The game-changing US AI Action Plan unveiled by President Donald Trump last week basically opens up exports of US-made AI chips to all US allies in a bid for global AI dominance. The US strategy is simple: disperse US-made advanced AI chips and software to US allies as fast as possible so that China will not be able to catch up.
But guess what? China has apparently caught up, to some extent.
According to Saanya Ojha, Partner at Bain Capital Ventures, in her weekly newsletter last Friday, China is leading in open source AI models. “This month, China shipped the two best open-source LLMs released to date,” she said.
- Moonshot’s Kimi 2: A 400B+ MoE model with 2M-token context (proprietary) and 128K public window.
- Alibaba’s Qwen3: Smaller, faster, and top of the leaderboard across MMLU, GSM8K, ARC, and HumanEval.
Ojha added, “They’re better than anything the West has open-sourced.
According to her, China’s AI strategy is diverging fast:
- It’s building from scratch, not fine-tuning Western models;
- It embeds AI in superapps, not chatbots;
- It builds for efficiency and deployment, not just benchmark glory.
- And it does all this under tight chip constraints, strict regulation, and with intense state-enterprise coordination.
“While the West chases AGI, China is quietly operationalizing AI across logistics, finance, education, and government. If you’re only tracking the western labs, you’re missing half the map. And the half you’re missing is moving fast.”
Nvidia’s CEO Jensen Huang has earlier said that AI scientists in China are world-class and overall, China is doing “fantastic” in the AI market, with models from Chinese-based companies — such as DeepSeek and Manus — emerging as powerful challengers to systems designed in the US.
Trump must have been apprised of this – hence the dramatic change in tactics. If restrictions don’t work, then how about deluging the world with US-made AI chips and software? If most of the world use US-made AI ecosystems, then US wins.
But the AI race isn’t over yet.
According to Ojha, three players are sprinting to build the U.S. backbone of AI, namely:
- OpenAI is orchestrating a hyperscaler stack through Oracle and CoreWeave, trying to control the AI supply chain without owning it. It’s scaling through partnerships, not property.
- xAI is building everything in-house – chips, data centers, power – on a mountain of debt, betting that speed and vertical control outweigh risk.
- Meta is quietly out-building both – committing over US$ 100 billion in capex and constructing the 5 GW Hyperion campus. Its models run inside search, feed, Ray-Bans, and WhatsApp – embedding intelligence at the edge.
And not to mention Meta’s multimillion-dollar Superintelligence Labs team which has the ability to build ChatGPT from scratch. Chief scientist of the team, Shengjia Zhao, was one of the co-creators of ChatGPT.
Ultimately perhaps, it doesn’t matter who wins the AI race as long as the whole world receives the benefits of AI. As it is, half the world’s population or more, are not even connected to the internet yet.
