IDC Releases 2026 China AI Compute Management Platform Assessment: Fanxiang Intelligence Tops Overall

IDC released the 'China AI Compute Management Platform Technical Capability Assessment, 2026,' listing resource management, service management, operations, technical architecture, business scenario support, and delivery services as six major assessment dimensions. Paradigm Intelligence (formerly Fourth Paradigm) ranked first overall and scored full marks in four dimensions: resource management, operations, technical architecture, and delivery services.

The ranking itself is a conclusion, but the real pain point the report touches on is: how much hardware is added to an enterprise's books versus how much compute the business can actually mobilize—there is an entire set of underlying engineering capabilities in between. Around this assessment, what is worth questioning is—what kind of capabilities does an enterprise need so that the hardware compute on its procurement list truly becomes a service the business can use at any time.

For supply chain security, enterprise data centers often mix and match compute cards from multiple brands. But hardware entering the facility is only the first step: underlying drivers, compilation environments, operator libraries, and inference engines all differ. Due to the lack of a cross-chip unified abstraction layer, different cards can only each build independent 'chimney clusters.'

The consequence is that clusters cannot dynamically support each other—models running on chip A are difficult to migrate to chip B; one side queues under overload while the other idles. Enterprises cannot afford the risk of binding to a single hardware vendor. The platform must, downward, smooth over driver and runtime differences across chips, and upward, allow mainstream models to run without code changes.

Traditional scheduling mostly uses 'whole cards' as the minimum allocation unit. A lightweight debugging or inference task, even if it uses very little memory, must monopolize a large compute card, causing 'false resource overflow' where the books show full but in reality most is idling.

Distributed training scenarios are even more sensitive: without strong cross-node collaborative scheduling, once some nodes wait for resources, the entire expensive cluster is dragged into inefficient internal friction. This requires introducing a gang scheduling mechanism at the scheduling layer, packaging all nodes of a task (i.e., the PodGroup pattern in cloud native), strictly enforcing 'either all satisfied or none dispatched,' and fundamentally eliminating the resource internal friction of stop-and-go execution.

In the past, large models were mostly steady-state computations with fixed batch processing. But after entering the Agent era, business calls evolved into pulse-like tidal surges with long chains, high-frequency reflection, and multi-model linkage.

If the underlying compute still relies on static allocation and manual reservation, either one blindly stacks cards to cope with unpredictable peaks and bears huge idle costs, or when sudden traffic floods in, the lack of dynamic elasticity directly causes business lag or even downtime.

Against the assessment dimensions given by IDC, the keys to engineering implementation can be grouped into three layers of capability. The first layer is fine-grained partitioning and secure isolation: coarse whole-card allocation inevitably leads to waste, but partitioning is by no means simply proportional slicing; the core difficulty lies in strong isolation—both squeezing idle compute through resource sharing and preventing task preemption from dragging down core business.

The open-source scheduler HAMi was promoted to a CNCF (Cloud Native Computing Foundation) incubation-level project in July of this year. The article states that the project is mainly contributed by the Paradigm team, focusing on dynamic elastic partitioning, eliminating resource misjudgments during scheduling, and enabling coordination with batch scheduling engines.

The second layer is gang collaboration for large-scale tasks, i.e., the aforementioned PodGroup-style packaged scheduling. The third layer is cross-chip neutral abstraction and multi-track scheduling. The article states that Paradigm materialized this logic into three resource pools: 'exclusive, shared, and overcommit': core training runs in the exclusive pool for stability, high-concurrency inference introduces the overcommit pool to extract elasticity, and then dynamically dispatches based on topology and load. The article also mentions that it is said a large policy bank, through this architecture, opened up a heterogeneous compute pool under the high compliance requirements of two locations and three centers.

When heterogeneous chips are pooled and partitioned, how to measure the business value of the rescued compute becomes the next question. The article states that the metering logic of AI infrastructure is shifting from 'renting machines by card-hours' to 'billing by Token usage'; in Agent high-frequency concurrency scenarios, business departments care more about task invocation throughput, latency, and the Token cost per inference.

The article states that Paradigm introduced dual-track metering of 'card-hours + Tokens' through an AI gateway: downward, it clarifies hardware occupancy consumption; upward, it measures the invocation output and quota circuit-breaking of model services. The article also states that in the first half of 2026, Paradigm Intelligence's API business revenue reached 464 million yuan, a year-on-year increase of 860.8%; and that relying on HAMi vGPU virtualization technology, comprehensive GPU utilization increased from less than 30% to 70%-90%. These figures come from the article's citations, with no first-party financial report original text seen, and the measurement caliber and scenarios are not explained.

The article judges that the era of competing on hardware specifications is passing, and the more critical threshold for enterprise AI implementation lies in using software to activate idle chips; and believes that Paradigm is shifting from delivering privatized software to becoming an efficiently operating 'Token factory' relying on the scheduling foundation, activating heterogeneous chips into standardized Token capacity. These two points are the article's judgments and inferences, not independently verifiable facts.

The boundaries that need to be preserved are equally clear: the specific scoring method of the IDC assessment, the range of participating vendors, and the complete ranking are not disclosed in the article; the specific score standards for full marks in four dimensions are also not explained; the exact date and review details of HAMi's promotion to a CNCF incubation-level project are not given; the 464 million yuan revenue and 860.8% growth rate, and the GPU utilization range are all cited by the article; the policy bank case is only expressed as 'it is said,' with the bank's name and verification details missing. Therefore, what this article can confirm are the report's dimensions and ranking conclusions, as well as engineering constraints such as heterogeneous barriers, whole-card waste, gang scheduling, and Agent tidal surges; parts involving specific benefit figures and customer cases should be regarded as source claims rather than established facts.