Analysis: China’s DeepSeek Proves Limited Resources Are No Barrier to Catching Up With U.S. AI Rivals

05 Feb 2025

By Liu Peilin and Kelsey Cheng

Chinese AI company DeepSeek has been the talk of the town since launching its open-source AI model DeepSeek-R1 on Jan. 20. Photo: VCG

As Chinese artificial intelligence (AI) startup DeepSeek shakes the global AI industry with its powerful and cost-effective model, domestic industry insiders attributed the firm’s breakthrough to its top-notch engineering and “fast follower” advantage.

Hangzhou-based DeepSeek launched its open-source AI model R1 on Jan. 20, with benchmark tests showing that it can go toe-to-toe with OpenAI’s o1 model launched in September and Meta Platform Inc.’s Llama 4, scheduled to be released this year.

The new product is an upgrade from DeepSeek-V3, a 671-billion-parameter model launched in late December which required just 2,048 Nvidia H800 chips and $5.58 million to train, according to the company’s technical paper, many times less than what OpenAI and Meta spent to train similar-sized models.

Nvidia Corp. designed the H800 graphics processing units (GPUs) specifically for the Chinese market after the U.S. banned advanced semiconductor exports in October 2022. However, the chips were subsequently restricted by Washington in October 2023, prompting Chinese firms to race to use the available chips as efficiently as possible.

Describing DeepSeek-V3 as “a frontier-grade (large language model) trained on a joke of a budget,” Andrej Karpathy, one of OpenAI’s founders, said on social media platform X that it was believed a model of this scale should require at least 16,000 GPUs, while the most advanced models typically consist of around 100,000 GPUs.

Based on preliminary tests, Karpathy said that the model is “a highly impressive display of research and engineering under resource constraints.”

Founded in 2023 by Liang Wenfeng — who also established quantitative hedge fund High-Flyer Quant in 2015 — DeepSeek’s ability to achieve top-notch performance at a fraction of its rivals’ costs was hailed by many as a testament to Chinese tech companies’ resilience in the face of U.S. restrictions and their increasing maturity.

Significantly, its achievements have put its rivals in an awkward position — if DeepSeek was able to spend just a few million dollars to produce V3, why do they need to spend hundreds of millions to do the same?

‘Follower strategy’

DeepSeek has benefited from its fast-follower advantage, a domestic AI training and deployment software entrepreneur said. “It has been repeatedly proven that large models are highly vulnerable to replication. The first innovator pays a high price, while followers can replicate 90% of the results at just 1% of the cost. And the fast-follower strategy remains effective in the long term,” he told Caixin.

“Building a 1-million-GPU cluster like in the U.S. is nearly impossible for us, but the follower strategy ensures we won’t lag too far behind,” he said.

Later entrants benefit from the trial-and-error efforts of predecessors, leading to lower training costs for comparable models, said one technical executive at a leading Chinese LLM company.

But he also asserted that DeepSeek’s reported $5.58 million cost only accounts for a single training session. A complete training cycle includes additional steps such as pre-experimentation, as well as data generation and cleaning, which could easily double the cost, he said.

While resource-rich OpenAI can explore multiple technical routes, Chinese large model companies cannot afford to take risks and are therefore forced to follow, according to AI startup StepFun founder and Microsoft veteran Jiang Daxin.

“OpenAI doesn’t share its methods, but when they release something, it proves the approach works,” Jiang previously told Caixin. “Our strategy is to catch up within six months: first by following, then securing our position, and finally innovating by integrating applications.”

The follower strategy works in China’s favor as this round of AI innovation, driven by feeding as much data as possible into models, has begun to culminate.

OpenAI co-founder and former chief scientist Ilya Sutskever said at the NeurIPS 2024 conference in December that data was the “fossil fuel” of AI — arguing that its supply is finite and close to exhaustion. Sutskever suggested that new approaches are needed to achieve the next level of AI advancement.

Efficient use of resources

Chinese AI companies like DeepSeek have been focusing on optimizing their algorithms to train models with fewer resources. These strategies have allowed Chinese firms to achieve training and inference costs at least an order of magnitude lower than their international peers.

The inference cost of an AI model is the expense incurred each time the model processes data to generate a response, typically measured in cost per million tokens for large language models.

The inference cost of DeepSeek’s latest R1 model is approximately 1/25th to 1/50th of OpenAI’s pricing, according to Caixin’s calculations.

“In the era of large models, with computational resources much more limited than in Silicon Valley, Chinese companies must have strong AI infrastructure capabilities,” said Kai-Fu Lee, former head of Google China who founded AI developer 01.AI in March 2023.

Lee believes Chinese large model companies have advantages in cost, labor efficiency and resource utilization. Although 01.AI’s funding is less than 10% of OpenAI’s, its training costs are only 3% of OpenAI’s, and inference costs were 1/40th, he told Caixin in a recent interview.

Lee also predicted in his 2018 book “AI Superpowers” that while the U.S. will lead breakthroughs, China will boast superior engineering. The argument was met with much criticism at the time, Lee said in a LinkedIn post Monday.

“With the recent DeepSeek releases, I feel vindicated,” Lee said.

Han Wei contributed to this story.

Contact reporter Kelsey Cheng (kelseycheng@caixin.com) and editor Joshua Dummer (joshuadummer@caixin.com)

caixinglobal.com is the English-language online news portal of Chinese financial and business news media group Caixin. Global Neighbours is authorized to reprint this article.

Image: Monster Ztudio– stock.adobe.com