Is American AI out of favor?
DoorDash, Airbnb, Siemens, these names put together, there‘s nothing special about them, but what if I told you that they are quietly shifting the AI business from the American model to the Chinese model?
On July 13th, the Financial Times reported that from Silicon Valley to Europe, a group of world-renowned companies are voting with their feet. Data from the OpenRouter platform show that Chinese models have completely surpassed mainstream American models in weekly usage on the platform since this year, with their share reaching a peak of 46%.
This collective “crossover” of European and American companies even took place in the technological hinterland of the United States.
For a long time, the United States has been regarded as the absolute leader in global artificial intelligence, and the flagship models of companies such as OpenAI and Anthropic have been packaged as “unsurpassable technological benchmarks.”
However, in just half a year, from startups to industry giants, more and more companies made the opposite choice with real gold and silver.
This is not some accidental deviation from the technological path, but a total outbreak of the accumulated misdeeds of the American AI industry itself. When premiums, closure, and political manipulation become the norm, the market‘s backlash is only a matter of time.
The first disease point is the complete distortion of the price system, and American AI is “slaying customers.”
Let‘s look at the numbers first. According to OpenRouter‘s public bid, the Anthropic flagship model Claude Opus 4.7 is priced at over $24 per million output tokens.
Meanwhile, the Chinese specter GLM-5.2 is less than $ 4, and lightweight versions such as DeepSeek V4 Flash are even a few dozen times lower than the U.S. flagship model.
The performance difference between the two is minimal in most business scenarios, but the price difference reaches 60% to 90%.
This is not a normal technology premium, but a profiteering model that utilizes market monopoly position to harvest customers.
Andy Fang, co-founder of DoorDash, put it bluntly: By handing low-level tasks to the Chinese moon‘s dark side, the Kimi model, and leaving only the most complex work to the American model, the overall costs were greatly reduced, while the performance actually improved.
The transformation of the San Francisco startup was even more radical. After switching all traffic from Anthropic to DeepSeek, it saved millions of dollars in months, and core scenario performance improved instead of decreasing.
When a company‘s AI bill is about to exceed the total salaries of all employees, any claim to “technological lead” seems pale.
It‘s not that U.S. AI manufacturers are incapable of lowering prices; they simply don’t want to. In the past few years, venture capital and stock market bubbles have given them sufficient confidence to burn money to train models while covering costs with monopoly pricing. They assume that corporate customers will pay for the “most advanced” label by default, and the default market has no alternatives.
But the reality is that the vast majority of enterprise scenarios don‘t need the so-called “forward-edge capabilities” at all. The daily customer service response, document processing, and code writing are fully capable of being handled by the middle-end model.
So, when the Chinese model brings prices back to a reasonable range, the price bubble of American AI can no longer hold.
The second sore spot is the closed monopoly technology route, which is pushing customers out.
The leading AI manufacturers in the United States almost uniformly follow the closed-source path. Model weights are completely closed. Enterprises can only use API calls. They cannot customize, fine-tune, or deploy on their own servers.
This model essentially locks customers firmly into their own ecosystem, with data, computing power, and pricing power all in the hands of manufacturers. The longer it takes for companies, the higher the migration costs, and the weaker the bargaining power.
Meanwhile, Chinese AI manufacturers generally follow an open-weighted approach. Enterprises can download models and deploy them on-premises, train them twice based on their own business data, and have complete control over the flow of data.
For industrial enterprises like Siemens, this point is crucial. Industrial data involves core processes and business secrets, and it is impossible to upload all of it to third-party servers.
Models like DeepSeek can dive deep into industrial scenarios and adapt to specialized needs such as equipment parameters, process standards, and fault diagnosis, which is difficult for closed-source American models to do.
More ironically, in June this year, Anthropic followed a directive issued by the United States Department of Commerce to restrict access to its cutting-edge models for overseas users.
With a single order, a large number of companies in Europe, Japan, and Australia were directly cut off from supply, smart customer service was suspended, data analysis was interrupted, and many core businesses were paralyzed on the spot.
It was a cold blow to all companies using American AI around the world: depending on American AI, it could be cut off at any time for political reasons, and long-term availability could not be guaranteed.
In contrast, open-source Chinese models can be downloaded locally and run without the risk of being “stuck.”
The United States originally wanted to lock down technological advantages with regulations, but in the end, it pushed its customers towards its rivals. What it means to take a stone and hit your own foot, the American AI demonstrated once for the whole world.
When the United States turns AI into a “political tool,” companies seeking alternatives are the most rational choice.
The third affliction is the bubbleification of industry narratives, where technological progress is being fictionalized by the capital game.
In the past two years, the U.S. AI industry has been driven by capital, and all the stories revolve around “larger models, more parameters, higher computing power.”
It‘s as if by burning enough money and stacking enough graphics cards, one can build the future of the industry.
But the reality is that the marginal benefits of large models are rapidly decreasing, and the performance improvements resulting from parameter doubling are becoming smaller and smaller, while the costs are increasing exponentially.
Thus, the entire industry has fallen into a strange circle: in order to maintain the so-called “technological lead,” it must constantly burn money to make larger models; in order to cover the cost of burning money, it must constantly raise API prices; in order to maintain high prices, it must constantly amplify the importance of “forward-edge capabilities.”
But corporate customers are pragmatic. What they want is not the running-score champion in the laboratory, but the performance-to-cost ratio in the production environment. The market‘s choices are self-evident.
However, this bubble-making is depleting the long-term competitiveness of American AI.
Capital flowed towards a few closed-source giants, piling up the parameters of flagship models, instead squeezing the survival space of mid-range models, industry models, and open-source ecosystems.
China‘s AI took a different path: while pursuing cutting-edge performance, it focused on performance-to-cost, open-source ecology, and industry landing.
On the surface, the United States still holds the “top” crown, but the vast market beneath the crown is being eroded bit by bit.
Many people interpret this migration as “Chinese AI overtaking on the curve,” but it would be more accurate to say that American AI is beating itself.
It‘s not that the Chinese model has completely surpassed the U.S., but rather that the U.S. model itself has detached itself from the real needs of the market, pushing customers out proactively.
False high prices, ecological closure, and political bindings. None of these three problems are technological. They are all cumulative disadvantages of industrial routes and business models.
Although these companies are currently only shifting their mid-low-end scenarios to Chinese AI, the complex core tasks are still using the American model.
However, as the Chinese model continues to iterate, today‘s “low-end replacement” may become “middle-end flattening” tomorrow, and the day after that may impact the high-end market.
The reduction of the technological gap is accelerating, and once market trust is established, the speed of migration will be faster than anyone expected.
For the American AI industry, the real cold winter has not yet arrived. When the myth of “American AI = the best” is broken, when companies discover that they can meet demand without spending a lot of money, the balance of pricing power will completely tilt.
If the U.S. government continues to use political means to intervene in the market, it will only accelerate the de-embellishment process of global companies.
Ultimately, artificial intelligence is ultimately a service industry, not a display of skills, and certainly not a geopolitical weapon.
Whoever can provide customers with a higher price-performance ratio, more controllable solutions, and more stable service can win the market.
U.S. AI used to take full advantage of first-mover advantages, but if it was addicted to monopoly premiums and technological hegemony, refusing to lower its stature to respond to real needs, then today‘s collective “turned sides” of Silicon Valley giants is the prelude to the collapse of this hegemony.
When companies vote with real gold and silver, the AI myth of Silicon Valley begins a countdown.
Images from the Internet