Fuquan Wang and the price of changing his mind

A company party helped persuade Fuquan Wang to give up the business he had spent years building. In January 2018, he attended a friend's chip company's annual gathering and encountered an uncomfortable comparison: a small team selling specialized crypto chips was making money on a scale his processor business had never approached. The invitation came with an unexpectedly expensive lesson.

Wang had founded Zhongke Shenglong, the business associated with Sunlune, in 2009. For nearly a decade, he worked on commercial uses for Loongson, China's homegrown processor project. At the party, the attraction was the direct relationship between useful computing and a customer's willingness to pay. A chip that performed a particular job efficiently could find a buyer without first persuading the world to adopt a new general computing ecosystem.

That distinction mattered to someone who had lived with the difficulty of selling processors. Wang began moving toward specialized computing in 2018 and subsequently closed the Loongson business. The decision took him into crypto hardware, a field that some old colleagues regarded as a departure from more respectable work. Respectability, unfortunately, does not arrive with a purchase order.

Today, the Sunlune founder also leads SUNMMIO, a Beijing company developing AI inference chips. The route between those businesses runs through repeated changes of technical direction. Wang's story becomes more interesting when those changes are treated as decisions with costs, rather than tidy stepping stones placed conveniently in advance.

02Before the miner, the laboratory

Wang's technical formation began at the Chinese Academy of Sciences. At its Institute of Acoustics, he pursued doctoral research in high-throughput array signal processing under acoustician Zhang Renhe. He then undertook postdoctoral work in computer architecture at the Institute of Computing Technology, working with Loongson chief scientist Hu Weiwu. These were overlapping worlds of signals, processors and systems.

A 2012 research paper gives that background a more tangible shape. Wang, Ming Liu and Chao Yan studied how to reduce the hardware resources needed for a multichannel digital filter. Their approach allowed multiple streams of sampled data to share a single filter implementation on an FPGA, a chip whose logic can be configured after manufacture.

The paper examined savings in registers, multipliers and adders. Those are modest nouns beside today's talk of AI revolutions, but they describe the daily substance of computing: deciding how much hardware a task actually requires. A circuit designer cannot pay for a second multiplier with a stirring account of the future.

Liu later became Wang's co-founder and chief technology officer. Their connection therefore predates the current AI venture by years. The laboratory work also supplies a useful way to read Wang's subsequent career: changing the arrangement of a system can matter as much as adding more components. The problem in that earlier paper was precise: process several streams without multiplying the hardware assigned to them. Long before the miner, Wang and Liu were already examining what a calculation could do without.

03The work he agreed to throw away

The crypto venture began with an unusually precise headcount. Wang described a team of five and a half people, because one member worked part time. The group explored high-bandwidth memory, then an approach based on on-chip SRAM, then connections between multiple chips. Each offered a way around the same difficulty: getting enough data to the calculation.

The SRAM effort consumed 18 months before the team abandoned it. Another approach made the circuit board too complicated. By late 2019, they had settled on stacking memory and logic vertically. Seen from the end, that sequence can look inevitable. Inside a small company, 18 months represents salaries, investor patience and work that people would quite reasonably prefer to finish.

There was another decision after the design work. Wang committed 50 million yuan to a domestic memory supplier when the company had roughly 70 million to 80 million yuan available. He later explained the commitment in terms of securing a supply chain for volume production. A working design would be of little commercial use if the company could not reliably obtain the parts.

The Jasminer X4 reached the market in the fourth quarter of 2021. Its importance to Wang's later career was practical: the team had worked through the journey from architecture to manufacturing and delivery. Supplier commitments, production problems and customer orders had joined the research agenda. The chip business had acquired the untidy obligations of a business.

04A chip with an upstairs

Inside the X4, DRAM memory was bonded to the logic die. The arrangement created a short vertical route for data. Imagine a calculation waiting for its next inputs: making the arithmetic faster helps only so much if the inputs are still arriving slowly. Bringing storage closer addresses a different part of the delay.

The X4 was examined in 2023, and its hybrid-bonded DRAM construction was documented independently. Sunlune advertised memory-access bandwidth of one terabyte per second, five gigabytes of storage and up to 65 million hashes per second at 23 watts. Those were the company's mining-chip specifications, with a particular algorithm and purpose. They should be read in that context.

