shows like this stay ad-free thanks to listener support. If it's been useful, you can find us at buy me a coffee dot com slash fexingo. Lucas: Exactly. That direct support is what keeps us independent. So back to the data — one of the key drivers here is the yuan. It's been stable around 6.79 per dollar for weeks, which actually makes Chinese AI services more predictable on pricing than some alternatives. Luna: Right, because if the yuan starts appreciating, that price advantage erodes. But for now, U.S. companies are locking in contracts. Which specific models are winning? Lucas: Mainly two: DeepSeek's latest model, which is built on a mixture of experts architecture that keeps inference costs low, and Baidu's Ernie 4.5. Ernie actually scored competitively on some industry benchmarks against GPT-4, but at roughly half the API cost per token. Luna: Half the cost. That's hard for CFOs to ignore, especially when enterprise AI budgets are under scrutiny. Are there any catches — data privacy, latency, export controls? Lucas: Sure, there are catches. The Biden-era export curbs on advanced chips haven't been fully rolled back, so these Chinese models are running on less powerful hardware. That can mean higher latency on complex tasks. Also, data sovereignty is a concern — some U.S. clients are requiring that all inference happens on domestic servers. Luna: But that's something Chinese providers have been accommodating. I've read that DeepSeek set up a U.S.-based inference node through a partnership with a cloud provider. So they're adapting. Lucas: They absolutely are. And look at the market reaction. The KWEB — that's the China internet ETF — jumped over 5 percent in the last five trading days. PDD Holdings, which owns Pinduoduo and Temu, surged more than 9 percent. That's partly broader tech optimism, but the AI narrative is feeding into it. Luna: That's a big move for PDD in particular. Their e-commerce business is strong, but are they actually an AI play? Lucas: More than you'd think. They've been using AI for supply chain optimization and personalized recommendations, and they recently launched a small language model for merchant tools. It's not their core story, but it adds to the perception that Chinese tech companies are innovating. Luna: So the U.S. trade deficit in services — which is already at about 55 billion dollars — could actually narrow if these AI exports keep growing. That's an interesting twist. Lucas: Exactly. The trade-weighted dollar index dipped slightly to 120.7, and a weaker dollar makes U.S. imports of Chinese services cheaper in relative terms. That's a tailwind for this trend. Luna: And all of this is happening while the U.S. is still the leader in foundational AI research. But the cost curve is shifting fast. What does this mean for the broader AI industry? Lucas: I think we're heading toward a bifurcated market. On one side, you have premium models like GPT-5 or Claude 4 for high-stakes tasks where accuracy is everything. On the other, cost-efficient models from China and open-source alternatives that handle 80 percent of use cases at a fraction of the price. Luna: That sounds a lot like what happened with cloud computing. AWS and Azure kept the high end, but Alibaba Cloud and others carved out a massive middle market. Lucas: Exactly the parallel. And I think it's why we're seeing U.S. companies like Google invest in nuclear fusion — they're betting that AI's energy demands will eventually overwhelm current power grids, and they want a clean, cheap source ready. Luna: That Google fusion story is wild. They backed a startup targeting Europe's first commercial plant. But that's a long play. In the near term, Chinese AI is here, and it's winning clients. Lucas: It is. And if the yuan stays stable and the cost gap remains, more U.S. companies will test these models. The question is whether quality concerns will limit adoption to less critical tasks. Luna: Or whether the models improve fast enough to close the gap. DeepSeek's latest benchmark scores were within a few points of GPT-4 on reasoning tasks. That's not nothing. Lucas: Not nothing at all. And it's worth noting that China's overall export machine is still humming. Exports of goods and services hit 3.5 trillion dollars in the first quarter, up from 3.35 trillion a year ago. Tech is a growing slice of that. Luna: So if you're a U.S. tech procurement manager, you might be looking at Chinese AI and thinking, 'Can I afford not to consider it?' Especially when your board is asking why the AI budget doubled. Lucas: Right. And that's the tension. National security concerns, export controls, data privacy — those are real. But the cost pressure is equally real. Companies are already voting with their wallets. Luna: I'd love to check back in six months and see how many of those trial contracts convert to long-term deals. That'll be the real signal. Lucas: Absolutely. We'll track it. For now, the trend is clear: Chinese AI is no longer just a domestic story. It's going global, and the U.S. market is the biggest prize.