DeepSeek's Founder Just Bought Into China's Memory Chip Champion

The Chinese internet is obsessing over a five-character financial headline that tells you exactly where China's AI industry is heading next. Liang Wenfeng (梁文锋) — the reclusive billionaire founder of DeepSeek (深度求索), the AI lab that shook Silicon Valley in January — has personally subscribed to shares in Changxin Memory Technologies (长鑫科技), China's domestic DRAM champion and the country's best hope for breaking Korea's stranglehold on memory chips.

The Toutiao hot board lit up with 23.7 million engagements on this story, and for good reason. This isn't some random billionaire buying lottery tickets in the stock market. This is the most-watched man in Chinese AI making a very public bet that the next bottleneck in the AI arms race isn't software — it's silicon.

Let me set the scene for you.

Liang Wenfeng doesn't give interviews. He doesn't do Davos panels. He barely exists on Weibo (微博). The 40-year-old Zhejiang native built his fortune running High-Flyer (幻方量化), one of China's largest quantitative hedge funds, before pivoting hard into AI with DeepSeek in 2023. When DeepSeek R1 dropped on January 20th of this year and proceeded to make OpenAI's o1 look overpriced and lazy, Liang became the closest thing China's tech world has to a folk hero — the guy who proved Chinese engineers could match American AI labs at a fraction of the compute cost.

So when this guy moves his money, you pay attention.

Changxin Memory Technologies — often known as CXMT — is China's only serious domestic DRAM manufacturer. Founded in 2016 and based in Hefei, the company has burned through an estimated $10+ billion in state and private capital trying to catch up with Samsung, SK Hynix, and Micron, the three companies that control roughly 95% of the global memory chip market. CXMT's current DDR4 and LPDDR4X products are still a generation or two behind the Koreans, but they're good enough to land in Chinese smartphones, servers, and — critically — AI training clusters.

And that's the connection everyone on the Chinese internet is making right now.

The logic is brutal and simple.

Training large language models requires absurd amounts of memory bandwidth. HBM — high-bandwidth memory — is the secret sauce behind Nvidia's H100 and H200 GPUs, and it's almost entirely controlled by SK Hynix and Samsung. Huawei's Ascend (昇腾) AI chips, China's domestic alternative to Nvidia, are hobbled partly because Chinese companies can't reliably source advanced HBM. If China can't solve the memory problem, it doesn't matter how clever DeepSeek's architecture gets — the hardware ceiling will eventually bite.

Liang Wenfeng buying into CXMT reads as a hedge against exactly that ceiling. It's the AI software guy saying: I need to own a piece of the physical layer.

This tracks with a broader pattern in China's tech ecosystem right now. We're watching the great vertical integration land grab of 2025. Alibaba (阿里巴巴), which owns the Qwen (通义千问) model family, has been pouring capital into its in-house chip design unit T-Head (平头哥). ByteDance (字节跳动), which operates the Doubao (豆包) AI assistant, has been quietly hiring semiconductor talent and designing its own inference chips. Huawei, of course, controls both the Ascend silicon and the MindSpore software stack.

DeepSeek is the outlier — a lean, research-first lab with maybe 200 employees that punched above its weight through algorithmic ingenuity rather than brute-force compute. But scale changes your incentives. Once you're serving millions of API calls daily and planning next-generation models, you start caring deeply about who makes your memory, who makes your interconnects, and whether your supply chain can survive another round of US export controls.

The Chinese internet's reaction has been predictable in the best way. On Toutiao (今日头条), commenters are split between two camps: the "Liang Wenfeng is a genius who sees three moves ahead" camp, and the "CXMT is a money pit that will never catch Samsung" skeptics. Both are probably right. Memory chips are notoriously capital-intensive and cyclical — CXMT could easily burn another $20 billion before it reaches true competitiveness on DDR5 and HBM3. But China's government has made semiconductor self-sufficiency a national priority on the level of the space program, and patient capital is the one thing the Chinese system is actually good at mobilizing.

What I find most interesting is the symbolism. For the past two years, the narrative around Chinese AI has been "great software, terrible hardware." DeepSeek R1 was the ultimate proof point: a world-class model trained on constrained compute, a triumph of software engineering over silicon limitations. But Liang Wenfeng buying into a memory chip company suggests even the software heroes know that story has a limited shelf life.

You can only cheat physics for so long.

The other subtext here is about talent flows. China's chip industry has historically struggled to attract the kind of elite engineering talent that floods into AI and internet platforms. Having someone with Liang Wenfeng's cachet — the DeepSeek halo — publicly validate CXMT sends a signal to every ambitious engineer in Hangzhou and Shenzhen: memory chips are not a backwater, they're the frontier. That kind of narrative shift matters more than any single investment.

The broader market read is that Chinese AI's next phase won't be defined by model releases or benchmark scores. It'll be defined by who controls the stack — from the silicon up through the inference layer to the consumer app. DeepSeek just told you which direction the wind is blowing.

Watch the follow-on moves. If other AI lab founders start taking stakes in chip design houses, foundries, or packaging companies, you'll know the great Chinese AI vertical integration is officially underway. And the Americans should probably stop arguing about model weights and start paying attention to memory bandwidth.