For Wang, the useful inheritance was experience building a physical structure that combined memory and computation. An application-specific integrated circuit, or ASIC, accepts a narrower job in exchange for the possibility of doing that job efficiently. The commercial question is whether the chosen job stays valuable for long enough to justify the design.

In September 2022, Sunlune announced a Series A investment from Intel Capital worth tens of millions of dollars. Wang said the money would support chip research, hiring and international expansion. Ethereum completed its Merge that same month, ending proof-of-work mining on its main network. Compatible networks, including Ethereum Classic, remained, but the market had delivered a reminder that a network's rules can change faster than a chip.

05The AI business he once opposed

Wang had been skeptical about entering AI chips. Before large language models became prominent, the variety of models and the strength of established GPU software made a narrowly specialized startup look difficult to justify. His earlier experience with processors had already supplied an education in the cost of an ecosystem.

During his time in Silicon Valley in early 2023, he reconsidered. Large models placed heavy demands on memory capacity and bandwidth, giving his team a familiar problem in a new market. SUNMMIO had been incorporated in November 2022; its subsequent work concentrated on chips for large-model inference, the process of running a model to generate responses.

His shorthand for the new design is TokenPU. Tokens are the units a language model processes and generates. Wang wants the chip's value judged against that output: how much useful inference can a customer obtain for the money and electricity consumed? A processor specification becomes a business proposition only after someone can connect it to a workload.

The first-generation A4E entered tape-out on June 15, 2026. Tape-out sends a completed chip design into the manufacturing process. The announced structure stacks eight memory wafers above a logic wafer, using mature domestic processes. It turns Wang's architectural preference into a literal cross-section: storage upstairs, calculation underneath. The diagrams are neat; producing the stack consistently is the difficult part.

06Customers, capital and a court order

In February 2026, SUNMMIO disclosed two funding rounds totaling nearly one billion yuan. Source Code Capital and Stony Creek Capital co-led the Pre-A round; Xianghe Capital led the Pre-A1 round. The money was intended for development and production. By October, the company had announced further rounds, bringing its fundraising over the preceding year to nearly two billion yuan and its reported valuation to 12 billion yuan.

Those numbers belong to the company. They describe the resources available for an industrial project and the price investors accepted for an ownership stake. Wang's customer argument remains more workmanlike: deploy a chip around a specific inference requirement and make the operating economics persuasive. Capital can pay for the engineering needed to attempt that argument; customer use has to complete it.

Sunlune's history also includes a consequential software dispute. In November 2024, a US federal court entered default judgment against Sunlune Corporation in a case brought by Synopsys over unauthorized access to chip-design software. The judgment included $27.5 million in statutory damages and a permanent injunction. In May 2025, the court denied the company's motion to vacate it.

Wang had signed filings as Sunlune's director and CEO while the corporation was attempting to represent itself. The court struck those filings because a corporation must appear through licensed counsel. The defendant was Sunlune Corporation. The episode belongs in an account of Wang's leadership alongside the engineering and fundraising: running a chip company carries obligations well beyond getting the silicon to work.

“I think I have become a practical idealist.”

Fuquan Wang / translated from Chinese

07Keeping the ambition, revising the answer

Wang has described himself as a “practical idealist.” It is a revealing combination for a researcher who left processor commercialization for mining hardware and then reconsidered his opposition to AI chips. The ambition survived successive changes in the product. His preferred technical answer remained open to revision, sometimes after a considerable amount of money and time had already been spent.

There is a small piece of continuity in the paperwork. A US patent issued on October 20, 2026 names Wang alongside Kai Cai and Ming Liu for a proof-of-work chip method. It concerns managing data used by a mining calculation. Even as the public conversation has moved toward tokens, the engineering record still includes the older workload that helped develop the team's experience.

SUNMMIO's aim is international competitiveness in AI computing. That aspiration is larger than a funding round or a completed design, and Wang has spoken about the need to remain open to what comes next. His career offers several concrete examples of what that can require: leaving an established business, rejecting completed design work and buying into a manufacturing relationship before the outcome is secure.

The next stage is to turn an AI chip design into repeatable production and useful customer output. Wang has arrived here with a team, manufacturing experience and money to continue. The question now is wonderfully prosaic for such an elaborate piece of engineering: can the customer get enough worthwhile work out of it? After the laboratories, the party and the changes of mind, there is still an invoice to earn.