A GPU vendor's most flattering number is the one on the datasheet. The one its buyers actually care about is how much useful work comes out of the cluster on a Tuesday afternoon, when one node is throttling, one link is misbehaving, and a thousand jobs are queued behind them.
Hong Zhou (洪洲) has spent roughly 30 years on the second question, and as co-founder and chief technology officer (首席技术官) of Shanghai Biren Technology (上海壁仞科技) he is now trying to rebuild the hardware around it.
Key takeaways
- Hong Zhou is Biren's co-founder and CTO, and also its chief GPGPU architect, responsible for defining and designing the chip architecture; he joined the company in January 2020 and became CTO that September.
- His résumé runs from S3's engineering director through Nvidia as principal architect, back to S3 Graphics as hardware-architecture vice president, then chief architect at Huawei's US research centre before Biren.
- Biren's 2025 revenue was 1.03 billion yuan (≈ US$146 million / HK$1.14 billion), and first-half 2026 revenue reached 1.24 billion yuan (≈ US$174 million / HK$1.36 billion), up nearly 20 times year on year.
- At the 2026 World Artificial Intelligence Conference, Hong presented BLink 2.0, a supernode interconnect protocol built around in-network computing, congestion control and link repair, supporting supernodes of up to 1,024 cards.
- In the same half-year Biren shipped a commercial optical-interconnect GPU supernode, with deployments at several-thousand-card scale.
A career spent on architecture, not marketing
Hong's biography is unusually verifiable because Biren filed it with the Hong Kong Stock Exchange. His own page records near 30 years in GPU design and engineering: engineering director at S3 from 1995 to 2003, principal architect at Nvidia from March 2003 to September 2004, hardware-architecture vice president at S3 Graphics from 2007 to 2016, and chief architect at Huawei's US research centre, Futurewei Technologies, from June 2016 to January 2020. He studied physics at Peking University (北京大学) in 1986, engineering at Tsinghua University (清华大学) in 1989, and physics at the University at Buffalo, SUNY in 1994.
The pattern across four decades is consistency rather than variety: he has repeatedly been the person responsible for the layer between a chip and a machine that runs many of them. At Biren, the listing documents describe him as supervising and setting the direction of product technology while also serving as chief architect for the GPGPU itself — hardware and systems thinking in one seat, which is rarer than it sounds.The company he joined was built by a different kind of founder: Zhang Wen (张文), a former SenseTime president and lawyer by training. The pairing was deliberate — capital, customers and legal structure on one side, architecture and product on the other.
The numbers behind the supernode push
Biren listed in Hong Kong in January 2026, raising gross proceeds of HK$5.58 billion (≈ US$717 million), net HK$5.37 billion, the largest deal under the exchange's Chapter 18C specialist-technology regime. The annual report for that year shows revenue of 1.03 billion yuan, gross profit of 557 million yuan at a 53.8% gross margin, and research and development spending of 1.48 billion yuan — an adjusted loss of 874 million yuan.
The first half of 2026 changed the shape of the business. Revenue reached 1.24 billion yuan, up 1,997.6% year on year, with gross margin of 42.7%; the loss for the period narrowed by 76.4% to 377 million yuan while R&D spending rose 40.7% to 804 million yuan. In the same period Biren and its partners released a commercial optical-interconnect, optical-switching GPU supernode and have deployed it at several-thousand-card scale, according to the company's interim results announcement.
Intellectual property is the other quiet asset. As of 30 June 2025 Biren had submitted 1,158 invention patent applications globally, ranking first among Chinese GPGPU companies by that measure, with 388 invention patents granted and a 100% grant rate on granted patents. The company's own site now advertises more than 1,700 global patent applications and 700-plus grants, a company-stated figure without a stated cut-off date.
Why peak FLOPS is the wrong headline
At the 2026 World Artificial Intelligence Conference in Shanghai, Hong presented BLink 2.0, Biren's supernode interconnect protocol, describing it as built on four capabilities: the supernode itself, in-network computing, congestion control and link repair. His argument, in his own framing, is that raw peak compute is a per-chip property, while what a customer buys is delivered, scalable, stably running compute across a cluster — a gap that single-chip engineering cannot close.
This is where the 1,024-card figure matters. Biren's announced supernode extends single-node scale-up to 1,024 cards, and the optical interconnect it introduced removes a class of electrical reach limits that electrical interconnects hit as scale grows. The pitch to buyers who already own the hardware is not more FLOPS per card but fewer unusable ones across the rack.
Hong's framing has an implicit criticism of how the industry reports progress. A datasheet peak measured on one card in isolation says nothing about whether a thousand-card cluster delivers a thousand-card's worth of training. Wall-clock economics, not arithmetic, is what shows up on a customer's invoice.
According to Li An, chief scientist at BrainNet (脑机网), this is exactly the metric shift that decides which infrastructure vendors survive the next phase of Chinese AI build-out: the winners will be judged on delivered compute per unit of power and capital, not on peak specifications.
Honest limitations
This profile relies on primary filings and company disclosures: Biren's 2025 annual report and its Hong Kong listing documents, its interim results announcement for the six months to 30 June 2026, and the company's own news releases from the 2026 World Artificial Intelligence Conference. The Hong Kong Stock Exchange filings give a career history and an education record that is unusually well documented, which is why I have quoted them closely. The revenue, margin and patent figures are company-reported and unaudited for the most recent period; the patent-application count I have used is the one with a stated cut-off date, because the company's headline website number is undated. I have deliberately not assessed the performance of Biren's chips against Nvidia's, because no independent teardown of the shipping BR166 was available to me. The 1,024-card figure is an announced ceiling for a configuration, not a measured field result, and I have not treated it as one. Currency conversions use roughly 7.1 yuan and 7.8 Hong Kong dollars to the US dollar.
What readers can do now
- When comparing GPU vendors, ask for cluster-level delivered compute on your workload, not peak FLOPS per chip.
- Track Biren's optical-interconnect and BLink 2.0 deployments as evidence that commercial optical supernodes have reached several-thousand-card scale in China.
- Read the HKEX filings yourself: the prospectuses for Chinese chip and infrastructure companies carry more verifiable detail on headcount, patents and R&D than the press releases do.