Sunrise (曦望)
A dedicated inference-only GPU architecture bet — the Hangzhou-based chip spinout from SenseTime’s semiconductor division
Sunrise (legal entity: Hangzhou Xiwang Chip Technology Co., Ltd., headquartered in Hangzhou) traces its origin to SenseTime’s large-chip R&D unit, Shanghai Zhenliang Intelligent Technology, established in May 2020 as a wholly-owned SenseTime subsidiary. Under SenseTime’s “1+X” reorganization strategy announced in December 2024, the capital-intensive, long-cycle semiconductor business was carved out into an independent legal entity, formally launching in July 2025 alongside a wave of external capital raises. In our assessment, this positions Sunrise as a comparatively lower technology-risk spinout: unlike most early-stage fabless entrants, it inherited a mature engineering asset base — two generations of already-taped-out silicon — rather than starting design work from a blank sheet.
Some English-language coverage romanizes the company name directly as “Xiwang” or conflates it with its brand name “Sunrise,” while the flagship product line is marketed separately under the “Qiwang (启望) S-series” name. This report follows the company’s own usage — “Sunrise” (曦望) for the corporate entity, and “Qiwang S1/S2/S3” for the product line — to avoid ambiguity.
Co-founded SenseTime alongside Professor Tang Xiao’ou while a PhD candidate at the Chinese University of Hong Kong (CUHK). During his tenure at SenseTime, he is credited with leading over $1 billion in cumulative capital raises and playing a central role in the company’s 2021 Hong Kong Stock Exchange listing (HKEX: 20). He stepped down as SenseTime’s executive director in May 2025 to focus full-time on Sunrise’s semiconductor business.
A veteran chip architect with more than 20 years of design experience, previously at AMD and Baidu’s Kunlunxin unit. He joined SenseTime in 2020 to lead R&D, overseeing development of the S1 (multimodal vision inference chip) and S2 (large-model inference GPU), both of which self-reportedly achieved first-silicon functional success (“first-time bring-up”). At Sunrise, he continues to lead next-generation chip R&D and mass-production supply chain execution.
An early Baidu team member and former senior vice president who led the build-out of Baidu’s “Fenghao” search advertising system, earning him a reputation within the industry as an architect of Baidu’s ad-monetization engine. He left Baidu in April 2016 and was in semi-retirement before returning to lead operations, commercialization strategy, and external fundraising at Sunrise.
A specialist in end-to-end IPO process management, listed-company financial infrastructure, and industrial capital integration. He is credited with leading the listing of Runze Technology, cited as the first successful reverse-merger listing on the ChiNext board. He works alongside Chief Marketing Officer Liu Ying as part of the executive team.
In our view, the leadership composition combines three distinct capabilities that are unusually deep for an early-stage fabless entrant: (i) proven chip design and mass-production execution under SenseTime (Wang Yong), (ii) large-scale commercial and organizational growth experience (Wang Zhan), and (iii) direct capital-markets and listing-process expertise (Ren Yuan, Xu Bing).
Sunrise is a fabless semiconductor company designing GPGPUs purpose-built for AI inference workloads. Under SenseTime, the team had already taped out and shipped two chip generations — the S1 multimodal vision inference chip and the S2 large-model inference GPU — before formally launching its flagship product, the Qiwang S3, in January 2026 as an independent entity. We view this pre-existing production track record as a meaningful execution-risk mitigant relative to pure R&D-stage peers, though S3’s performance at scale-up delivery volumes remains an unproven variable at this stage.
Rather than pursuing a general-purpose GPU roadmap that serves both training and inference, Sunrise’s operating model concentrates its transistor and power budget specifically on inference workloads.
Developed and mass-produced under SenseTime, these first- and second-generation chips reportedly saw the substantial majority of delivered compute consumed by inference workloads — a real-world usage pattern the company cites as the design rationale for making S3 inference-only by architecture.
Substitutes LPDDR (low-power memory typically used in mobile devices) for HBM to lower cost and improve supply reliability, while low-precision compute reduces per-token data movement. Self-reported as the first domestic GPU to adopt LPDDR6 and PCIe Gen6.
