From quant fund to AI lab: how Liang Wenfeng's High-Flyer gave rise to DeepSeek

Long before DeepSeek released its first model, the hedge fund behind it was running stock trades chosen by neural networks on GPUs. Liang Wenfeng founded the quantitative trading firm High-Flyer in 2015 and launched DeepSeek, the Hangzhou-based AI company, in 2023, partly funding it with money from High-Flyer, according to Forbes. In between, High-Flyer built two computing clusters of its own and ran them on software it developed in-house. That path, from trading models to a lab that publishes its papers and model weights, is an early case of the overlap between quant trading and AI research.

Liang holds a bachelor's and a master's degree in engineering from Zhejiang University, Forbes says. The university wrote in March 2025 that Liang had completed a master's thesis at its College of Information Science & Electronic Engineering fifteen years earlier, and quoted its acknowledgements in English. In the university's rendering, Liang credited the supervisor, Xiang Zhiyu, with an introduction to "the field of machine vision". Forbes says Liang set up High-Flyer with two college classmates and used maths and AI techniques to make it "one of the country's largest and best-performing quantitative trading firms". Polstob found no verifiable published figures for High-Flyer's returns or assets under management.

The engineering behind Fire-Flyer II is documented in a 2024 paper by DeepSeek-AI researchers, with Liang among the co-authors and listed as its submitter on arXiv. The authors write: "For DL training, we deployed the Fire-Flyer 2 with 10,000 PCIe A100 GPUs, achieved performance approximating the DGX-A100 while reducing costs by half and energy consumption by 40%." Those are the paper's own comparisons with Nvidia's DGX-A100 systems. The team also wrote its own software for the cluster: High-Flyer's history page lists a parallel file system (3FS) and a communication framework for distributed training (hfreduce), and the paper credits its software stack with letting computation and communication overlap.

DeepSeek's V3 technical report, first posted on 27 December 2024 with Liang as submitter, describes a mixture-of-experts language model with 671 billion parameters in total, 37 billion of them active for each token, pre-trained on 14.8 trillion tokens. The report says the full training took 2.788 million H800 GPU hours, and that each trillion tokens of pre-training took 3.7 days on a cluster of 2,048 H800 GPUs. DeepSeek published the model checkpoints on GitHub.

High-Flyer's money got DeepSeek started, but the company has since taken outside capital, according to news reports that DeepSeek has not confirmed. A Bloomberg report carried by The Straits Times on 15 July 2026 described a US$7.4 billion round in June 2026 that valued DeepSeek at US$50 billion, with Liang investing US$3 billion personally. On 22 October 2026, Reuters cited two people as saying a new round had raised about 100 billion yuan; one of them named CATL and Tencent among the biggest investors. Forbes puts that round at US$14.9 billion, valuing DeepSeek at US$74.7 billion.

On 15 April 2026 CoreWeave announced that Jane Street had committed about $6 billion to use its AI cloud platform and had separately bought $1 billion of CoreWeave Class A shares at $109.00 each; CoreWeave's 8-K filing records 9,174,311 shares issued to Jane Street Global Trading, LLC. In September 2026, Reuters reported, citing Bloomberg News, that Jane Street had also signed a five-year cloud contract with Crusoe worth around $13 billion. The figure is unconfirmed. In CoreWeave's April release, Max Hjelm, CoreWeave's Senior Vice President of Revenue, said Jane Street "operates like a frontier lab".

Hudson River Trading (HRT) signed a multi-year agreement with CoreWeave, announced on 20 August 2026, to run AI-driven trading research on Nvidia Vera Rubin systems. CoreWeave called it a multi-billion dollar deal but gave no value or term. In May 2026, Dell said HRT had recently opened a data centre for AI research at Lefdal Mine Data Centers in western Norway. Dell's release gives no investment figure.

XTX Markets, an algorithmic trading firm, says on its website that its research cluster has more than 25,000 GPUs and more than 1 exabyte of usable storage. In January 2025 it announced plans to invest over €1bn in a data centre complex in Kajaani, Finland, with the first 22.5MW centre due for completion in 2026, according to the City of Kajaani. Its chief technology officer, Joshua Leahy, said then: "Our need for compute has outgrown available leasing options."

Jane Street describes its work in terms of large, complex models trained on noisy data, and XTX says its teams include people with backgrounds in mathematics, physics, programming and machine learning. High-Flyer built its clusters and software in-house before Liang launched a lab that publishes its methods and releases its model weights. Like High-Flyer, which built Fire-Flyer I and II, the trading firms are lining up compute: Jane Street and HRT through cloud deals with CoreWeave, the latter also through the data centre it opened at Lefdal in Norway, and XTX through the data centre it is building in Finland.

Much remains undisclosed. High-Flyer's returns and assets are not published in any source Polstob could verify, DeepSeek has not given a full cost for building its models, and the trading firms' contracts are known only in part: CoreWeave gave Jane Street's commitment but not HRT's, and the Crusoe figure rests on Bloomberg's sources.