Chinese AI Infrastructure Providers Push Token Productivity and On-Demand Supercomputing for Enterprises

Tang presented the formula “Token productivity = capacity × utilization × delivery quality.” Capacity is determined by chip hardware and model optimization; utilization depends on traffic governance and intelligent compute scheduling; delivery quality covers inference accuracy, tool-call success rates and SLA stability in production. SiliconFlow says it works on the conversion layer between raw compute and tokens, managing compute uniformly and using its own inference engine to run models faster, both in public-cloud token factories and in token factories built inside central state-owned enterprises.

SiliconFlow has designed a four-layer decoupled architecture to connect model operations, model maintenance, inference execution and compute scheduling. The model gateway handles metering, billing and authentication for all model calls and routes requests by response-time requirements, performance priority or cost priority. The model service platform manages model listings and a model repository, with one-click deployment practices and adapters for nearly 100 models. The high-performance inference framework works with chip vendors on operator and communication libraries; with Huawei Cloud, SiliconFlow launched what it described as the industry’s first token service for DeepSeek R1/V3 on domestic chips and optimized inference for CloudMatrix clusters, reaching more than 1,920 tokens/s decode throughput per card on DeepSeek-R1. It also deployed Kimi-K2 on a Muxi C550 cluster and DeepSeek-V3 on Moore Threads MTT S5000 chips, with prefill throughput above 4,000 tokens/s and decode throughput above 1,000 tokens/s per card, according to the presentation. The compute scheduling layer provides elastic scaling and supports time-of-day pricing, with discounts for nighttime use.

In one central state-owned enterprise group, about 50 machines served subsidiaries that applied for machines and deployed their own models, creating 40 to 50 model instances and low utilization. SiliconFlow and the client consolidated repeated models to about 10 and replaced machine allocation with Serverless model services called through APIs. The same 50 machines then carried existing small models and freed capacity for larger models including GLM-5.2 and DeepSeek-V4-Flash-0731. Output speed stayed above 30 tokens/s across tests from single to higher concurrency, and same-model tests showed 40% to 100% performance gains depending on input and output length. Cost per million tokens for a Qwen model fell from more than 100 yuan to below 10 yuan, according to the presentation.

SiliconFlow is also working with customers on application and infrastructure co-optimization. In coding, poorly constructed context can mean hundreds of thousands of input tokens miss the cache; better construction can hit about 90%, raising concurrency from 100 to 200 without adding hardware. Preprocessing code repositories during nightly builds and storing results in KV Cache can shift work away from daytime peaks.

Lü told Leiphone.com that 70% to 80% of Parallel Technology’s business came from research and universities five or six years ago, while this year more demand comes from enterprises. IDC’s 2026 China Artificial Intelligence Computing Power Development Assessment Report, cited in the article, expects China’s intelligent computing scale to reach 2,576.5 EFLOPS in 2026, up 87.9% year on year, and exceed 10,000 EFLOPS by 2030. Self-built clusters are costly and hard to scale for peak research needs, while university supercomputing centers have limited resources; one car company’s CAE engineer faced simulation tasks backlogged for two weeks and queues starting at 12 days, according to the article.

Parallel Technology promotes a “compute at night, review in the morning” model, in which engineers submit long-running calculations after work and see results the next morning. Lü described the service as being like electricity: users pay for what they use and are not charged when idle. The company has also launched a token factory for inference, providing compute and platform services without entering customers’ models or applications. For latency-sensitive applications, compute is deployed near users; for coding and other tasks that can tolerate delay, it can run in western data centers where electricity is cheaper.

Parallel Technology has connected about 63 computing centers, including more than 40 intelligent computing centers and a dozen or so supercomputing centers, into a unified pool through scheduling software. Its application runtime analysis aims to determine within four hours whether a new customer’s workload is compute-intensive or memory-intensive and to find the most cost-effective resource. Lü said the model requires pricing and settlement arrangements that let partners profit, and Parallel Technology positions itself at the IaaS and PaaS layers rather than in customer applications. A neutral supplier that does not make models is valuable to model developers worried about losing compute access if a large vendor competes with them, he said.

Auto manufacturing and new energy account for nearly 40% of the business unit’s revenue, Lü said. CAE simulations for vehicle design sometimes require 512 to more than 1,000 cores working at once, while AI adds GPU training for autonomous driving, AI-assisted coding and embodied intelligence. Parallel Technology is proposing to combine supercomputing and intelligent computing for these workloads. Lü said AI’s efficiency gains are still at an early stage and may take three, five or ten years to fully appear.