Frost & Sullivan Report: SenseTime's AIDC Captures Nearly Half of China's Neocloud Market

It was surprising, but also not surprising.

Here's what happened—

Frost & Sullivan and LeadLeo Research Institute recently released a report titled "China Full-Stack AI Cloud Service Market Report 2026H1," and in the Neocloud (new cloud) segment, nearly half of the market share was actually captured by a single Chinese company.

And this segment is also very hot overseas; star companies like CoreWeave have backlog orders alone reaching $104.2 billion; in addition, domestic comprehensive cloud giants are also fully pivoting to AI.

And just today, Musk also commented on the Neocloud segment, but what he poured was a bucket of cold water.

This means more than 200 Neoclouds will face a wave of large-scale bankruptcies and mergers. Some will survive, but not many.

On one hand, overseas Neoclouds are predicted to face a major reshuffle; on the other hand, domestic comprehensive cloud giants are fully pivoting to AI. Under such a landscape, a Chinese company being able to take nearly half the share is therefore very surprising.

So who exactly is this company?

Without keeping you in suspense, it is SenseTime, or more precisely, its AI infrastructure platform SenseCore.

Although the name Neocloud is new, SenseTime has been doing this for years. From serving internal AI R&D, to providing cloud services to external customers in 2022, to today organizing large-scale inference itself and continuously supplying Tokens, the product form of SenseCore has been constantly changing.

Nowadays, rather than saying SenseCore is a computing power service platform, it is more like a continuously expanding Token factory.

Because at the beginning of this year, SenseCore was serving about 400 billion Tokens per day; by the end of July it had risen to 2.42 trillion; by the end of September, it had already reached a scale of 6 trillion. According to the plan, in the fourth quarter of this year it is set to surge to an average of 10 trillion per day!

But if you understand the development history behind SenseCore, you will find that its ability to capture nearly half the share is actually not surprising.

To answer this question, we might first recall the OpenClaw that went viral at the beginning of the year.

At that time, although OpenClaw remained highly popular, news like this frequently emerged:

An IT professional asked OpenClaw to help search for materials and write a 2,000-word Word document, and after all the back-and-forth, 7 million Tokens were gone...

The reason it burns Tokens so heavily is that Agents doing tasks today are not in a question-and-answer mode; they need to break down tasks themselves, look up materials, and then repeatedly confirm execution, with every step consuming Tokens.

As Frost & Sullivan pointed out in the report, at the same Token generation speed, Agentic AI has more than 10 times the computing power requirements of traditional large models.

And this number becomes even more exaggerated when viewed nationwide. For example, according to statistics from the National Data Administration, in March this year China's average daily Token calls had already exceeded 140 trillion, an increase of more than a thousandfold in two years.

According to Xuan Shanming, CTO of SenseCore, throughout last year almost everyone was busy training models; then around this year's Spring Festival, OpenClaw made Agents popular, and coupled with the obvious narrowing of the gap between domestic models and overseas leading models, Token demand suddenly took off.

What changed was not only the volume of demand, but also its direction.

Previously, the main battlefield for large models was training, and everyone competed on whose cluster was larger and who had more computing power. But as AI workloads expanded from training to inference and Agent applications, what customers care about also changed: unit Token cost, inference latency, model deployment efficiency, and whether multiple workloads can be flexibly scheduled began to matter more than simply how much computing power one could rent.

In other words, the unit of measurement for the computing power business is changing from "Flops" to "Token."

Neocloud became the new favorite amid this wave of change. Unlike comprehensive cloud vendors, which first have cloud and then layer AI capabilities on top, Neocloud serves AI only from the very beginning. In the words of Lu Guoqiang, general manager of SenseCore products, the former is "cloud + AI," while the latter is "AI + cloud."

According to Synergy Research statistics, the global Neocloud market reached $25 billion in 2025 and is expected to approach $400 billion by 2031.

However, although both are called Neocloud, overseas and China are taking two completely different paths.

Take CoreWeave as an example. It started out in "mining" and already had a large stock of high-end GPUs on hand. Riding the tailwind of large model training, it transformed accordingly, and its main business is renting large-scale high-end computing power to model companies and tech giants.

This model is asset-heavy, and expansion is highly dependent on financing. Once the capital environment tightens, pressure will quickly transmit through the chain, which is also why Musk predicts that few of the more than 200 Neoclouds will survive.

In addition, overseas Neoclouds currently focus more on AI training and are relatively weak at the inference layer. The biggest differences between SenseCore and them lie in two points.

First is inference-layer capability, that is, the ability to continuously convert complex and diverse computing power into low-cost, stable Tokens, which is exactly what customers need most in the Agent era.

