By Jin Lei, from Ofeisi
QbitAI | Official account: QbitAI
It's surprising, but also not surprising.
Here's what happened —
Frost & Sullivan and Leadleo Research have just released a report called"China Full-Stack AI Cloud Services Market Report 2026H1", and in theNeocloud (new cloud)track, nearly half of the market sharehas surprisingly been swept up by a single Chinese company。
. Neocloud, as the name suggests, is not like traditional comprehensive cloud vendors that built the cloud first and stacked AI on top; it is the kind that has been deeply bound to AI from its founding.
And this track is also very hot overseas — star companies like CoreWeave have backlogs reaching $104.2 billion; meanwhile, domestic comprehensive cloud giants are also shifting fully toward AI.
And just today,Muskalso commented on the Neocloud track, but what he poured on it was a bucket of cold water.
He said that in just the past few weeks, for companies without substantial revenue outside of AI data centers, the debt market has almost completely shut its doors to financing. It has a certainMatthew effectfeel to it:
"This means that 200+ Neoclouds will face a wave of large-scale failures and consolidation. Some will survive, but not many."
On one side, overseas Neoclouds are predicted to undergo a major shakeout; on the other, domestic comprehensive cloud giants are shifting fully toward AI. In such a landscape, it is indeed surprising that one Chinese company has taken nearly half the share.
△ Image generated by AI
So who exactly is this company?
No suspense — it isSenseTime, or more precisely, its AI infrastructure platformSenseCoreSenseTime's AI infrastructure platform.
Although the name Neocloud is new, SenseTime has been doing this for years. From serving its internal AI R&D, to offering cloud services to external customers in 2022, to organizing large-scale inference and continuously supplying tokens today, the product form of SenseCore has kept evolving.
Nowadays, rather than being a computing power service platform, SenseTime's SenseCore is more like a continuously expandingToken factory。
Because at the beginning of this year, SenseTime's SenseCore was serving roughly400 billion; by the end of the 7th month (July) it had risen to2.42 trillion; by the end of the 9th month (September), it had already reached6 trillion. According to plan, in the fourth quarter of this year it is set to surge to10 trillion per day！
△ Image generated by AI
But if you understand the development history behind SenseCore, you will find that it is actually not surprising that it captured nearly half of this share.
How did Neocloud become the new darling?
To answer this question, we might first recall the little crayfish (OpenClaw) that went viral at the beginning of the year.
At that time, although the crayfish's popularity remained sky-high, news like this kept coming to light:
An IT professional asked the crayfish to help search for information and write a 2000-word Word document, but after all the back-and-forth, 700 ten-thousand Tokens were gone…
The reason it burns through Tokens so fast is that Agents today do not operate in a one-question-one-answer mode; they need to break down tasks on their own, look up materials, and then repeatedly confirm and execute, consuming Tokens at every step.
As Frost & Sullivan pointed out in its report,At the same Token generation speed, Agentic AI requires more than 10 times the computing power of traditional large models.
△ Image generated by AI
And viewed nationwide, this number is only more exaggerated; for example, according to statistics from the National Data Administration, in March this year China's daily average Token call volume had already exceeded 140 trillion, an increase of more than a thousand times in two years.
According to SenseCore CTOXuan Shanming, for all of last year, almost everyone was busy training models; around this year's Spring Festival, the crayfish made Agents popular, and combined with the gap between domestic models and leading overseas models narrowing significantly, Token demand suddenly took off.
What has changed is not just the volume of demand, but also its direction.
Previously, the main battleground for large models was training, where everyone competed on whose cluster was bigger and who had more compute. But as AI workloads expand from training to inference and Agent applications, what customers care about has changed too: cost per Token, inference latency, model deployment efficiency, and whether diverse workloads can be flexibly scheduled have started to matter more than simply how much compute you can rent.
In other words, the unit of measurement for the compute business is shifting from "Flops" to "Token".
Neocloud has become the new rising star amid this shift. Unlike integrated cloud providers that had the cloud first and then layered AI capabilities on top, Neocloud has served AI exclusively from the very beginning. In the words of SenseTime Lijiang (AI Big Device) Product DirectorLu Guoqiang, the former is "cloud + AI", while the latter is "AI + cloud".
According to Synergy Research statistics, the global Neocloud market size reached 25 billion USD in 2025 and is expected to approach 400 billion USD by 2031.
△Image generated by AI
However, even though they share the same name, Neoclouds overseas and in China are traveling two completely different paths.
Take CoreWeave as an example: it started from "mining", already holding a large batch of high-end GPUs, and rode the wave of large model training to transform, with its main business being renting out large-scale high-end compute to model companies and tech giants.
