Ex-SenseTime chip unit Sunrise banks ~US$420M as inference GPUs take the spotlight

A chip team that began life inside SenseTime (商汤) is now racing toward its own public debut — and investors are writing nine-figure checks to get there.

The wager is on 推理GPU (inference GPU), the quiet workhorse that powers every chatbot reply, ad render, and recommendation. Sunrise (曦望), spun out of China's best-known AI lab at the end of 2024, just closed another heavy funding round, and the money is earmarked for turning prototypes into shipped silicon.

From an internal unit to an independent bet

Sunrise was not born in a garage. Its roots trace to Shanghai Zhenliang Intelligent Technology (上海阵量智能), the in-house large-chip department that SenseTime built to design its own accelerators. When SenseTime reorganized under a "1+X" structure in late 2024, the chip group was pushed out as an independent company and later rebranded Zhejiang Sunrise Intelligent Technology (浙江曦望智能科技股份有限公司) after a joint-stock reform in mid-2026.

The leadership reads like a roll-call of China's AI and chip establishment:

• Chairman Xu Bing (徐冰), a SenseTime co-founder who steered the lab's own Hong Kong listing and several fundraising records.
• Co-CEO Wang Yong (王勇), a two-decade chip veteran who previously served as chief GPU architect at AMD and worked on Baidu's Kunlun (昆仑芯) line.
• Co-CEO Wang Zhan (王湛), an early Baidu (百度) executive who built the "Phoenix Nest" ad system.

That pedigree matters because the inference-GPU market is short on proven operators and long on ambitious claims. A seasoned capital-markets chairman plus silicon engineers with shipping experience is exactly the combination investors say they want to see before underwriting a hardware-heavy float.

The round that set the tone

The most consistently reported financing event is a strategic round announced in January 2026, sized at roughly RMB 3 billion (≈ US$423M / HK$32.7亿). Reporting from China Securities Journal and corroborated by Sina Finance and Tencent Securities places a broad investor mix at the table:

1. State-linked capital such as China Chengtong's mixed-ownership fund and industrial funds from Hangzhou.
2. Established venture names including IDG Capital, Gaorong Ventures, and Cowin Capital.
3. Corporate strategics spanning manufacturing, gaming, and travel.

A follow-on tranche of more than RMB 1 billion (≈ US$140M) was reported in April 2026 with Hangzhou municipal and domestic VC participation. Multiple rounds since the 2024 spin-off have pushed the company's valuation above RMB 10 billion (≈ US$1.4B) according to Sina Finance. The company has not publicly detailed a listing date, but it has begun preparatory work for a Hong Kong IPO and is widely described as competing for one of the exchange's AI-chip listing slots.

Why inference, why now

Training a model grabs the headlines, but inference is where the recurring cost lives. As Chinese enterprises move large model (大模型) deployments from demo to production, demand for cheap, compatible inference silicon has surged, and the center of gravity is shifting from "how fast can we train" to "how cheaply can we serve."

According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the inference-GPU segment is drawing an outsized share of the country's AI-chip venture capital because it faces fewer fabrication bottlenecks than frontier training chips while still riding the same national-supply narrative. That mix — strategic relevance plus a nearer path to revenue — is what lets pre-profit hardware names keep raising.

Sunrise's product story supports the pitch:

• The S1, aimed at cloud-edge vision inference, reached mass production in 2022 with cumulative shipments past 20,000 units.
• The S2, a general large-model inference GPU, shipped at scale in 2024 with broad CUDA-ecosystem compatibility.
• The S3 flagship, unveiled in January 2026, is marketed as delivering more than 3x the inference performance of the S2 while cutting per-token cost by up to 90%.

A window that is also a squeeze

Hong Kong's Chapter 18C regime for specialist technology companies is the obvious venue for a Sunrise listing: it tolerates pre-profit issuers that can show meaningful R&D and a credible commercialization path. That same regime, however, has lately drawn tougher scrutiny, with regulators and sponsors more willing to question whether "embodied AI" and "domestic substitution" narratives rest on real orders.

For inference-GPU makers, the squeeze is two-sided. On one side, hyperscalers and enterprises want a credible alternative to imported accelerators. On the other, price competition is brutal and software ecosystems are hard to clone. A company can post a flattering benchmark and still lose the account to a rival that plugs more cleanly into existing toolchains.

The open questions

Plenty of risk sits under the hood. SenseTime remains a meaningful shareholder, so Sunrise's fortunes are partly linked to its former parent's strategy and capital needs. The business also carried a tiny revenue base in its earliest standalone year, a reminder that "national AI-chip champion" stories often outrun the income statement.

Investors should watch three things:

• Whether S3 volume shipments convert the 90% cost-cut claim into real customer wins rather than pilot deployments.
• Whether the Hong Kong listing actually proceeds, given tighter scrutiny of pre-profit tech floats.
• Whether gross margins hold as price competition in inference silicon intensifies and customers consolidate spend.

What readers can do now

1. Track Sunrise's HKEX filings directly rather than relying on secondary round-size headlines, since private-round figures are frequently restated.
2. Compare the inference-GPU thesis against listed peers such as Cambricon (寒武纪) and Hygon (海光信息) to gauge whether the "cheaper inference" premium is already priced in.
3. Treat any "IPO imminent" claim as provisional until the company publishes a hearing-approved prospectus with audited numbers.

Honest limitations

This article relies on Chinese financial media (China Securities Journal, Sina Finance, Tencent Securities, Jiemian) and secondary aggregators; it did not access Sunrise's audited financial statements or a published prospectus, which were not public at writing. Private-round sizes and valuations are self-reported or media-sourced and can be revised. The competitive claims about S3 performance come from the company's own marketing, not an independent benchmark. Dimensions not covered include detailed customer concentration, supply-chain exposure to advanced packaging, the precise terms of SenseTime's remaining stake, and how the firm's books compare with rivals Biren (壁仞科技) and Iluvatar (天数智芯) also pursuing Hong Kong listings.