SenseCore Takes Nearly Half of China's Neocloud Market, Report Says

Neocloud providers are built around AI from the start, unlike general cloud vendors that add AI services on top of existing cloud businesses, according to the report. QbitAI reported that CoreWeave, a prominent overseas Neocloud company, has a backlog of $104.2 billion. It also reported that Elon Musk said in recent weeks the debt market had almost completely closed to companies without substantial revenue outside AI data centers, a situation QbitAI said points to a wave of closures and mergers among more than 200 Neocloud companies, with few expected to survive.

SenseCore's token volumes illustrate the demand. QbitAI reported that the platform served about 400 billion tokens per day at the start of this year, 2.42 trillion by the end of July and 6 trillion by the end of September. The plan is to reach 10 trillion tokens per day in the fourth quarter. SenseTime began offering cloud services to external customers in 2022 after using the platform for internal AI research and development. The product has changed from an internal compute platform into what QbitAI described as a growing token factory.

The report said Agentic AI requires more than 10 times the compute of traditional large models at the same token generation speed. The National Data Administration's statistics showed China's daily token calls exceeded 140 trillion in March this year, more than 1,000 times higher than two years earlier, according to QbitAI. Xuan Shanming, CTO of SenseCore, recalled that the industry spent most of last year training models. Around this year's Spring Festival, OpenClaw popularized agent applications and the gap between Chinese models and leading overseas models narrowed, and token demand rose quickly, he said, according to QbitAI. Customers began to care more about unit token cost, inference latency, model deployment efficiency and flexible scheduling across workloads than about how much raw compute they could rent. The unit of measurement in the compute business was shifting from Flops to tokens.

Synergy Research data cited in the article put the global Neocloud market at $25 billion in 2025 and projected it to approach $400 billion by 2031. CoreWeave started in mining, accumulated high-end GPUs and shifted to renting large-scale high-end compute to model companies and technology giants, a capital-intensive model that relies heavily on financing. China's environment differs. High-end GPU supply is limited, and domestic chips from Huawei, T-Head and Cambricon follow different routes with different strengths; their software ecosystems are still maturing, leaving compute fragmented. SenseCore began adapting and tuning domestic chips in 2018 and has tested clusters built on almost all mainstream domestic chips, Xuan said. The team learned chip by chip what each device does best and where it fits.

SenseCore's approach is an end-to-end system spanning compute, platform, solutions and services, connecting power, energy storage, AIDC and heterogeneous hardware on one side and models, applications and industries on the other, according to the article. A central method is PD disaggregation, which separates prefill, the compute-heavy reading of input, from decode, the memory-bandwidth-sensitive generation of output. Domestic chips handle prefill while high-end GPUs handle decode, with KV Cache carrying the intermediate data between them. SenseCore said the overall inference cost-performance of this arrangement is 1.25 times that of mainstream high-end GPUs. Long-context requests are still handled by high-end GPUs to maintain service quality. The company is validating cross-region PD separation, which would let one city perform prefill and another decode, and is researching AF separation with hardware vendors. Chen Xi, head of the SenseCore platform, said the same code and hardware configuration may not deliver the same results at another company because network flow control, KV Cache transmission and other links all affect cluster performance.

On power costs, SenseCore launched a compute-power coordination Agent at this year's WAIC. It links an energy model with CATL's energy storage system, forecasts every 15 minutes and corrects every five minutes to decide which tasks run now and which can be deferred, while adjusting storage charging and discharging. During a demand response initiated by the Shanghai power grid in July, the Lingang data center released 75 percent of its power capacity within two hours. SenseCore said token output per unit of electricity cost rose 80 percent, and average electricity costs were 10 percent lower than those of data centers in the same region. It also proposed TPW, or tokens per watt, as a metric for how much useful token output each watt produces.

The token factory serves SenseTime's own models and several external foundation model providers. Xuan compared the business to wholesale: lower margin per unit than retail, but direct access to model developers and competition based on long-term reputation, technical capability and stable delivery. For industry customers, SenseCore sends forward-deployed engineers, or FDEs, to work on-site with clients. In a project with the Shanghai Planning and Natural Resources Bureau, the two sides built Yunyu Xingkong, described as China's first foundation model for planning resources, with 600 billion parameters. The model is embedded in planning workflows and covers knowledge retrieval, drawing understanding, spatial analysis and decision support. SenseCore sums up delivery as AI123: one day to clarify requirements, two days to produce a demo and 30 days for a stable version to go online. Xuan said SenseCore does not want to become a labor-intensive systems integrator. The report said that over the next three years, full-stack AI cloud vendors will move from resource competition toward efficiency and ecosystem.