An American tech founder in San Francisco just announced he moved 100% of his company's AI traffic to a Chinese model because it was ten times cheaper than using Anthropic, the U.S. darling backed by billions in American investment. He says it saved him millions. He says it was a very simple business decision. He is not wrong, and that is exactly the problem.
When Anthropic Costs More Than Your Entire Payroll
Flo Crivello runs Lindy.ai, a San Francisco startup that builds AI assistants to manage your email and calendar. He had more than two dozen employees. He had rent. He had every normal cost a growing tech company racks up. And according to NPR, his single largest expense, bigger than all of it, was Anthropic.
Not Anthropic in a 'wow, AI is getting expensive' kind of way. Anthropic more than payroll. Anthropic more than rent. Anthropic more than anything else the company spent money on. So last month, Crivello announced Lindy had migrated completely to DeepSeek-V4, a Chinese AI model. The math was not complicated. 'It was just 10x cheaper,' he told NPR.
He also said that every founder he knows working in the AI space is either thinking about switching to Chinese models or has already done it. Every. Single. One.
This Is Not Just a Startup Problem
Before you write this off as a bunch of scrappy little tech companies pinching pennies, consider what NPR also dug up: Uber CEO Dara Khosrowshahi said on a podcast last month that his company blew through its entire annual AI budget in a single quarter. 'It is forcing us to adjust,' he said. Uber, a company with a market cap north of $150 billion, got caught off guard by how fast AI costs spiral.
Bloomberg reported that Airbnb CEO Brian Chesky said his company leaned on Alibaba's Qwen model last year, describing it as 'good,' 'fast and cheap.' Perplexity and Nvidia have also used Qwen. These are not fringe players experimenting with bootleg software. These are flagship American technology companies quietly routing work through Chinese AI because the American alternative is financially brutal.
The pattern is clear enough that OpenRouter, a platform where companies can access multiple AI models, told NPR that usage of DeepSeek has nearly doubled since January, jumping from around 9% to close to 20%. Chinese models from MiniMax, Xiaomi, and Tencent are also climbing.
The Ferrari vs. Honda Argument, and Why It's Winning
Eugene Cheah runs Featherless, a San Francisco company that gives developers access to roughly 30,000 AI models. His take on the situation is blunt and kind of beautiful in its simplicity. NPR quotes him saying Chinese models are like a Honda and American frontier models are like a Ferrari. 'You can have the best luxury car, or you can just have a Honda at scale that works.'
Here is the thing about that analogy: most businesses do not need a Ferrari. Most businesses need a vehicle that shows up every day and does not cost more to fuel than the business makes. For repetitive, high-volume tasks like coding, customer service automation, or processing data, a model that performs at one-tenth the cost is not a compromise. It is the rational choice.
Victor Su-Ortiz from MiniMax, a Shanghai-based AI company, attended an AI engineers conference in San Francisco recently and told NPR the same thing from the supply side. Companies are moving away from maximizing AI usage at any cost toward actually thinking about cost per token. 'A lot of repetitive tasks can be done with a model that's just as performant but has much lower cost per token,' he said. The market is responding to basic economics, which tends to happen whether governments like it or not.
The Political Sensitivity Nobody Wants to Talk About
NPR reports that many companies are wary of publicly announcing their use of Chinese AI models because of the political environment around it. Which, given the current administration's ongoing trade war with China and the general hysteria around anything with a Chinese IP address, is extremely understandable and also extremely telling.
So they route traffic through U.S.-based hosting companies like Featherless and OpenRouter, keeping the data on American servers, and nobody has to have an awkward press conference about it. The Chinese model does the work. The American middleman holds the data. Everyone gets plausible deniability. It is an elegant arrangement born entirely out of the gap between what American AI companies charge and what the market is actually willing to pay.
American AI experts told NPR that Chinese models are roughly six to twelve months behind U.S. frontier models in raw capability. But when the gap keeps shrinking and the price difference stays at ten-to-one, the capability gap starts to matter a lot less to a startup trying not to run out of runway.
What the Open-Source Piece Actually Means
One detail in NPR's reporting deserves more attention than it typically gets. Crivello told them flatly: 'The open-source scene right now is absolutely dominated by the Chinese. It's not even close.' Chinese AI companies have made aggressive bets on releasing models openly, meaning developers can download, modify, and deploy them without paying ongoing licensing fees to the original company.
This is not an accident. It is a deliberate strategy to gain adoption and trust in global developer communities, and it is working with remarkable efficiency. While American companies like Anthropic and OpenAI have largely kept their best models locked behind APIs with steep per-token pricing, Chinese companies flooded the open-source ecosystem. Now they own it.
The downstream consequences of that strategic difference are exactly what NPR's reporting describes: American startups making million-dollar decisions to migrate entirely to Chinese models because the economic case is overwhelming and the capability gap is shrinking every quarter.
The Dingo Take
Let's just sit with the full absurdity of this for a moment. The United States government has spent the last several years treating Chinese technology companies like an existential national security threat, banning TikTok, blacklisting chip suppliers, and building an entire political movement around the idea that anything touching China is a vector for catastrophe. Meanwhile, in San Francisco, the physical heart of American technological supremacy, startup founders are migrating their entire operations to Chinese AI because the American product costs more than their payroll. The market has spoken. It said 'ten times cheaper, please.'
This is what happens when you build an industry on the assumption that being first and being American is enough of a moat. OpenAI and Anthropic raised staggering sums of money and charged prices to match their ambitions. Chinese companies, partly by design and partly out of competitive necessity, flooded the open-source market with models good enough for most real-world tasks at a fraction of the price. And now, quietly, the switchover is happening at scale. Not because American founders hate America. Because they have a finance guy in a recurring meeting pointing at a spreadsheet.
The rich irony is that the political figures screaming loudest about Chinese technological threats are the same ones whose trade policies and deregulatory instincts helped create the cost pressures driving this shift in the first place. American AI is expensive because American AI companies need to recoup genuinely enormous infrastructure investments. Chinese models are cheap because of very different economic incentives and state backing. Nobody in Washington seems interested in grappling with that actual dynamic. It is much easier to hold a hearing about TikTok.
