For years, the AI race looked like a clear contest between a handful of leading US companies and everyone else. That gap is becoming harder to defend.
China’s AI industry is moving quickly, and the latest models are showing that being cheaper does not necessarily mean being far behind on performance. In some areas, Chinese models are already competing with the best systems coming out of the US.
That matters because the next phase of the AI race may not be decided simply by who builds the most powerful model. It could come down to who can make capable AI cheap, accessible and widely adopted.
China is closing the capability gap
The latest example is Kimi K3, developed by Chinese AI company Moonshot.
The model has performed strongly across several independent benchmarks, particularly in coding, reasoning and general-purpose tasks. In some tests, it is getting close to the performance of leading models from companies such as Anthropic.
The important part is not just the score.
It is the cost of achieving that performance.
Chinese AI companies have been aggressively competing on price, making their models attractive to developers and businesses that care about getting useful AI without paying premium prices.
That combination of improving capability and lower cost is putting pressure on the traditional US advantage.
The numbers tell an interesting story
On specialized software tasks, the leading US models still occupy the top spots.
According to the Terminal-Bench 2.1 leaderboard, models from OpenAI and Anthropic remain ahead, while Kimi K3 has entered the top tier.
The gap, however, is no longer enormous.
Kimi K3 scored 85%, compared with 89.5% for the highest-ranked model in the benchmark.
That is a meaningful difference, but it is also a very different picture from the one many people had a year or two ago, when Chinese models were generally viewed as a step behind the strongest American systems.
The bigger story is that China is catching up while offering much lower-cost alternatives.
Price could become China’s biggest advantage
AI development has become an expensive business.
Training frontier models requires enormous computing resources, while running those models at scale requires even more infrastructure. For companies building AI products, the cost of using a model can have a direct impact on margins.
This is where Chinese AI companies are becoming increasingly competitive.
Developers do not necessarily need the absolute best model available. They need a model that is good enough, reliable and affordable.
If a Chinese model can deliver most of the performance of a more expensive US model at a fraction of the price, developers have a strong reason to consider it.
And once developers build products around a model, switching becomes less straightforward.
Open-weight models change the game
There is another important difference.
Many Chinese AI models are released as open-weight models, allowing developers to download them and run them locally rather than relying entirely on a company’s cloud service.
That creates a much wider distribution network.
A developer can experiment with the model.
A company can deploy it on its own servers.
Another developer can modify it for a specific use case.
That makes controlling access much harder than it would be with a closed model that can only be accessed through an API.
It also helps explain why Chinese models are gaining traction among developers around the world.
Chinese models are already finding users
The adoption story is becoming difficult to ignore.
Chinese models account for the largest share of generative AI model downloads on Hugging Face, a major platform for hosting and sharing AI models.
Some US companies are also using Chinese models in their own operations.
Companies including Airbnb, DoorDash and Coinbase have adopted Chinese models hosted on local servers.
Alibaba has also said its open-weight AI models have surpassed 3 billion downloads in six months, putting its model family ahead of major competitors in terms of downloads.
That is an important signal.
AI leadership is not only about benchmark scores anymore. Distribution and adoption matter just as much.
The US has a difficult policy problem
Washington has already tried to slow China’s AI progress by restricting access to advanced US-made chips.
The logic is straightforward: powerful AI models require enormous amounts of computing power, and limiting China’s access to advanced chips should make it harder for Chinese companies to train frontier systems.
But the results are becoming more complicated.
Chinese companies have continued releasing increasingly capable models despite those restrictions.
Now there is another question: should the US also restrict access to Chinese AI models?
Nearly 200 US companies have reportedly opposed such a ban, arguing that it could increase their costs and make them less competitive globally.
There is also a practical problem.
You can restrict a company’s cloud service. It is much harder to restrict an open-weight model that can be downloaded and run locally.
The AI race is becoming a pricing war too
The conversation around AI has often focused on who has the most advanced model.
But the market is starting to ask a different question:
How much does that intelligence cost?
If two models can complete a similar task, the cheaper one can become extremely attractive.
That could have major implications for OpenAI, Anthropic and other US AI companies that are being valued on the assumption that their technological lead will translate into enormous future revenues.
Anthropic is reportedly considering an IPO as soon as October, while OpenAI is looking at going public next year.
Anthropic was valued at around $965 billion in its May funding round, while OpenAI’s most recent funding round valued it at around $852 billion.
Moonshot, by comparison, is valued at roughly $35 billion.
The valuation gap is enormous.
But so is the speed at which the competitive landscape is changing.
A real-world test offers another clue
Benchmark scores are useful, but they do not always tell us what happens when people actually use these models.
Bloomberg, working with AI evaluation platform Vals.ai, tested seven leading Chinese and US models by giving them the same task: build a fictional coffee e-commerce website called Brewberg.
Most of the models achieved 100% functional accuracy.
The differences were more visible in design and, importantly, price.
That is exactly where the AI market could become increasingly competitive.
If several models can successfully perform the same practical task, businesses may start choosing based on cost, speed, flexibility and ease of deployment rather than simply picking the model with the highest benchmark score.
DeepSeek already showed how quickly things can change
China’s AI progress is not happening in isolation.
In January 2025, DeepSeek’s R1 shocked the AI industry and financial markets by demonstrating that a Chinese company could produce a highly capable reasoning model while challenging assumptions about how much computing power and money were required to build advanced AI.
That release changed the conversation.
Kimi K3 is now another reminder that Chinese AI development is moving quickly.
The lesson for investors and businesses is simple: the competitive landscape can change much faster than expected.
There may not be one permanent winner
The biggest takeaway is that the AI race probably will not have a single winner for very long.
New models are being released at an extraordinary pace. A company can lead one benchmark today and lose that advantage after the next major release.
That makes the idea of a permanent US technological lead increasingly difficult to rely on.
The US still has enormous advantages in AI research, capital, infrastructure and leading companies.
But China has shown that it can move quickly, compete aggressively on price and distribute models at massive scale.
The two countries are now competing on more than raw intelligence.
They are competing on cost, accessibility, distribution, developer adoption and speed of innovation.
What this means for the AI market
For investors, this could become one of the most important shifts in the AI story.
The first phase was about building the most powerful models.
The next phase could be about making those models economically useful at massive scale.
That creates both opportunities and risks.
US AI companies may need to spend heavily to maintain their technological lead while also dealing with lower-cost competition from China.
Meanwhile, Chinese companies have an opportunity to gain global market share if their models continue improving and businesses keep adopting them.
The bigger question is no longer whether China can compete in AI.
It clearly can.
The question now is how quickly the gap will continue to narrow, and what that means for the enormous valuations being placed on America’s AI leaders.
The AI race is no longer a one-sided sprint.
It is becoming a much closer competition, and the next winner may be decided as much by economics and adoption as by technology.