Company disclosure indicates recent financing proceeds are earmarked for S3 scale-up delivery, full-stack software ecosystem development, and S4/S5 R&D. We read the multi-generation roadmap, rather than a single-product bet, as an attempt to establish a repeatable execution track record.
Memory Strategy — LPDDR in Place of HBM: HBM offers substantially higher bandwidth but commands a cost premium, with advanced packaging capacity concentrated among a small number of suppliers such as TSMC, creating a persistent supply bottleneck. LPDDR trades lower bandwidth for larger achievable capacity, lower cost, and a more diversified supply base — reinforced by domestic memory maker CXMT’s 2025 entry into LPDDR5X mass production and its advancing LPDDR6 development. As agentic workloads increasingly demand larger working memory for long-context and transient state, we view this as a rational trade-off that prioritizes capacity, cost, and supply security over peak bandwidth.
Certain local media outlets have reported cumulative revenue in the low hundreds of thousands of yuan (approximately RMB 240,000), a figure the company has not officially confirmed and one that stands in stark contrast to roughly RMB 6 billion in cumulative capital raised. We classify this revenue figure as unverified, single-source data and recommend treating it as directional only pending official financial disclosure.
In the roughly thirteen months since its formal July 2025 launch, Sunrise has reportedly completed seven or more financing rounds totaling approximately RMB 6 billion (~$840 million) in cumulative proceeds. Of particular note, post-money valuation is reported to have nearly doubled — from roughly RMB 10 billion to roughly RMB 20 billion — between April and August 2026, a four-month window. In our assessment, this pace of re-rating is unusually rapid even by the standards of China’s current domestic AI-chip capital cycle, and we flag it as a risk factor in its own right (see Section 05).
Regarding the August 2026 round (approximately RMB 2 billion), a Sunrise representative stated the company was “not in a position to make further comments for the time being.” Accordingly, the round’s size, valuation, and investor roster in this section derive from exclusive media reporting cross-referenced against corporate registry filings rather than an official company announcement, and specific terms could differ once formally disclosed.
Established with Wang Yong as legal representative, wholly owned by SenseTime. At this stage the entity operated as an internal compute-chip R&D project rather than an independent startup, with headcount and budget fully absorbed within SenseTime’s own organization.
SenseTime CEO Xu Li announced the “1+X” strategy to carve the capital-intensive, long-cycle semiconductor business out into a standalone entity, initiating the process that would allow the chip organization to raise external capital independently.
An early-2025 angel round with participation from SenseTime and Legend Star was followed by a Series A of roughly RMB 250 million led by Beijing Lier in early May, establishing the newly independent entity’s first external fundraising track record.
Sunrise formally launched under its “曦望 Sunrise” brand alongside disclosure of an approximately RMB 1.0 billion round, with participation from Sany-affiliated Huaxu Capital, 4Paradigm, Yoozoo Interactive, Beijing Lier, Cowin Capital, and Haitong Kaiyuan. Per company disclosure, cumulative 2025 proceeds had by this point exceeded RMB 1.5 billion.
Self-confirmed cumulative proceeds of approximately RMB 3.0 billion within the entity’s first year of independent operation. That same month, the flagship Qiwang S3 chip launched, aligning product and capital-markets momentum.
A round exceeding RMB 1.0 billion, led by Hangzhou Capital, pushed post-money valuation above RMB 10 billion. IDG Capital, Gaorong Ventures, and CICC Capital are reported to have collectively invested approximately RMB 3.0 billion over the preceding twelve months. Cumulative rounds reached seven, with total proceeds of approximately RMB 4.0 billion.
Per exclusive media reporting, Sunrise completed a new round of approximately RMB 2.0 billion, with post-money valuation reaching roughly RMB 20 billion (~$2.8B) — nearly double the April figure within a four-month window. New investors include industrial capital CP Group, Andon Health, Infore Environment, and Tongcheng Travel, alongside state-affiliated “national team” funds PICC Capital and CCB Principal Asset Management. Janchor Partners, CAS Star, and returning investor Cowin Capital are also reported to have participated.
