The US AI ecosystem is not only galloping down a different (and, I believe, a wrong) hardware configuration road, but it is also barrelling down a different (and, I believe, a wrong) software road. China’s software model – open-weight, far more efficient, lower cost in energy use, data centre-light, quicker, leaner, better structured with its Mixture-of-Experts core supporting an Edge-compatible periphery, self-improving… this is what makes the Chinese LLM models so attractive to customers.
And oh yes, I forgot to add: THEY ARE 100% FREE FOR USERS TO ACCESS. No subscriptions are required as with most advanced US closed-weight models.
More advanced open-weight Chinese models are regularly being introduced, with upgrade costs running at a fraction of those incurred by new US closed-weight models. Take for example the August 2025-released ChatGPT 5: it took 29 months to build at an estimated $2.5-billion and produced a model that hardened users dismissed as “underwhelming”.
By contrast, the September 2025-released Qwen3-Omni took five months to build at a cost of $500,000 and is now, in its category, #1 on Hugging Face downloads. Collectively the various Qwen models – all produced by Alibaba’s Tongyi Lab – are the most widely used family of open-weight LLMs in the world today.
There are currently 12 thoroughbreds in Qwen’s stable, each trained to “run in a different race”, all free to use. Tongyi Lab does not just produce Qwen LLMs; it also designed the very popular Falcon 180B, a model which “comes” from the UAE.
Then there is the cost of the hardware support: ChatGPT 5 was built on Nvidia chips using 200,000 H100/B200 units, each costing $30,000, implying a total hardware cost of about $6-billion. Qwen3-Omni was built on 2,048 Huawei Ascend 910C chips, each costing $3,100, implying a total hardware cost of $6.35-million.
And then there are the comparative running costs: the inference cost per token for ChatGPT 5 ($0.001) is 100 times more expensive than Qwen3-Omni ($0.00001).
Here is the trillion-dollar question few in the US will dare ask: Why is China spending just $30-billion on data centres in 2025 when the US is spending $400-billion, with $1-trillion forecast to be spent before the end of 2027? After all, China is not short of money, as its infra-spend budget testifies. So why is China seemingly underinvesting in data centres?
The simple answer is China is not underinvesting. And this cuts to the core of the argument I am making. While markets are distracted by the prospect of a financial bubble, underneath the noise a technological arm wrestle is taking place… and China is winning. The blunt answer as to why China is building so few data centres is that the distributed design of the Chinese AI ecosystem requires far fewer data centres.
China’s AI ecosystem is, comparatively, very hardware light. Their open-weight LLMs like Qwen and DeepSeek do a far greater share of the “heavy lifting” in the overall Chinese AI ecosystem than does the associated Chinese hardware, with much of that lifting being done on the Edge. In the US, the hardware plays a far weightier role in their overall AI ecosystem, partly precisely because so little lifting is done on the Edge: rather it is nearly all done by the data centres and clouds back at the network’s core. China, by coupling its proto-brain and body in its hardware map with its open-weight LLM software, is “outbraining” the US.
This is not to say China is not investing elsewhere… and the most relevant other comparative statistic does not so much expose China as further blow the US’s cover. China’s big investment priority in 2025 remains energy transition with a forecast expenditure of $625-billion compared with the US’s $338-billion. Aha, you might say: China is “hiding” its AI expenditure in other areas… Not so when you read the assessment of the April 2025 report of the US Center for Strategic and International Studies, which concluded: “Today, access to electricity supply is the binding constraint on continued US leadership in AI.”
The assassin’s stiletto in the MIT Report is to be found in the answer to the question: “Why is US AI not working for US companies?” The MIT paper replied: “The core barrier to scaling is not infrastructure, regulation or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time.”
Why is this assessment so damning? Because Chinese LLMs do retain feedback, do adapt to context and do improve over time. No wonder 80% of entrepreneurs seeking finance from a16z are using Chinese open-source models!
I leave you with two comments that suggest that, under all the “there is a financial bubble in US AI” brouhaha, the most pervasive bubble in today’s US AI world is the mistaken belief that the US has technological superiority:
- Scott Galloway, NYU Stern: “Nvidia’s valuation assumes perpetual scarcity. But China has made AI abundance inevitable. This isn’t innovation. It’s a Ponzi scheme wrapped in FLOPS”; and
- Ben Thompson, Stratechery: “The US is betting on closed systems. China is winning with open ones. That’s not a technical gap. It’s a civilisational one.”
Fittingly, my last comment belongs to one of the two founding partners of Andreessen Horowitz, the firm which supplied the opening insight to this essay. a16z is a firm which – recall – maintains software is eating the world.
Marc Andreessen recently said: “We follow the Edge. Wherever it leads.” And today – even if a16z has no investments outside of the US – the Edge unequivocally leads to China… from where free-to-use Chinese open-weight LLM software operating off a proto-AI brain and body is rapidly eating the AI world. DM
*Read part one here and part two here.