Second is the domestic computing power environment. The supply of high-end GPUs is limited. Although domestic chips have developed rapidly in recent years, with Huawei, T-Head, Cambricon and other manufacturers all bringing out their own products, the routes are diverse, the strengths differ, and the software ecosystem is still gradually maturing, so computing power as a whole shows a fragmented character.

With these two premises combined, what China's Neoclouds need to solve is not merely whether cards are available, but how to screw different chips, different models and different customer needs into the same system that can run stably, and make every Token cheap enough.

SenseCore began laying out its plans for this early on. Xuan Shanming mentioned that they started doing adaptation and tuning for various domestic chips in 2018, and have tried out the clusters of almost all mainstream domestic chips. What each chip is good at and which link it is best suited for were all found out through testing one by one.

With these two premises, looking again at SenseCore's approach makes things much clearer.

What it does is not single-point computing power rental, but an end-to-end systematic capability of "computing power - platform - solution - service." Downward, it connects power, energy storage, AIDC and heterogeneous hardware; upward, it connects models, applications and industries, with the goal of helping enterprises truly turn AI into value with a higher return on investment.

In this system, two things best reflect the characteristics of China's Neoclouds: one is heterogeneous mixed inference, and the other is computing-power-electricity coordination.

First, heterogeneous mixed inference. Facing domestic chips with different strengths, the core method adopted by SenseCore is PD separation (Prefill-Decode Disaggregation).

PD separation can be understood as dividing the process of a large model answering a question into two steps. First is "reading the question," reading the input in one go, called Prefill, which mainly consumes computing power; then "writing the answer," generating it word by word, called Decode, which places more emphasis on memory bandwidth.

Since the two steps have different requirements for hardware, PD separation splits them apart and assigns them to different chips. Domestic cards handle reading the question, high-end GPUs handle writing the answer, and in between the "reading notes" are passed over through KV Cache.

In this way, the strengths of domestic cards are fully used, and their weaknesses no longer hold them back.

Behind this is a set of ideas SenseCore has explored over many years. Xuan Shanming said that they have dealt with domestic chips since 2018 and know very clearly which point of each chip is a strength and which is a weakness, and also know that some weaknesses are actually not affected in certain scenarios.

For example, in PD separation, if a domestic card is only responsible for Prefill, then as long as its computing power is sufficient, other aspects need not be demanded too strictly. What the team needs to do is overcome all the technical difficulties in this single point of "using domestic computing power to run Prefill."

That is to say, there is no need to wait for a domestic card to catch up with high-end GPUs in every dimension before using it. Many domestic cards that outsiders regard as having low return on investment, when placed into this division of labor, can instead leverage the output of high-end GPUs and multiply overall Token capacity.

According to data provided by SenseCore, the overall inference cost-effectiveness of this solution has reached 1.25 times that of mainstream high-end GPUs.

Of course, this division of labor also has boundaries. Requests such as ultra-long context are still backed up by high-end GPUs to ensure service quality.

After the solution was proven, SenseCore is already looking at the next step. Xuan Shanming mentioned that they are currently focusing on two directions.

One is cross-domain PD separation, that is, "pulling PD far apart," for example doing Prefill in city A and Decode in city B. Intelligent computing centers are now being built everywhere, and a considerable amount of existing domestic computing power sits idle and cannot be used. If it could be connected with computing power in other regions, these resources could be revitalized.

This was unthinkable before. Xuan Shanming said that thanks to changes in model architecture, the volume of transmitted data has dropped sharply over the past few months, and coupled with improvements in communication technology, cross-domain PD separation has become possible. Of course, not all requests are suitable for sending far away; which need to be processed locally immediately and which can be handed over remotely must first be classified. At present, this technology is still in the validation stage.

Another is AF separation, that is, on the basis of PD separation, also splitting the Attention and FFN computation inside the model and handing them to more suitable hardware. In Xuan Shanming's view, the entire industry is still in the preliminary research stage for this direction, because it cannot be separated from support at the hardware level, and doing only software is of little significance, so SenseCore is also advancing it jointly with hardware manufacturers.

Whichever direction it is, many problems "only emerge when you actually go and do large-scale heterogeneity." Chen Xi, head of SenseCore's SenseCore platform, also mentioned that from network flow control to KV Cache transmission, if any link is not tuned well, it will affect the performance of the entire cluster, and "the same set of code and the same hardware configuration, placed at another company, may not be able to produce this effect."

Next, look at computing-power-electricity coordination. For Neoclouds, electricity bills are an unavoidable account.

