The three pressures that broke DeepSeek's self-funding model

For three years, DeepSeek, the Chinese AI lab founded by quantitative hedge fund manager Liang Wenfeng, turned away every outside investor, including Tencent, the Chinese internet and gaming conglomerate, and Alibaba, the e-commerce and cloud giant. Now Reuters and The Information report that Liang is in active talks with both at a valuation exceeding $20 billion. Something wore him down.

Three things broke at once. SemiAnalysis, a chip and infrastructure research firm, estimated DeepSeek's total server capital expenditure at nearly $1.6 billion, with over $900 million going to operating the compute cluster. That's a number that outlasts most hedge fund thesis cycles. At the same time, five researchers who built DeepSeek's core systems left for rivals: Luo Fuli, who designed the V2 model architecture (the structure of the underlying neural network), went to Xiaomi, the Chinese smartphone maker; Wang Bingxuan, lead author of the first-generation large language model, went to Tencent; Guo Daya, who developed GRPO, a training algorithm central to the reasoning model R1, went to ByteDance's Seed AI team (ByteDance is the parent company of TikTok); Ruan Chong, who led multimodal research (combining text with images and other inputs), went to Yuanrong Qixing, a Chinese autonomous-driving startup; Wei Haoran, who worked on OCR (optical character recognition), is reportedly leaving, according to the South China Morning Post and 36Kr. 36Kr also reported that DeepSeek's next major model, V4, was originally planned for February 2026 and has been postponed repeatedly, partly because the team is migrating from Nvidia's CUDA software framework (the standard toolkit for building AI models) to Huawei's Ascend chip architecture, which requires rewriting foundational training code.

The self-funding model had a specific logic. High-Flyer Capital Management, Liang's quant fund and the sole owner of 99 percent of DeepSeek, insulated the lab from commercial pressure by design: no investors meant no board, no quarterly targets, no push to monetize. That logic held while training costs stayed manageable.

The $5.6 million training cost for V3, run on 2,048 Nvidia H800 GPUs, became a famous number when it surfaced in January 2025, suggesting DeepSeek could do far more with far less. What the number obscured was the infrastructure bill underneath it. The $5.6 million covered one training run; the $1.6 billion in server capex is what made runs like that possible. The cluster doesn't pay for itself.

The funding talks have a specific wrinkle. Reuters reported that Tencent proposed acquiring up to a 20 percent stake, but DeepSeek is resisting that level of control. Wang Bingxuan, one of the five departed researchers who helped build the original model, now works at Tencent. The company seeking 20 percent of DeepSeek is already employing one of the engineers who built it.

For context on the $20 billion valuation: OpenAI is at $852 billion, Anthropic at $380 billion. Alibaba's Qwen family of open-source models commands over 50 percent of global open-source model downloads, with more than 940 million cumulative downloads as of March 2026. Both Tencent and Alibaba have competitive reasons to want a stake in DeepSeek that go beyond the financial return.

No terms are final. Reuters reported the initial target was at least $300 million at a $10 billion valuation; investor interest has since pushed discussions above $20 billion. 36Kr noted that the raise may partly be about establishing a market-based valuation anchor for employee stock options, not just capital. What hasn't changed is the tension: whether Liang can raise enough to keep the lab running without giving the companies he disrupted a seat at the table.