DeepSeek, a Chinese company, was the first to demonstrate the expected AI cost reductions, which had never before happened with a Chinese company.

The recorded date (2022-10-13) appears inconsistent with the subject, since DeepSeek's V3/R1 releases came in late 2024 and early 2025. Judge the claim as of when DeepSeek's results became public (around January 2025). The claim is that DeepSeek was the first company to show the expected trend of falling AI training or inference costs at its level of capability, and that this was new in that earlier such firsts had not come from Chinese companies. True: credible analyses show DeepSeek's V3/R1 were the first to reach near-frontier performance at the expected lower cost point, and earlier cost-curve firsts were by non-Chinese (mainly US) labs. Mostly true: broadly so, but DeepSeek was among the first rather than clearly first, or the scale of its cost reduction was somewhat overstated. Mixed: comparable cost reductions had been shown around the same time by other labs, or by earlier Chinese labs. Mostly false: DeepSeek's cost reductions were real but had clearly been shown first by others. False: DeepSeek did not show the claimed cost reductions, or a Chinese company had clearly done so before.

Set before the research began.

1. E1 Anthropic CEO downplays DeepSeek models, calls for more Chinese chip sanctions DatacenterDynamics

   “that means it would be totally normal, totally 'on trend,' if DeepSeek-V3 training cost ~8x less than the current US models developed a year ago.”

2. E2 Anthropic's Dario Amodei on AI Competition ChinaTalk · 5 February 2025 Own record

   “Another was that the model was an efficient and strong model on the cost reduction curve, but it’s the first time the point on the cost reduction curve was made by a Chinese company.”

3. E3 DeepSeek and the generative AI perception dilemma DatacenterDynamics

   “Even if DeepSeek is telling the truth, the reality is the model did not cost $5.6m - that is only the cost of the final run.”

4. E4 Anthropic's Dario Amodei on AI Competition ChinaTalk · 5 February 2025 Own record

   “the model was an efficient and strong model on the cost reduction curve, but it’s the first time the point on the cost reduction curve was made by a Chinese company.”

5. E5 Anthropic's Dario Amodei on AI Competition ChinaTalk · 5 February 2025 Own record

   “This wasn’t making something for 6 million that cost the other companies billions. This was more like we’ve seen a 4x cost decrease per year — roughly that, give or take, relative to models that had been trained half a year to a year ago.”

6. E6 GitHub - deepseek-ai/DeepSeek-V3 GitHub (deepseek-ai repository) · 26 December 2024

   “At an economical cost of only 2.664M H800 GPU hours, we complete the pre-training of DeepSeek-V3 on 14.8T tokens, producing the currently strongest open-source base model.”

7. E7 GitHub - deepseek-ai/DeepSeek-V3 GitHub (deepseek-ai repository) · 26 December 2024

   “DeepSeek-V3 achieves the best performance on most benchmarks, especially on math and code tasks.”

8. E8 How much does it cost to train frontier AI models? Epoch AI · 3 June 2024

   “The cost of training frontier AI models has grown by a factor of 2 to 3x per year for the past eight years, suggesting that the largest models will cost over a billion dollars by 2027.”

The only direct support for the claim's central point, that the first company to reach this point on the AI cost-reduction curve was Chinese, is Amodei's own interview remarks (E3, E5). DeepSeek's README (E7, E8) gives a pre-training compute figure and performance claims, but it compares DeepSeek-V3 with neither earlier labs' cost-performance points nor earlier Chinese labs. E4 says the $5.6m figure covers only the final training run, and E9 covers frontier training costs before DeepSeek. Neither shows who first reached the expected lower cost at near-frontier capability. No independent analysis in the evidence confirms or refutes the claim that DeepSeek was first, or that earlier such firsts came from non-Chinese labs.