In response, SenseCore's approach is the "compute-power coordination Agent" released at this year's WAIC. It links an energy large model with CATL's energy storage system, making a prediction every 15 minutes and a correction every 5 minutes to decide which tasks run now and which can be pushed back; at the same time, it dynamically adjusts the charging and discharging strategy of energy storage across different time periods, coordinating and optimizing computing load and electricity costs without affecting task SLAs or incurring additional overcapacity costs.

Simply put, this approach is similar to us rushing to run the washing machine during off-peak electricity hours. Tasks such as offline inference and model evaluation are moved to times when the power load is low, while the energy storage system acts like a giant power bank, charging when electricity prices are low and discharging when prices are high.

In July this year, when the Shanghai power grid initiated demand response, the Lingang data center gave up 75% of its electricity capacity within two hours. A major electricity consumer in normal times became, at a critical moment, a pool of flexible capacity that the grid could call upon.

Calculated this way, Token output per unit of electricity cost increased by 80%, and the average electricity bill is 10% lower than that of IDCs in the same region. SenseCore also took the opportunity to propose a new metric, TPW (Tokens Per Watt). Unlike PUE, which focuses only on whether a data center saves electricity, it measures how many effective Tokens each watt of electricity actually yields.

The accounts for computing power and electricity bills are now clear, and more and more Tokens are being produced. The next question is whether anyone uses these Tokens and whether they can truly be applied in business.

In addition to serving SenseTime's own models, SenseCore's Token factory already provides services to a number of external foundation model vendors.

Xuan Shanming likens this to "wholesale services." Compared with "retail," wholesale may earn less per transaction, but it can connect directly with the "original manufacturer" and build competitiveness through long-term reputation, technical capability, and stable delivery. Serve the leading foundation model vendors well first and establish a solid reputation; then when talking to major clients, they will naturally believe that the goods you supply are genuine and not watered down.

However, selling Tokens to model vendors is one thing; getting AI truly into every industry is another.

This is also why SenseCore does FDE. When facing industry clients, SenseCore dispatches FDEs (forward-deployed engineers) to be on site, working with clients to sort out requirements, tune models, and drive deployment. According to Zhou Mingjun, director of SenseCore TokenHub, FDE is neither as miraculous as it is rumored online nor traditional outsourcing.

Take the "Quantum City" project of the Shanghai Municipal Planning and Natural Resources Bureau as an example. Urban planning and resource management involves a large number of maps, planning drawings, statistical data, and professional reports, and information across different businesses is complex, with a very high professional threshold. They hope AI can not only answer policy and business questions, but also truly understand maps and planning drawings, complete spatial analysis and data statistics, and assist in generating professional reports.

SenseCore and the Shanghai Municipal Planning and Natural Resources Bureau jointly built the country's first foundation large model in the planning and resources field, the "Yunyu Xingkong Large Model." The model has 600 billion parameters and is deeply embedded in the planning and resources business processes. It has capabilities such as "stumping any question, able to adjust maps, good at statistics, able to recognize images, and able to write reports," and can cover the complete workflow from knowledge retrieval and drawing understanding to spatial analysis and decision support.

Around such projects, SenseCore has summarized a set of "AI123" delivery rules: clarify requirements in 1 day, produce a Demo in 2 days, and launch a stable version in 30 days. Zhou Mingjun said this speed is not just talk; behind it are a pre-built AI middle platform and multimodal data processing pipelines.

So, will the FDE model turn SenseCore into a systems integrator?

Xuan Shanming said that FDE goes deep into the business frontline, "accompanies" clients throughout the process, quickly produces benchmark cases, and then distills them into standard products. "We do not want to be a labor-intensive enterprise."

In addition, the word efficiency was also a keyword repeatedly mentioned during exchanges with the SenseCore team.

According to Frost & Sullivan's judgment, over the next three years, full-stack AI cloud vendors will move from a resource competition to a competition over efficiency and ecosystem. Xuan Shanming made an analogy: in the past, doing training computing power was somewhat like real estate—buy cards, rent them out, and collect a fixed return every month; now, doing Tokens, returns fluctuate greatly—you may lose money, or you may earn a lot—and what creates the gap is efficiency.

So, looking back now at SenseCore's achievement of taking nearly half the share this time, it is clear that it was not built by stuffing cards into data centers.

From adapting domestic chips one by one starting in 2018, to using PD separation so that different chips each perform their own roles, to calculating how many Tokens each kilowatt-hour of electricity can yield—looking at it this way, it is no surprise.

As for the next direction, Xuan Shanming said they will put more energy and resources into domestic chips and move forward together with domestic chips and the domestic ecosystem.

At this moment, the rows of domestic cards in the Lingang machine room are probably reading questions for some Agent.