The trading floor fell quiet for a beat, then the tape lit up. On one February afternoon in 2026, a single share of Cambricon (寒武纪) traded higher than a share of Moutai — something no domestic chip stock had managed before. Behind that number sits a soft-spoken researcher who spent ten years being told that designing AI chips (AI芯片) in China was a fool's errand.
Chen Tianshi (陈天石) didn't set out to run a company. He set out to prove a thesis.
The kid from Nanchang who skipped grades
Chen was born in 1985 in Nanchang, Jiangxi. He and his older brother — also a chip researcher — both entered the University of Science and Technology of China's Special Class for the Gifted Young (中科大少年班) as teenagers, earned PhDs, and landed at the Chinese Academy of Sciences' Institute of Computing Technology.
The brothers shared a stubborn idea in the early 2010s, when artificial intelligence was still a backwater: that general-purpose CPUs were a poor fit for the math of neural networks, and that a purpose-built processor could do the job orders of magnitude better. They taped out an early AI chip inside the academy — a bet that, at the time, had no clear customer and no clear market.
A company named after a geological epoch
In March 2016, Chen left the lab and founded Cambricon (寒武纪) — the name deliberately evokes the Cambrian explosion, the moment life on Earth suddenly diversified. The metaphor was the pitch: a "Cambrian explosion" for machine intelligence.
He took the classic full-stack route: design the chip, write the instruction set, build the software stack, and chase the customers training large models (大模型). The first public validation came fast. In 2017, Huawei's Kirin 970 — the first mobile chip with a dedicated neural processing unit — licensed Cambricon's NPU IP. For a one-year-old startup, that was a loud arrival.
Cambricon listed on Shanghai's STAR Market (科创板) under ticker 688256. The listing gave it capital, but not yet profits.
Eight years in the red
From 2017 through 2024, Cambricon lost money every single year. The AI chip business is brutal: you burn cash on R&D and wafer tape-outs long before volume arrives, and you compete against Nvidia, which owns both the hardware and the CUDA software ecosystem that most engineers already know.
For most of that stretch, Cambricon was a story investors loved to argue about and analysts loved to doubt. The bears had a point — a Chinese AI chip designer with no foundry of its own, squeezed by U.S. export controls that limited access to the most advanced manufacturing, looked structurally disadvantaged.
The year it finally clicked
Then 2025 happened. On 27 February 2026, Cambricon filed its 2025 performance flash (业绩快报). The numbers were not a rounding error:
- Revenue: ¥6.497 billion (≈ US$915M / HK$7.1B), up 453% year on year.
- Net profit attributable to shareholders: ¥2.059 billion (≈ US$290M / HK$2.3B) — the company's first annual profit after eight consecutive years of losses.
- According to the filing, the swing was driven by surging demand for AI compute and deeper tie-ups with large model (大模型) developers and internet platforms.
None of this was a mystery. As Nvidia's most advanced chips grew hard to buy in China, domestic cloud and internet giants went looking for local alternatives, and Cambricon was one of the few with a mature, shipping product.
Why a chip stock out-priced Moutai
On 27 February 2026 Cambricon closed at ¥1,178 a share (≈ US$166 / HK$1,295), valuing the company at about ¥496.7 billion (≈ US$70B / HK$546B). At its intraday peak the stock touched ¥1,595.88 (≈ US$225 / HK$1,754), briefly making it the most expensive stock on the A-share market — the symbolic "股王" title long held by Moutai.
Chen, who holds roughly a 28.57% stake, saw his paper wealth balloon. On the 2025 Hurun Global Rich List his fortune was pegged at about ¥87 billion (≈ US$12.3B / HK$95.6B), ranking him 195th globally.
The "costliest stock" label is partly a quirk of Cambricon's thin free float and partly a referendum on how urgently China wants a homegrown AI chip. Either way, the market was voting with both feet.
What Chen actually built
Strip away the stock chart and the substance is real. Cambricon ships a cloud-to-edge portfolio — training and inference cards, edge modules, and the software that ties them together. Its chips are now used to train and serve large models (大模型), the exact workload Chen imagined back in the academy "small dark room."
That is the genuinely interesting part for outside observers: a Chinese design house reached the point where domestic model builders treat its silicon as a default option. It is also a reminder that silicon is only half the battle. Nvidia's moat is the CUDA software ecosystem, and Cambricon has had to rebuild an equivalent from scratch — a slower, less glamorous grind than a quarterly earnings beat.
Honest limitations
A few things this profile does not claim:
- The 2025 figures come from a preliminary performance flash (业绩快报), not the final audited annual report; line items can shift slightly when the audit lands.
- One profitable year does not erase the structural gap with Nvidia. Cambricon still depends on external foundries, and the most advanced nodes remain constrained by export controls.
- Customer concentration is a real risk. A large share of demand traces to a handful of hyperscalers whose procurement can turn on policy.
- The "most expensive stock" milestone reflects price per share and a tight float, not market dominance.
- Personal-wealth figures are ranking estimates (Hurun), not audited balances.
What readers can do now
- Track the audited filing. When Cambricon publishes its full 2025 annual report, check the customer-concentration footnote and gross-margin trend — those reveal whether the profit is structural.
- Watch STAR Market (科创板) disclosures for capex and wafer commitments; they signal how much real capacity is behind the headline.
- Follow the large model (大模型) builders' supply chains. If domestic clouds keep naming Cambricon in their compute stacks, the thesis holds; if they quietly revert to imported silicon, it doesn't.
- Treat the "股王" headline as sentiment, not fundamentals. Price-per-share records tell you about scarcity, not moats.