IDC and Frost & Sullivan Name China AI Infrastructure Leaders as Baiyao Releases AI Virtual Cell Report

IDC evaluated mainstream AI compute management platform providers in China across six dimensions: compute resource management, compute service management, compute operations, technical architecture, business scenario support and delivery services. It said that as AI applications enter large-scale deployment, agent technologies are reshaping the core paradigm of cloud infrastructure, and AI infrastructure is evolving from a basic compute platform into core IT infrastructure for enterprise intelligent transformation. Paradigm received full scores in four dimensions, the report said.

Paradigm founder and CEO Dr. Dai Wenyuan said the AI industry is moving from a 'model race' to a new stage of 'large-scale deployment.' He said the real challenge for enterprises is no longer whether they can use large models, but how to continuously turn agent capabilities into business productivity under controlled costs and stable operations. He added that unified management, flexible scheduling and refined operations of heterogeneous compute are becoming important capabilities for enterprise AI infrastructure. Paradigm said its vendor-neutral management software supports mainstream models running on chips from different brands and has adapted to more than ten mainstream chip vendors. It supports mixed deployment of large-scale training and high-concurrency inference. Its core capabilities include zero-intrusion multi-chip virtualization, which allows a single AI accelerator card to be divided among multiple tasks; support for PyTorch, TensorFlow, vLLM and SGLang without modifying business code; and resource quota mechanisms. It also offers three resource pools—exclusive, shared and overcommitted—and four-dimensional scheduling based on task priority, hardware topology, real-time load and resource demand. An AI gateway manages model routing, API calls and auditing, with dual-track metering based on card-hours plus tokens. The platform packages heterogeneous compute into standardized services for agents, supports on-demand partitioning, reuse and elastic scaling, and uses token quotas and circuit breakers to control cost and risk. Its operations cover nodes, accelerator cards, tasks and model services, supporting fault isolation, task rescheduling and private deployment without public internet access. The IDC report mentioned that a policy bank used Paradigm's three-pool architecture to unify compute management across two locations and three data centers. Paradigm's platform has adapted to and managed compute cards from more than ten vendors, including Ascend, Hygon DCU, Cambricon, Iluvatar CoreX, MetaX, Moore Threads, Kunlun Core, Enflame, T-Head and Biren, according to the company.

IDC said AI infrastructure is undergoing a deep transformation from the traditional 'cloud + AI' architecture to a 'cloud-native + Agent collaboration' architecture. AI compute management platforms will gradually evolve from a focus on compute management to comprehensive capabilities covering compute pooling, partitioning, scheduling and operations, according to the report.

In the Neocloud report, Frost & Sullivan and Head Leopard Research classified full-stack AI cloud vendors into two groups. One is hyper-scalers such as Alibaba Cloud and Baidu AI Cloud, which are adding AI to comprehensive cloud businesses. The other is Neocloud providers that were built with AI from the start and have developed into full-stack AI cloud vendors, represented by SenseTime, CoreWeave and Nebius. The report said 2026 is the first year of Agentic AI. Compared with traditional large language models, Agentic AI requires more than ten times the compute at the same token generation speed, placing new demands on power supply, cooling, large-scale parallel computing and scalability. It said Neocloud providers combine high-performance computing, high-speed interconnect, flexibility, security and scalability, and can carry AI training, inference and agentic applications. SenseTime's Dazhuangzhi was built around AI computing needs and has iterated with models and applications, the report said. It offers end-to-end capabilities covering compute, platform, solutions and services. For large-scale clusters, it uses breakpoint-resume training, automatic fault tolerance, task migration and topology-aware scheduling to maintain stability during node fluctuations, resource switching and dynamic expansion. It also introduced a compute-power coordination agent to align compute, energy and business operations for system-level cost and efficiency improvements in intelligent computing centers. On domestic compute, the report said SenseTime's Dazhuangzhi uses PD-disaggregated heterogeneous hybrid inference, achieving inference price-performance 1.25 times that of NVIDIA H-series and 2.5 times the token output of domestic homogeneous inference at the same cost. Its platform served an average of 4.5 trillion tokens per day as of August 2026 and is expected to reach 10 trillion tokens per day in the fourth quarter, according to SenseTime. The company also uses an 'end-to-end multimodal solutions plus FDE' model to work with customers on scenarios, models and goals.

Frost & Sullivan predicted that over the next three years, full-stack AI cloud vendors will move from a 'resource race' into an 'efficiency and ecosystem race.' SenseTime said it launched the Galaxy Plan with nearly 20 ecosystem partners at the WAIC 2026 forum to promote domestic AI infrastructure ecosystem development, and released a series of innovations at WAIC 2026 aimed at cost-performance, adaptability and energy efficiency.

On the same day, Baiyao Technology released the AI Virtual Cell (AIVC) Technology Trends, Industry Ecosystem and Application Prospects Research Report—New Infrastructure for Life Sciences in the AI Era. The report was initiated by Baiyao Technology, researched and written by MIT Technology Review China, and completed under the guidance of the Zhongguancun Science City Management Committee, according to the company. Baiyao said the report, based on interviews with dozens of research institutions, technology companies and pharmaceutical companies worldwide, defines AIVC as new infrastructure for life sciences in the AI era and maps its technology routes, industry landscape and future trends. The report asks how AIVC can move from capability metrics in papers to real infrastructure in drug development and life-science industries. It says that while AI tools such as AlphaFold have turned protein structure prediction into a computational problem and improved molecular-level drug R&D efficiency, success at the molecular level does not guarantee effectiveness at the cellular level. AIVC aims to build a 'digital mirror' of cells in computational space, moving research from the molecular level to the cellular level. In this paradigm, cells can be represented, simulated, intervened and continuously calibrated, and the system can predict how a cell changes after a gene is knocked out or a drug is added. The report says AIVC could become reusable computational models and tool platforms for pharmaceutical companies, research institutions and medical institutions, covering basic research, target discovery, drug development and cell engineering. As data quality, model generalization, experimental validation and standards improve, it could expand into clinical research, patient stratification and clinical decision support.

The report identifies three trends. Model competition is shifting from scale expansion to generalization and task adaptation; simply increasing data or parameters does not necessarily improve zero-shot generalization. Private data and newly produced data are becoming key competitive factors, as public data alone cannot support differentiation. And dry-wet closed loops—data production, model prediction, experimental validation and data feedback—are becoming a core threshold for industrialization. The report lists Baiyao Technology among innovative companies using AIVC as a core product or R&D engine, alongside Xaira, Recursion, Cellarity, Noetik and Turbine. In June, Baiyao released AURA CellOS, described as the world's first AI virtual cell world model based on an LLM-JEPA architecture and one of the largest single-cell foundation models publicly reported. The company said CellOS's main breakthrough is architectural innovation, with expression and perception perspectives working together to learn stable features of cell state changes in latent space. It said public evaluations show CellOS leads mainstream models by multiple times in prediction accuracy and perturbation modeling. Baiyao is building a data factory with real samples and self-produced perturbation data, including a TM-Perturb-seq high-throughput multimodal perturbation experiment platform and clinical cooperation resources for tumor microenvironment and clinical data. It plans to complete a dry-wet closed loop and start validation in tumor clinical patient stratification, CAR-T optimization and skin anti-aging target discovery with pharmaceutical and research partners.