Twelve years dissecting SenseTime (商汤): the true state of a veteran AI company

A bright quarter, a cold market

SenseTime (商汤) delivered what the company called a landmark 2025: annual revenue topping RMB 5.01 billion and a 33% year‑on‑year rise, with EBITDA and operating cash flow turning positive in the second half — a first since listing. Yet the stock market has been unmoved. It has been reported that SenseTime’s market value stood below HK$77 billion in early April 2026, more than 80% below its peak. Improved operating metrics meet investor skepticism. Why the disconnect?

Growth gains, but questions over quality

The headline shift is clear: generative AI now accounts for RMB 3.63 billion, or about 72% of revenue, up from under 35% in 2023, marking SenseTime’s move from computer‑vision specialist to large‑model player. Reportedly, its NEO multimodal architecture and SenseNova SI benchmarked highly in domestic tests and — according to some claims — compared favourably to international models such as GPT‑5 and Gemini‑3 Pro. But the profit improvement relied in part on one‑off gains: over RMB 1.9 billion in “other gains,” including RMB 1.313 billion from selling a subsidiary. Headcount fell by more than 30% and R&D spend dipped for the first time in three years. Growth has decelerated too: generative AI revenue growth slowed from nearly 200% in 2023 to 51% in 2025, lagging the 70.8% market average, while public‑cloud share slipped as Baidu (百度) and Alibaba Cloud (阿里云) gained ground. Margin pressure persists as compute costs jumped 163.5%, pushing gross margin down to 41%.

Spinning out talent and the 1+X gamble

SenseTime has been both a talent factory and a talent drain. Alumni have founded MiniMax, Biren (壁仞科技) and Momenta, among others, some of which have drawn far richer market valuations than their old parent. To retain teams while unlocking capital, SenseTime adopted a 1+X strategy: keep core generative and vision AI centralized (the “1”) while spinning out autonomous driving, medical AI, chips and robotics as independent entities (the “X”) with outside funding and independent CEOs. The early returns are tangible — several spin‑outs raised sizable rounds and have repaid cash to the parent — but independence creates potential conflicts of interest and cedes direct upside in large markets that SenseTime once targeted. Can the company capture enough future value from ecosystem partners rather than from owning the whole stack?

A compute moat — and a capital mill

SenseTime’s “big‑device” compute base is a rare strategic asset: roughly 40,400 PetaFLOPS of heterogeneous capacity and an early shift to domestic chips such as Huawei’s Ascend, Haiguang (海光) and Cambricon (寒武纪). In an era of tightening U.S. controls on chip supply, an onshore, heterogeneous compute stack gives SenseTime a foothold in government and sensitive enterprise workloads that foreign models cannot easily serve. But that moat is expensive. The market applies a valuation penalty to heavy‑asset, capital‑intensive models; SenseTime’s price‑to‑sales sits well below what its technology might warrant, and it has raised over HK$100 billion in placements since mid‑2024. If investor sentiment cools or funding tightens, the capital drain could blunt the very compute advantage investors say they prize.

SenseTime’s 2025 results are both vindication and cautionary tale: a company that has engineered a technical pivot and reached a cash‑flow inflection now faces structural questions about business model, talent incentives and capital intensity. Can a veteran heavy lifter translate a tested compute and model stack into sustained, high‑margin software value — and convince markets to pay for that durability? The AI marathon continues, and SenseTime has begun another twelve‑year leg with its fortresses intact but under new scrutiny.