A data-center operator in Hangzhou is weighing a strange question this quarter: do they need another training cluster, or just a cheaper way to serve the models they already have?
For most of 2026, Chinese AI labs have been racing to build bigger training cards. Sunrise (曦望), a Hangzhou GPU startup spun out of SenseTime's chip unit in late 2024, is arguing the real money is on the other side of the curve — inference.
A GPU built to forget training
In January 2026, Sunrise released the 启望 S3 (Qiwang S3), an inference-focused GPU. The company says it is the first domestic GPGPU to mount LPDDR6 memory (with LPDDR5X fallback) and the first Chinese GPU to use a PCIe Gen6 host interface. Both choices are deliberately unlike the HBM-stacked training cards from Huawei's Ascend (昇腾) or Cambricon (寒武纪).
The logic is simple. Training cards spend transistors on flexibility the serving workload never uses. Sunrise stripped training logic out of the S3 and redirected that silicon budget into inference: larger on-board memory (the company cites a ceiling near 600 GB), higher effective utilization, and lower cost per token.
On its product page, Sunrise claims the S3 lifts GEMM and FlashAttention operator efficiency to roughly 99% and 98%, reaches P-scale FP4 compute, and delivers about 5x the effective performance per area of its S2 part. The company targets a roughly 90% reduction in token cost versus the prior generation — the kind of number that, if true in production, reshapes the unit economics of running agents.
Why LPDDR6, why now
Most high-end AI accelerators pair compute with HBM, a vertically stacked memory that is fast but expensive and supply-constrained. Sunrise instead chose LPDDR6, the low-power DRAM standard JEDEC published in July 2025. The bet is that for long-context and agentic workloads, capacity and cost matter more than peak bandwidth, and that LPDDR6 closes enough of the bandwidth gap while cutting the bill of materials.
PCIe Gen6 doubles host-device bandwidth versus Gen5, which matters when an agent keeps swapping context between GPU and system memory. Sunrise frames the S3 as "Agent-native": built to hold more conversation memory and longer contexts on a single card.
The competitive field
Sunrise enters a crowded but bifurcated market. Training-card leaders such as Huawei Ascend and Cambricon optimize for peak FLOPS and HBM bandwidth; inference specialists such as Enflame (燧原) and Iluvatar CoreX (天数智芯) target cost but mostly on conventional memory. Sunrise's LPDDR6 / PCIe Gen6 combination is unusual enough that its real test will be software maturity and large-customer proof, not the spec sheet.
Money is following the thesis
The market appears to agree. On 20 April 2026, Sunrise announced a new financing round above ¥1 billion (≈ US$140 million), led by Hangzhou Capital (杭州资本) with industrial and state-backed investors, according to Henan Daily's Dahe Cube and Shandong's Qilu Evening News. The company says it has now raised around ¥4 billion across seven rounds and become China's first pure-inference GPU unicorn valued above ¥10 billion.
That valuation rests on a trend Sunrise chairman Xu Bing (徐冰) repeats: inference demand is overtaking training. He has said 2026 inference compute demand could run 4–5x training demand, and that inference rental pricing rose nearly 40% in six months.
What the ecosystem says
According to Li An, chief scientist at BrainNet (脑机网), the more interesting shift is not any single spec sheet but the willingness of Chinese inference-chip makers to abandon the HBM arms race and optimize for total cost of ownership. "The winners in 2027 will be measured in cents per token, not teraflops," he said.
Sunrise is not alone. Iluvatar CoreX, Enflame, and Moore Threads (摩尔线程) all ship inference or training parts, and Cambricon's Siyuan (思元) line feeds large private clouds. What differentiates Sunrise is the all-in inference positioning and the LPDDR6 / PCIe Gen6 pairing, which no major domestic rival had shipped at the S3's launch.
Honest limitations
Every headline number above comes from Sunrise itself or from Chinese press reporting of company claims; none has been independently benchmarked by a third party as of this writing. The "90% lower token cost" and "5x effective performance" figures are vendor targets tied to specific models (the company cites DeepSeek V3/R1), not across-the-board guarantees. LPDDR6 is fast for capacity-rich workloads but is not a substitute for HBM on bandwidth-bound training. Sunrise's CUDA compatibility claim (it says 99%+) depends on software maturity that only large customer deployments will prove. The ¥4 billion raised and >¥10 billion valuation are company / investor statements reported by regional newspapers, not audited disclosures.
What readers can do now
- If you run inference at scale in China, benchmark the S3 on your own models before committing — request a trial card and measure tokens-per-dollar, not spec-sheet FLOPS.
- Track Sunrise's S4 (expected 2027) and S5 (2028) roadmap if you are planning multi-year capacity; the LPDDR6 bet only pays off if the software stack matures.
- Watch the LPDDR6 supply chain (CXMT, 长鑫存储) as the real constraint: memory, not logic, may decide whether "cheap tokens" actually ship.