• Round size: ~RMB 2.0 billion (~$280 million)
• Post-money valuation: ~RMB 20 billion (~$2.8 billion), roughly 2x the April figure within four months
• Cumulative proceeds since spinoff: ~RMB 6.0 billion (~$840 million) across 7+ rounds
• New investor profile: industrial end-users (CP Group, Andon Health, Infore Environment, Tongcheng Travel) plus state-affiliated funds (PICC Capital, CCB Principal)
• Company’s official position: “not in a position to make further comments for the time being” (no comment)
In our view, Sunrise’s competitive positioning rests on four pillars: (i) a clearly differentiated inference-only architecture, (ii) a memory strategy aligned with domestic supply-chain localization by avoiding HBM dependence, (iii) a proven mass-production track record inherited from its SenseTime era, and (iv) a cap table weighted toward industrial end-users rather than purely financial investors. Where many domestic GPU peers pursue general-purpose architectures serving both training and inference, Sunrise has staked out a clear inference-first positioning from the outset.
As industry compute demand shifts from training toward 24/7 inference and token generation, S3’s decision to forgo a combined training-and-inference design in favor of concentrating transistor and power budget on inference efficiency reads as a well-timed bet. That said, this is also a trade-off: it could constrain product-portfolio flexibility should training demand re-accelerate.
Unlike HBM supply, which is structurally bottlenecked by advanced-packaging capacity concentrated at a handful of suppliers such as TSMC, LPDDR sourcing benefits from CXMT’s advancing LPDDR5X/LPDDR6 production ramp within China, offering a comparatively more secure domestic supply chain — and a partial hedge against geopolitical supply risk.
Both the S1 and S2 chips developed under SenseTime are self-reported to have achieved functional operation on first silicon. For a fabless startup, repeated successful tape-outs are a relatively scarce execution signal, and in our view underpin investor confidence in the subsequent S3 and S4/S5 R&D pipeline.
A cap table skewed toward end-user industrial capital — CP Group, Andon Health, Infore Environment, Tongcheng Travel — and state-affiliated funds such as PICC Capital and CCB Principal, rather than purely financial VCs, is open to interpretation as quasi-pre-order strategic capital that could translate into future compute purchase or strategic cooperation agreements. At the same time, this structure carries a latent governance risk of skewing toward the interests of specific industrial stakeholders.
We characterize Sunrise as a high-beta domestic-semiconductor story with clear product and fundraising momentum, but with commercialization proof points that remain at an early stage. Risk and opportunity factors are set out in balanced form below.
- Structural growth in 24/7 inference demand driven by proliferating agentic workloads — consistent with the broader Token Economics narrative
- A team with a proven execution record, including two consecutive first-silicon successes
- Cap table weighted toward industrial end-users and state-affiliated funds — potential for future pre-order or policy-linked support
- LPDDR-based memory strategy partially hedges geopolitical supply-chain risk via linkage to domestic memory production
- The 4-month valuation re-rating can be read as a strong market demand signal
- No official revenue disclosure; the self-reported revenue figure (~RMB 240,000) sits far below cumulative proceeds (~RMB 6.0 billion), indicating a very early commercialization stage
- Listing path remains unconfirmed relative to already-listed peers such as Moore Threads and Biren Technology, with Enflame and Kunlunxin also entering the IPO pipeline
- S3’s large-scale delivery stability and its targeted 90% per-token cost reduction remain unproven — even the sourced media coverage itself characterizes the innovation as “only half completed”
- A 2x valuation re-rating within four months raises the possibility that capital inflows are running ahead of fundamentals, creating downside re-rating risk should subsequent rounds fail to materialize on similar terms
- No official company confirmation of the latest round leaves residual uncertainty around the precision of reported terms and investor roster
- Broader exposure to export-control and geopolitical policy risk affecting China’s domestic AI-semiconductor sector