SenseTime's SenseCore Captures Nearly Half of China's Neocloud Market Share
nashnova research
A Frost & Sullivan report shows SenseTime's SenseCore seized nearly 50% of China's neocloud market in H1 2026 — not by stacking GPUs, but by building a heterogeneous system that puts domestic chips to work on production workloads.
Nearly half the market — how did SenseCore pull it off?
SenseCore captured nearly 50% of China's neocloud market in H1 2026, making it the segment's clear leader.
This means → even as Musk warns that over 200 global neocloud startups face bankruptcy or acquisition, concentration at the top of China's neocloud market is accelerating.
The key difference from offshore peers like CoreWeave: SenseCore does not just rent out GPUs — it delivers an end-to-end stack from power grid to model deployment.
How does an end-to-end model differ from pure compute leasing?
CoreWeave and similar offshore neoclouds sell large-scale GPU access. Customers get raw compute and handle the rest themselves.
SenseCore covers four layers: infrastructure (grid + storage + data center + heterogeneous hardware) → platform → turnkey solutions → managed services.
In plain terms = offshore neoclouds sell a "bare shell." SenseCore sells "move-in ready" — turning compute directly into measurable business output for enterprise clients.
Domestic chips vary widely — how do you make them work together?
Restricted access to high-end GPUs is China's neocloud hard constraint. Huawei, Cambricon, and T-Head are iterating fast, but their chips differ sharply in compute specs, processing traits, and software-stack maturity.
SenseCore's answer is PD disaggregation — prefill-decode separation. It splits model inference into two stages and assigns each to the chip type best suited for the job.
In plain terms = think of inference as a two-stage assembly line. The front half (prefill) is compute-heavy and goes to domestic accelerators. The back half (decode) is memory-heavy and goes to high-end GPUs. The two stages pass data through a KV cache — an intermediate handoff layer.
Can domestic chips really stand in for high-end GPUs?
CTO Xuan Shanming says his team has been tuning and benchmarking domestic chips since 2018, testing virtually every major domestic architecture at cluster scale.
His core argument: domestic chips do not need to match high-end GPUs on every metric. They only need to meet prefill-stage requirements to deliver commercial value in production.
SenseCore's own data shows this disaggregated inference stack achieves 1.25× the cost-performance of mainstream high-end GPUs. That said, specialized workloads like ultra-long-context processing still require high-end GPUs.
What is the strategic significance of this system?
This reflects a deeper shift: domestic AI accelerators — previously seen as "low-ROI" — are being pulled into commercially viable inference pipelines through heterogeneous orchestration.
This means → it addresses two problems at once: easing the high-end GPU supply bottleneck and opening a scaled-deployment window for domestic chip makers.
Whether SenseTime can keep expanding its share depends on the stability and replicability of this heterogeneous system — the next phase hinges on whether a technical moat can become a commercial one.
市场有风险，内容仅供研究参考，不构成投资建议。