This model is asset-heavy, and expansion depends heavily on financing; once the capital environment tightens, pressure quickly transmits through. This is also why Musk predicted that of the more than 200 Neoclouds, only a few would survive.
In addition, overseas Neoclouds currently focus more on AI training and are relatively weak at the inference layer. The biggest differences between SenseTime Lijiang and them lie in two points.
The first is inference-layer capability, that is, the ability to continuously convert complex, diverse compute into low-cost, stable Tokens—which is exactly what customers need most in the Agent era.
The second is the domestic compute environment. The supply of high-end GPUs is limited; domestic chips have developed rapidly in recent years, with Huawei, T-Head, Cambricon and other vendors each putting out their own products, but the routes are diverse, strengths differ, and the software ecosystems are still gradually maturing, so compute as a whole shows a fragmented character.
Stacking these two premises together, what Chinese Neoclouds need to solve is not just whether they have cards, but how to twist different chips, different models, and different customer needs into the same system that runs stably, while making every Token cheap enough.
SenseTime Lijiang began planning for this early. Xuan Shanming mentioned that they started doing adaptation and tuning for various domestic chips in 2018, and have gone through clusters of almost all mainstream domestic chips. What each chip is good at and which stage it is best placed in was all figured out by testing one chip at a time.
From a Single Card to a Kilowatt-Hour
With these two premises in mind, looking again at SenseTime Lijiang's approach makes things much clearer.
What it does is not single-point compute rental, but an end-to-end systematic capability of "compute-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 convert AI into real value with a higher return on investment.
Within this system, two things best reflect the characteristics of Chinese Neoclouds: one is heterogeneous mixed inference, and the other is compute-power coordination.
First, heterogeneous mixed inference. Facing domestic chips with different strengths, the core method adopted by SenseTime Lijiang is PD separation (Prefill-Decode Disaggregation).
As for PD separation, we can understand it as large models answering questions in two steps. The first is "reading the question", reading the entire input in one go, called Prefill, which mainly consumes compute; the second is "writing the answer", generating output word by word, called Decode, which values memory bandwidth more.
△Image generated by AI
Since the two steps have different hardware requirements, 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 along via the KV Cache.
In this way, the strengths of domestic cards are fully utilized, and their weaknesses no longer hold things back.
Behind this is a set of ideas that SenseTime Lijiang has figured out over many years. Xuan Shanming said that they have been working with domestic chips since 2018, know very well each chip's strengths and weaknesses, and also know which weaknesses actually don't matter in certain scenarios.
For example, in PD separation, domestic cards are only responsible for Prefill, so as long as their compute is sufficient, nothing else needs to be demanded; what the team has to do is conquer all the technical difficulties of the single point of "using domestic compute to run Prefill".
In other words, there is no need to wait until a domestic card matches high-end GPUs in every dimension before using it. Many domestic cards that, in outsiders' eyes, have a low return on investment can, when placed into this division of labor, actually leverage the output of high-end GPUs and multiply the overall Token capacity.
According to data provided by SenseTime Lijiang, the overall inference cost-performance ratio of this solution has reached1.25 times。
that of mainstream high-end GPUs. Of course, this division of labor also has boundaries: requests such as ultra-long contexts are still handled by high-end GPUs as a fallback to ensure service quality.
Having gotten the solution working, SenseTime Lijiang 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, "stretching" PD apart geographically, for example doing Prefill in city A and Decode in city B. Everywhere is now building intelligent computing centers, and many of the existing domestic compute resources are sitting idle, unused. If they can be connected with compute in other regions, these resources can be put to use.
This would have been unthinkable before. Xuan Shanming said that thanks to changes in the model architecture, the amount of transmitted data has dropped dramatically over the past few months, and combined with improvements in communication technology, cross-domain PD separation has become possible. Of course, not all requests are suitable for being sent to a remote location; which ones need to be processed locally immediately and which can be handed off remotely must first be classified. Currently, this technology is still in the validation stage.
The other is AF separation, which means, on top of PD separation, splitting the Attention and FFN computations inside the model apart as well, assigning each to the hardware best suited to it. In Xuan Shanming's view, for this direction the entire industry is still in the pre-research stage, because it cannot do without hardware-level support; software alone means little, so SenseCore is also working jointly with hardware vendors to advance it.
In either direction, many problems are ones 'you only discover when you actually go and do large-scale heterogeneous work'. Chen Xi, head of SenseTime's SenseCore platform, also noted that from network flow control to KV Cache transmission, if any single link is not properly tuned, it will affect the performance of the entire cluster. 'The same set of code, the same hardware configuration, placed at another company might not achieve this kind of result.'
