When most people outside China think of SenseTime (商汤科技), they still picture surveillance cameras and face recognition. Inside its 2025 annual report, a different company is visible: one that describes itself as running an "infrastructure–model–application" three-in-one strategy, where the plumbing (SenseCore), the models (SenseNova), and the products are sold as one stack.
That reframing matters. As China's model race crowded toward consumer chatbots, SenseTime leaned the other way — into the boring, capital-heavy business of renting compute and models to enterprises.
SenseTime was founded in 2014 by researchers from the Chinese University of Hong Kong and listed in Hong Kong in December 2021. Its earlier identity —城市感知 (city sensing) and facial-analysis software — made it a target of U.S. export restrictions, which is precisely why the domestic-chip story below is not a marketing choice but a survival constraint.
From algorithms to a computing grid
SenseCore is SenseTime's AI infrastructure platform. According to the company's 2025 annual results (filed with the Hong Kong exchange), SenseCore moved in 2025 "from technological strengths to an industrial closed loop." The concrete claim: it supported nearly one million model R&D tasks during the year.
The more notable engineering point is the chip strategy. Under export-control constraints that limit access to leading Nvidia GPUs, SenseCore was built to run on domestic accelerators — Huawei Ascend (昇腾), Hygon (海光), and Cambricon (寒武纪). SenseTime says it completed full adaptation to the Ascend 384 super-node and launched a "SenseCore Computing Power Mall" (算力商场) pooling resources from more than a dozen chip suppliers.
Total compute is reported at roughly 25,000 PetaFLOPS as of August 2025.
The model layer: SenseNova
On top of the grid sits SenseNova, SenseTime's multimodal foundation-model family. The report dates two key releases:
- SenseNova V6 — launched April 10, 2025. The company positions it as leading among domestic models on multimodal reasoning, with data-analysis performance it claims surpasses GPT-4o, and says it was the first Chinese model able to analyze medium-to-long videos up to about 10 minutes.
- SenseNova V6.5 — launched July 27, 2025, adding interleaved visual-linguistic chain-of-thought and multimodal reinforcement learning.
A separate inference system, LightX2V (a "world model" video generator), was highlighted as running real-time video generation on domestic hardware, and the company says its open-source release passed ten million downloads on Hugging Face.
Why a "native AI cloud" claim
In the China Full-Stack AI Cloud Service Market Report (H1 2025) by Frost & Sullivan and LeadLeo, SenseTime says SenseCore ranked among the top four in China's full-stack AI cloud market and first among "native AI cloud vendors" — i.e., vendors built around AI rather than retrofitted from generic cloud.
That distinction is the strategic bet: not to out-chat OpenAI, but to be the layer Chinese enterprises and research institutes rent when they need to train or serve models on local chips.
The overseas wrinkle
The 2025 report also states SenseTime launched what it calls China's first overseas domestic computing cluster, in Saudi Arabia — exporting its software stack rather than its chips. If accurate, it is a notable test of whether a China-built AI cloud can operate in a heterogeneous, non-domestic environment.
Certifications and the energy angle
The 2025 report lists third-party validations of the SenseCore platform. In January 2026, it says, the platform received what it describes as the industry's first "5A Excellence Level" certification in a CAICT–CTTL test of computing–model–application integration. In February 2026 it received an "Excellent Level" in MIIT's Software Supply Chain Security assessment, among the first enterprises nationwide.
SenseTime also frames SenseCore around efficiency, not just scale. It cites a "Computing Power Synergy × Energy Dispatch Engine" aimed at lowering the energy cost of large-model training and inference. For a grid-constrained market where power can be the real bottleneck on a training cluster, that claim — if it holds up — is as commercially relevant as raw PetaFLOPS.
The open-source counterweight
Not everything SenseTime builds is closed. The LightX2V world-model inference system was released as open weights (开源权重), and the company says the release passed ten million downloads on Hugging Face, at times ranking among the global top ten there. Open-sourcing inference tooling is a recruiting and ecosystem play as much as a product — it gives SenseCore a reason for developers to stay inside the stack.
Honest limitations
The core figures here come from SenseTime's own HKEX filings and the Frost & Sullivan / LeadLeo market report it cites — i.e., company-disclosed, not independently audited by a third party in this article. I did not verify benchmark comparisons (e.g., "surpasses GPT-4o") against independent tests. The export-control and domestic-chip narrative is consistent with the filing but is presented by the company in a favorable light. The Saudi Arabia cluster claim is reported by the company; I found no independent confirmation in the sources reviewed.
What readers can do now
- Read the primary document: SenseTime's 2025 annual results on the HKEX news site, not just secondary summaries, for the exact wording on SenseCore utilization and certifications.
- Separate "native AI cloud" market-rank claims from audited revenue — the rank is a commissioned report; the financials are in the filing.
- If evaluating the domestic-chip story, check whether SenseCore's Ascend-384 adaptation has independent customer references, not just the vendor's description.