Chinese AI labs are finding a cheaper way to close the gap

Chinese AI labs are narrowing the gap with US rivals by squeezing more useful work from less computing power. That matters because the contest is shifting from who can buy the largest pile of chips to who can make each token earn its keep.

DeepSeek Watch analysis

What happened

A Fortune report describes DeepSeek and other Chinese labs developing more efficient approaches to the attention mechanism used by large language models. The report links that push to tighter access to leading US chips and says the result is a different competitive advantage: models that may be cheaper to run while remaining good enough for a large share of business work.

Fortune cites Larridin co-founder Ameya Kanitkar saying Chinese models such as GLM and Kimi handled around 75% of the enterprise engineering tasks tracked by the company reasonably well, at about one-fifth of the cost of US alternatives. The report also notes that US frontier models still lead on the hardest tasks, so this is not a clean changing-of-the-guard moment. It is more interesting than that.

Key findings

• Efficiency over brute force
  Chinese labs are reported to have reduced the computational cost of attention rather than simply adding more hardware.
• Lower operating costs
  One analysis cited by Fortune put several Chinese models at roughly a fifth of the cost of US models for tracked engineering work.
• Open-weight flexibility
  Downloadable models can be run and adapted by organisations that do not want every prompt routed through a closed service.
• US models still lead at the frontier
  The report says the most complex tasks remain a US strength, even as everyday enterprise work becomes more competitive.

Why it matters

For companies buying AI, the useful question is no longer just which model tops a benchmark. It is which model delivers acceptable results at a cost the finance department will not describe as an interpretive art project.

Cheaper, adaptable models could make local or hybrid deployment more attractive, particularly where data control matters. They could also put pressure on US labs to improve efficiency rather than treating ever-larger training runs as the default answer.

Our read

This is a meaningful shift in the AI race, but not proof that Chinese models have overtaken the strongest US systems. The practical takeaway is to test models against your own workload, measure cost per successful task and keep open-weight options in the comparison set.

What to watch

• Whether the reported efficiency gains survive independent testing across more tasks.
• Whether DeepSeek and other Chinese labs continue releasing models that organisations can run and adapt.
• How much the hardest reasoning and agentic workloads still favour leading US systems.
• Whether chip restrictions accelerate new algorithmic gains or simply slow access to larger models.

Discussion spark: If a cheaper open-weight model handles most of your organisation’s routine work, how much frontier-model performance would you still pay for?

not affiliated with or endorsed by DeepSeek