Now let's turn to compute-power coordination. For Neoclouds, electricity bills are an unavoidable expense.
In response, SenseTime's SenseCore solution is theCompute-Power Coordination Agentreleased 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 get pushed back; at the same time, it dynamically adjusts the storage system's charge and discharge strategy across different time periods, coordinating and optimizing compute load and electricity costs without affecting task SLAs or adding extra supercapacitor costs.
△ Image generated by AI
Simply put, this approach is similar to running our washing machine during off-peak hours. Tasks like offline inference and model evaluation are moved to run when power load is low, while the energy storage system acts like a giant power bank, charging when electricity is cheap and discharging when it is expensive.
In July this year, when the Shanghai power grid initiated a demand response, the Lingang data center gave up 75% of its electricity capacity within two hours—a usually heavy power consumer became an elastic capacity pool the grid could call on at critical moments.
Working it out 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. SenseTime has accordingly proposed a new metric,TPW (Tokens Per Watt)which, unlike PUE that only focuses on whether a data center saves electricity, measures how many useful Tokens each watt of power actually delivers.
With both the compute and electricity bills accounted for, and Tokens being produced in ever greater numbers, the next question is whether anyone uses these Tokens and whether they can truly be put into business.
Getting Tokens onto the Industry Front Line
Besides serving SenseTime's own models, SenseCore's Token factory has already provided services to multiple external foundation model vendors.
Xuan Shanming compares this to a "wholesale service". Compared with "retail", wholesale may earn less per order, but it deals directly with "original manufacturers" and builds competitiveness through long-term reputation, technical capability, and reliable delivery. By first serving the leading foundation model vendors well and establishing a solid reputation, when negotiating with major clients later, they will naturally trust 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.
AI deployment has now entered deep water. Model capabilities are getting stronger and stronger, but truly integrating them into enterprises' actual production processes is far harder than imagined, and many industries face the awkward situation of "having a model but being unable to use it."
This is why SenseCore created the FDE role. Facing industry clients, SenseCore dispatches FDEs (Forward Deployed Engineers) on site to work with clients to sort out requirements, tune models, and drive deployment. According to Zhou Mingjun, director of SenseCore TokenHub, FDEs are neither as miraculous as online rumors suggest, nor outsourcing in the traditional sense.
Take the Shanghai Municipal Planning and Natural Resources Bureau'sQuantum Cityproject as an example. Urban planning and resources management involves large amounts of maps, planning drawings, statistical data, and professional reports; information across different businesses is complex and the professional barriers are high. They hope AI can not only answer policy and business questions, but also truly understand maps and planning drawings, perform spatial analysis and data statistics, and assist in generating professional reports.
SenseCore and the Shanghai Municipal Planning and Natural Resources Bureau jointly built the first foundation large model in the national planning and resources field—the "Yunyu Xingkong Large Model". The model has 600 billion parameters and is deeply embedded in planning and resources business processes, with capabilities such as "unanswerable questions, chart adjustment, statistics, image recognition, and report generation", covering the complete workflow from knowledge retrieval and drawing understanding to spatial analysis and decision support.
Around such projects, SenseTime has summarized a delivery rule called"AI123": clarify requirements in 1 day, produce a Demo in 2 days, and launch a stable version within 30 days. Zhou Mingjun said this speed is not just talk—it is backed by 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 FDEs go deep into the business front line, "accompany" clients throughout the process, quickly produce benchmark cases, and then distill them into standard products: "We don't want to be a labor-intensive enterprise."
Beyond that,efficiencyis another keyword repeatedly mentioned in conversations with the SenseCore team.
According to Frost & Sullivan's judgment, in the next three years, AI full-stack cloud providers will move from a resource race into a race of efficiency and ecosystem. Xuan Shanming used an analogy: training compute used to be a bit like real estate—buy the cards, rent them out, and collect a fixed return each month; now, doing Tokens, the returns fluctuate greatly—you might lose money, or earn a great deal, and what makes the difference is efficiency.
So, looking back now at SenseCore's result of capturing nearly half the market share, it is clear that it was not achieved by simply piling cards into data centers.
From adapting domestic chips one by one starting in 2018, to using PD disaggregation so different chips each play their role, to calculating exactly how many Tokens each kilowatt-hour can deliver—looking at it this way, the result is hardly surprising.
As for the direction going forward, Xuan Shanming said they will devote more energy and resources to domestic chips, moving forward together with domestic chips and the domestic ecosystem.
At this moment, the rows of domestic accelerator cards in the Lingang data center are probably busy reading problems for some Agent.