--- title: "SenseTime Bet Everything on SenseNova — and the Losses Just Got Smaller" date: 2026-10-06 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-sensetime-sensenova language: en --- # SenseTime Bet Everything on SenseNova — and the Losses Just Got Smaller > Once China's face-recognition champion, SenseTime (商汤) rebuilt itself around the SenseNova (日日新) generative model — and its 2025 books show the bet is finally showing revenue, if not profit. ## A camera company that had to learn a new language A strategy meeting in Shanghai three years ago would have been full of engineers talking about cameras, cities and surveillance tenders. Today the same company talks about tokens, agents and multimodal inference. The interesting question is no longer whether SenseTime (商汤) changed its story — it is whether the new story pays better than the old one. That question sat at the center of the company's full-year 2025 results, published in its audited HKEX annual report. The headline is blunt: the computer-vision pioneer that built China's "AI four dragons" reputation is now, by its own accounting, a generative-AI company that happens to still own a camera business. ## How the pivot actually happened SenseTime did not flip a switch. It spent 2023 gutting its four legacy business lines — smart business, smart life, smart auto, smart city — and re-stacking them into what it now reports as three: generative AI, visual AI, and "X" innovation businesses. By the first half of 2025 it had changed the disclosure labels again, a quiet admission that the camera-era categories no longer described where the money came from. The model behind the new identity is SenseNova (日日新), the large model (大模型) family. In 2025 alone the cadence was relentless: SenseNova V6 in April, V6.5 in July, the open-sourced spatial-intelligence model SenseNova-SI in November, and a brand-new native multimodal architecture called NEO in December. ## The number that surprised the skeptics The generative-AI segment is now the company's main engine, and the size of it is hard to dismiss. - Generative AI revenue reached **RMB 3.63 billion (≈ US$511M / HK$3.96B)** in 2025, up **51.0%** year on year. - That single segment is now **72.4%** of total revenue, up from 63.7% a year earlier. - Total group revenue hit **RMB 5.01 billion (≈ US$706M / HK$5.47B)**, up **32.9%**, the fastest growth in three years. For a company that three years running was written off as a government-contract dependent CV shop, a 51% surge in a self-funded model business is the part of the story worth taking seriously. ## Why the loss narrowed — and why it may not stay narrow SenseTime's net loss fell to **RMB 1.78 billion (≈ US$251M / HK$1.94B)**, a **58.6%** improvement. Adjusted net loss dropped to **RMB 1.96 billion (≈ US$276M / HK$2.13B)**, down 54.3%. More symbolically, second-half 2025 EBITDA turned positive at **RMB 376.4 million (≈ US$53M / HK$410M)** — the first time since its 2021 listing. But the narrowing came with asterisks. Research and development spend actually fell **8.6%** to **RMB 3.78 billion (≈ US$533M / HK$4.12B)**, mostly because headcount dropped from 3,756 to 2,472 employees. And a chunk of the "improvement" came from spinning off money-losing units — the autonomous-driving business left the consolidated accounts in August 2025. Strip those one-offs out, and the core model business is still burning cash to build capacity. The company's own computing base, SenseCore, ran at **40,400 PetaFLOPS (FP16)** by March 2026 — a capital-intensive moat that cuts both ways. ## The native multimodal bet SenseTime's most interesting technical wager is architectural. Instead of bolting a vision encoder onto a language model, its NEO architecture tries to fuse "left brain" logic and "right brain" spatial reasoning from the ground up. The company claims NEO reaches leading performance on roughly **one-tenth** the training data and compute of comparable models — a claim that, if true, matters enormously for a company whose biggest enemy is its own server bill. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the shift toward native multimodal architectures like SenseTime's NEO reflects a wider Chinese-lab conviction that fusing vision and language in a single model beats bolting specialist modules together — though commercial payback still hinges on enterprises agreeing to retool their workflows. ## The parts the bulls skip SenseTime is still not profitable on a full-year basis, and it has not been since 2018 — cumulative losses now exceed RMB 50 billion (≈ US$7.0B / HK$54.5B). Its market capitalization had shrunk to under **HK$75 billion** by late March 2026, a fraction of the HK$150 billion peak at its 2021 IPO, and well below the ~HK$300 billion commanded by newer model labs such as MiniMax (which was founded by a former SenseTime vice president). The generative-AI growth rate, while high, is also decelerating from the 199% and 103% explosions of 2023 and 2024. The "X" innovation businesses — robotics, healthcare, smart retail — actually shrank **5.9%** to **RMB 302 million (≈ US$43M / HK$329M)** as units were carved out. Visual AI, the old crown jewel, grew just **3.4%** to **RMB 1.08 billion (≈ US$152M / HK$1.18B)**. ## Honest limitations - **Primary source is the company's own audited filing.** The 2025 annual report is HKEX-filed and audited by PwC, but segment definitions were changed by management mid-year, so year-on-year comparisons require care. - **The "one-tenth compute" NEO claim is vendor-reported** and was not independently benchmarked at publication; treat it as a directional claim, not a verified figure. - **Profitability improvements are partly structural, not operational** — divestments and headcount cuts contributed to the narrower loss, so the underlying model-business margin is harder to read than the headline suggests. - **Valuation context (market cap, MiniMax comparison) comes from secondary Chinese financial media**, not the annual report, and stock prices move daily. - **No renminbi-to-USD/HKD figure is invented**; all conversions use approximate rates (RMB÷7.1 for USD, RMB×1.09 for HKD) and are flagged as approximations. ## What readers can do now - **If you evaluate Chinese AI vendors, separate SenseTime's model capability from its camera-era reputation.** The generative-AI segment is now real revenue, not a slide — benchmark SenseNova V6.5 / NEO on your own workloads before assuming the old "CV company" label still applies. - **Watch the second-half EBITDA figure as the real signal, not the topline growth.** Positive EBITDA that survives the loss of one-off divestment gains is the line that decides whether SenseTime reaches self-funding. - **Treat the NEO "10x efficiency" claim as a hypothesis to test**, not a fact — if you run inference at scale, a 10x training-data reduction is worth a controlled benchmark before you bet a roadmap on it. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-sensetime-sensenova Free to quote with attribution and a link to the original.