Bengaluru – On a video conference call from Hangzhou earlier in 2026, one of China’s hottest start-ups held a four-hour pitch meeting with potential venture investors that was unusual by almost any measure. Only two representatives per institution were allowed. The majority of those present were getting a glimpse of the company’s leader for the very first time.
“We’re a group of very ordinary people,” said Liang Wenfeng, DeepSeek’s elusive founder, according to some of the attendees.
His comment belies the extraordinary status of his company, a Chinese start-up that has shaken up the global tech industry with powerful models trained on a fraction of the resources used by OpenAI and Anthropic PBC. Liang ultimately raised US$7.4 billion (S$9.6 billion) in the largest private AI financing in Chinese history, valuing the three-year-old lab at more than US$50 billion. Just weeks later, DeepSeek is plotting an initial public offering in 2027 that will likely bring in billions more.
Liang is now poised to lead a fresh, potentially more threatening, challenge to Silicon Valley. With coffers full, DeepSeek has more resources to develop cutting-edge AI services and offer them globally at prices far below US rivals. Chinese low-cost, open-source services risk undercutting the business prospects for OpenAI and Anthropic in overseas markets – and increasingly in the United States.
China’s open-source models are already far cheaper than Silicon Valley rivals. The weighted average cost of performing a standardised intelligence task using DeepSeek’s V4 Flash model costs just 2 US cents, compared with US$2.75 for the same task on Anthropic’s Claude Fable 5, according to benchmarking site Artificial Analysis. DeepSeek’s fundraising signals that pricing and the start-up’s reach will only grow more aggressive.
For corporate customers, the math is undeniable. San Francisco-based Lindy AI had been using Anthropic’s Claude Sonnet, but switched to DeepSeek after six weeks of evaluation.
“We’re now paying about 10 per cent of what we used to pay,” said Flo Crivello, the company’s chief executive officer, adding the start-up is saving millions of dollars annually – more than the total cost of its entire 27-person workforce.
Big companies that jumped into AI in recent years are growing increasingly cost conscious. Corporations like Tesla and Meta Platforms that set up leaderboards to reward employees for “tokenmaxxing,” or using as many AI tokens as possible, are discovering that unbridled spending on AI doesn’t really pay off.
Now, they’re looking to control expenses and get a return on their money – a newfound practicality that’s leading them to services like DeepSeek and Zhipu. Chinese-built open-source models have surged to more than 30 per cent of the market on the AI aggregation platform OpenRouter some weeks in 2026, compared with less than 2 per cent in 2024.
The converts include some of the best-known companies in the US. Airbnb runs a mix of Chinese open-source models along with Western vendors. Pinterest touted their utility. Even Microsoft – one of OpenAI’s biggest backers – is in. CEO Satya Nadella said he’s considering using DeepSeek models in its Copilot Cowork product. Hundreds of US firms are quietly reserving premium models for the hardest reasoning tasks while routing simpler coding, automation and routine workloads to far cheaper Chinese alternatives.
TinyFish, a Palo Alto, California-based start-up that builds enterprise infrastructure for AI web agents, used to spend six figures a month on AI coding tools. Then it switched from frontier models to alternatives including Chinese open-weight models. Inference costs dived 90 per cent.
“This kind of bill is actually a wake-up call for a lot of companies,” said Chief operating officer Keith Zhai.
DeepSeek fits China’s broader tech agenda. Beijing has said it wants artificial intelligence to be low cost and accessible so it can be adopted quickly throughout the economy. It has pushed back on the prospect of proprietary AI models like Anthropic and OpenAI, which could reach trillion-dollar valuations as they seek to go public in the months ahead. Chinese AI leaders including Zhipu founder and Tsinghua University professor Tang Jie have argued that frontier AI should remain broadly accessible rather than controlled by select individuals – a view echoed by the state-backed People’s Daily.
Chinese AI models are also backed by government-supported efforts in everything from affordable energy to local semiconductor players, including Nvidia-rival Huawei Technologies and memory chip kingpin CXMT. The Chinese alternative to global AI services is aimed at supplying Chinese companies and government agencies with dependable technology at reasonable costs, with the potential to flood the world with cheaper AI that could undercut American rivals.
“DeepSeek is not simply one company,” said Tan Yinliang, a professor at Shanghai’s China Europe International Business School. “It is better understood as part of China’s broader effort to industrialise AI: models, chips, cloud infrastructure, applications, and deployment ecosystems moving together.”
Energy is emerging as a significant competitive advantage. China has shown it can build new power generation at a speed and scale unmatched anywhere in the world. In 2025 alone, it built 543 gigawatts of generating capacity, more than the total capacity of any country in the world outside of itself and the US.
While China cannot export that cheap electricity, it can leverage its massive domestic power to run local data centres and fuel a booming market for AI tokens.
“Maybe China doesn’t have the best compute, but China has lots of electricity,” said Robert Wu, chief executive officer of BigOne Lab, a company that tracks technology trends. “At the end of the day, the cost of AI will just go down and down and down to the cost of energy. China has a clear advantage in that.” BLOOMBERG
