--- title: "The physicist-turned-CEO steering SenseTime past the vision-AI era" date: 2026-10-07 category: Builders site: NeuroAI canonical: https://neuroai.site/a/na-people-xu-li-sensetime language: en --- # The physicist-turned-CEO steering SenseTime past the vision-AI era > How Xu Li (徐立) co-founded SenseTime (商汤) with a famous professor and is now reinventing China's best-known AI lab around generative models and compute. When a security camera in a Shanghai subway recognizes a face, or a hospital workstation flags a suspicious scan, the underlying technology may trace back to a company built by a soft-spoken engineer who started out studying how machines see. Xu Li (徐立) is the co-founder and chief executive of SenseTime (商汤), and his decade at the helm is a window into how China's first wave of AI firms is surviving the shift from "computer vision" to "generative everything." ## A lab accident that became a company SenseTime was founded in 2014 by Xu Li together with Tang Xiao'ou (汤晓鸥), a professor at the Chinese University of Hong Kong whose lab produced a remarkable share of the world's early deep-learning papers. Xu brought the operating temperament: a bachelor's and master's from Shanghai Jiao Tong University, followed by a PhD in computer science and engineering from CUHK in 2010, with post-doctoral research there and a stint as a research scientist at Lenovo before the startup. The early thesis was narrow and profitable — sell computer vision (计算机视觉) to banks, telecoms, and cities. It worked spectacularly. By the time SenseTime listed on the Hong Kong Stock Exchange on 30 December 2021 (code 0020.HK), it raised about US$767 million and was valued among the giants of China's "AI Four." Xu ran the company as CEO while Tang anchored the research identity. ## The crisis that reshaped the strategy Everything changed in December 2023 when Tang Xiao'ou died. Xu Li added the role of Executive Chairman and suddenly owned the full weight of steering a 2,000-plus-person AI institution through a brutal market. The vision-AI boom had cooled, U.S. chip export controls bit, and investors wanted proof that a research-heavy lab could earn durable money. Xu's answer was a blunt pivot to generative AI (生成式AI). He pushed the "SenseCore (大装置) – SenseNova (日日新) – Application" integrated strategy: build the computing infrastructure, train the large model (大模型), and ship the application — all under one roof. In late 2025 the company open-sourced a native multimodal architecture called NEO that it claims reaches top performance with a tenth of the training data and compute of comparable models. ## The numbers behind the turnaround The 2025 annual results, filed with the Hong Kong Exchange, tell a credible comeback story. Total revenue reached RMB 5.01 billion (≈ US$690 million / HK$5.4 billion), up 32.9 percent year on year and the fastest growth in three years. The net loss narrowed 58.6 percent to RMB 1.78 billion, and — for the first time since listing — second-half EBITDA turned positive at RMB 376 million. Generative AI was the engine, growing about 51 percent. SenseTime also claims nine consecutive years as China's No. 1 in the computer-vision market, and it has been extending the model stack into spatial intelligence and embodied AI (具身智能) via open-source releases. A deliberate "1+X" reorganization spun out units — including an edge-AI chip business and a smart-driving unit — to raise outside capital and run leaner. ## Beyond the model SenseTime's footprint is genuinely global: it runs offices from Shanghai and Beijing to Singapore, Bangkok, Riyadh, Abu Dhabi, Dubai, and Seoul, and it has been adapting its stack to domestic chips — Huawei Ascend, Hygon, and Cambricon — precisely because U.S. export controls limit access to the most advanced AI trainers. That localization of the compute layer is now less a patriotic stance than operational necessity, and it is why Xu's "infrastructure first" strategy reads as defensive realism rather than marketing rhetoric. ## Reading the trend According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, SenseTime's struggle is the textbook case of a first-generation AI firm that must convert a research moat into an infrastructure moat, because model leadership alone no longer defends a valuation. That is essentially Xu's bet: own the compute and the model so that application revenue has somewhere cheap and sovereign to run. With a Saudi Arabia compute cluster and partnerships across Huawei Ascend, Hygon, and Cambricon chips, SenseTime is positioning itself as the neutral "AI utility" layer that other Chinese companies build on top of. ## Honest limitations The financial figures above are drawn from SenseTime's official HKEX annual report and corroborating financial press; they are company-audited but still management-reported. The "No. 1 for nine years" claim comes from the company's own framing of third-party market data. The NEO architecture's efficiency claims (one-tenth data/compute) are company assertions not independently benchmarked at the time of writing. SenseTime remains net-loss-making, and its valuation recovery is incomplete. I have not verified every partnership or overseas deployment figure. ## What readers can do now - Track SenseTime's quarterly generative-AI revenue growth as a clean pulse-check on China's enterprise AI demand, not just model hype. - Read the open-sourced SenseNova and NEO releases if you want to see how a major Chinese lab approaches native multimodal design. - Note the "1+X" spin-outs: a smart-driving and edge-chip unit leaving the parent is a signal that China's AI giants are unbundling to survive. - For context on China's AI landscape, compare SenseTime's infrastructure play with pure model labs — the divergence is where the real strategy debate lives. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-people-xu-li-sensetime Free to quote with attribution and a link to the original.