Has the AI race shifted from U.S. vs China to open vs closed? — Fortune

RESOLVING THE LEDGER

Has the AI race shifted from U.S. vs China to open vs closed? — Fortune | AIPROPX

Another week, another new innovation from China’s AI labs. DeepSeek just released a new version of its V4 Flash model, which outperforms most top Western models, is priced far more cheaply than even its competitors, and is small enough to run on cheaper hardware.

Whether it’s GLM-5.2, Kimi K3, or DeepSeek V4, China’s open-source AI models are reshaping how we think about AI . In order to maintain its strategic technological advantage, the U.S. imposed strict export controls on China to limit global access to advanced compute and throttle foreign artificial intelligence development. Yet these policies instead acted as a massive stimulus for innovation. Denied unlimited access to top-tier hardware, global labs were pushed to optimize their algorithms and adopt open-source architectures.

Regulatory pressure effectively birthed a new generation of highly efficient, low-cost models that are now achieving capability parity with premium, closed systems.

This paradox shifts how we view the global technology landscape. The mainstream narrative loves to pit U.S. AI against Chinese AI. But the true race is really between open and closed.

Instead of looking at who is building the AI, we should examine who the AI is being built for, and how it’s being used—and, more importantly, whether world-leading powers can find common ground in embracing open-source technology.

Closed model labs have led on capability and benchmarks for years, but DeepSeek, Moonshot AI, and others are showing that open-source models may be no more than a few months behind. Then just a few days ago, DeepSeek pushed the question of whether frontier labs are really worth the price back to the forefront again.

According to independent evaluation platform Artificial Analysis, DeepSeek V4 Flash is only one Intelligence Index point behind GPT-5.6 Luna, and even after OpenAI’s 80% price cut, the new Chinese model’s cost per task is still 60% lower. So, from a business perspective, why would anyone want to pay more for similar-level quality?

U.S. commentators have accused Chinese labs of distilling frontier models , arguing it’s the only way Chinese labs could keep costs so low and performance so good. Others are going so far to suggest that the open-source approach is akin to China “dumping” low-cost models on the U.S., hoping to drive the labs out of business.

The truth is that U.S. policies forced Chinese AI companies into the open-source, low-cost model. Barred from accessing top-tier GPUs, Chinese labs had to innovate at the architectural level rather than brute-forcing scale.

China’s move towards open-source was not the result of some grand strategy, but instead emerged from how private companies adapted to hardware constraints. By releasing model weights, these firms could draw on the global AI research community to improve their systems more quickly. It also reduced the need to invest heavily in their own computing infrastructure. Instead of building and running costly GPU clusters, they relied on overseas cloud providers to handle much of the inference workload. This strategy helped Chinese developers gain global visibility while shifting much of the expense to Western infrastructure providers.

There’s a misconception that open-source models don’t make money. But that’s not the case. Users are still paying API providers for managed service, or paying the likes of Groq or Fireworks for managed inference. The monetization model is similar to open-source software: Paying for a managed service. Most users don’t want to self-host anyway, because that would require GPUs, security, monitoring, and maintenance.

Worries about data being sent to China don’t hold either. When you self-host or route through inference providers in the U.S., the data and API traffic stay within the U.S. The only reason to ban open-source models that are inching towards frontier-level capabilities may be an anti-competitive one. Open-source threatens the business model of the frontier labs: Charging a huge premium for “higher intelligence.”

An additional ban on open-weight models would hurt the U.S. more than China: U.S. companies will have to pay a premium for intelligence that those outside the country can get much more affordably.

The narrative is decisively turning against U.S. closed-model labs. In the past two weeks, U.S. AI labs and tech executives now argue that open source is the way forward. Former “AI czar” David Sacks and his fellow venture capitalist David Friedberg now point out how even Google first distilled Yahoo as it iterated its search product. CEOs rushed to join Jensen Huang’s call to support open-source models. Anthropic hasn’t signed onto Huang’s letter, but it’s still changed its tune on open source. It now says its main concern is safety, rather than its usual claims about IP theft.

If we take the concern that open-source models can be misused by bad actors at face value, then it makes even more sense for the U.S. and China to collaborate and embrace open-source. The May Trump-Xi Summit created space for the two sides to come back for safety guardrail discussions to prevent non-state bad actors from misusing advanced AI. With the coming September Xi visit to DC, there’s clear impetus to resume dialogue and find mutually agreeable models for cooperation on AI safety guidelines, evaluation models, and even long-term governance institutions.

Both countries will continue to pursue their own national security programs around AI. But for civilian AI use cases, they can view the technology as a global public good. This could reduce the arms race narrative that will suck up a trillion dollars this year. (In the U.S. alone, from 2026, expected capex spending by big tech will reach $1 trillion based on company filings from the Magnificent Seven.)

If open source were to continue to catch up but at a fraction of the cost to users, then diffusion adoption of open source will continue to rise, as recent adoption...