The meat of the article is behind the paywall.

My theory is that

1. China has better integration between academia and industry - I’m mostly basing this off how in the US the academia is chronically underinvested so that the private sector can get it’s plunder of smart graduates, but when it comes to AI development you actually need academic knowledge of intelligence otherwise you’re just approaching it like an engineering problem

2. US companies are incentivized to spend more, the more broke they are the smarter they seem, so there isn’t really any incentive to improve in smart ways because you can just rack up a bigger bill with your investors and keep the circular economy going.

Would love to know what the recipe says though.

Being the pioneer means doing the heavy lifting and making mistakes. So makes sense, China can learn for others mistakes

Weird how the article doesn’t mention a key factor: the Chinese labs have really low headcount compared to their Western competitors, and organizationally they’re much more focused on model building than sidequests. Deepseek had only a couple hundred employees until recently, and accordong to leaks only had one or two people maintaining their consumer facing app.

i distilled this article

Summary of the article “How China gets better bang for its buck than America in AI” (Aug 3 2026)

• U.S. AI spending is massive – Bloomberg Intelligence estimates U.S. data‑centre capital outlays could exceed $740 billion in 2026, with Nvidia alone negotiating a $250 billion financing deal for a $500 billion data‑centre run by OpenAI. Alphabet announced a $205 billion AI budget.

• China spends far less – Chinese tech firms are projected to invest less than one‑tenth of the U.S. amount in data centres. Yet their models perform only slightly behind U.S. equivalents. For example:

  • K3 (Moonshot AI) scores ≈ 95 % of Anthropic’s Fable 5 on common benchmarks while being 70 % cheaper to run.
  • Alibaba’s newly released model ranks among the world’s best on certain metrics.

• Why Chinese spending is efficient

  1. Lower input costs – Land, construction, equipment and labour are cheaper in China.
  2. Model distillation – Chinese labs often train models using outputs from expensive U.S. models, reducing the compute needed.
  3. Hidden spending – Some expenditures on high‑end chips are masked as “cost‑saving” techniques that make inferior hardware achieve higher performance (e.g., DeepSeek’s efficiency tricks).

• Export restrictions limit Chinese capital use – U.S. bans on advanced AI chips (Nvidia designs, TSMC manufacturing) prevent China from buying the most powerful hardware.

  • Chinese firms are pushed toward domestic alternatives (Huawei, SMIC).
  • Sanctions also block access to cutting‑edge chip‑making equipment, forcing costly work‑arounds and capping production capacity.

• Domestic demand constraints – Chinese enterprises spend < 10 % of what U.S. firms spend on IT, despite China’s GDP being two‑thirds of the U.S. (or a third larger in PPP terms). This throttles revenue prospects for AI providers, curbing their willingness to invest heavily.

• Strategic focus differs – The Chinese Communist Party emphasizes diffusing AI across the economy, not pursuing a race toward artificial general intelligence (AGI). Fewer than ten Chinese firms target AGI, compared with dozens of U.S. players.

• Investor attitudes – Chinese investors have historically punished over‑spending on AI, whereas U.S. investors once rewarded aggressive budgeting. This cultural difference keeps Chinese AI budgets modest.

• Potential bottlenecks for China – Despite restraint, China may face compute shortages:

  • ByteDance experiences ten‑hour processing times for some videos.
  • Alibaba Cloud, Zhipu AI, and Moonshot’s K3 have long waiting lists or quickly sell out capacity.
  • Over‑restriction could stifle growth if AI services cannot meet user demand.

Overall takeaway: China achieves comparable AI performance to the U.S. while spending a fraction of the capital by leveraging cheaper resources, model‑distillation techniques, and a strategic focus on wide‑scale diffusion rather than raw computational power. However, export bans, limited domestic chip capacity, modest corporate demand, and cautious investors together create both an efficiency advantage and a risk of under‑provisioned infrastructure.