A small but concrete GPU-rental award in China: Hubei University of Science and Technology selected China Telecom’s Xianning branch for a three-year GPU-server rental for its data-center compute buildout.
The university’s Nov. 17 award notice lists RMB 639,433 per year and says delivery is due within 15 calendar days after contract signing. The original solicitation set an annual budget of RMB 650,000. The public award does not name the GPU model or disclose capacity, so this is a service-contract datapoint, not a comparable price per accelerator.
Follow @9cloudai for public updates on China’s compute supply.
Source: hkcgzx.hbust.edu.cn/jgggjzxc…
5
China Telecom’s Zhenjiang branch has opened a five-year centralized procurement for “intelligent computing integration services,” with an estimated ceiling of RMB 7.96 million before tax.
The tender names one winning supplier for the full 100% share. The contract runs for five years or until the procurement ceiling is reached, whichever comes first; payment is based on service orders actually placed. The scope is described as intelligent-computing integration, but the public notice does not specify accelerator models or a guaranteed compute capacity. It also requires five years of free maintenance.
That makes this a local service-supply contract rather than a disclosed GPU capacity reservation: the eventual resource mix and realized spend will depend on orders under the framework. Bids are due Dec. 7.
Public tender notice: yobest.cn/bid/2842816.html
Follow @9cloudai for more public updates on China’s compute supply.
1
14
让中国 Token 供给,接上海外 AI 推理需求
寻找一位长期经营海外客户与合作的业务合伙人
AI 应用不断发展,推理需求也在持续增长。我们正在探索一项长期业务:把中国的 Token 供给与海外 AI 推理需求连接起来。
我们希望找到一位愿意长期投入、一起把业务做起来的合伙人,负责拓展和经营海外客户,了解他们真正需要什么,并与国内供给方协作,推动合作逐步落地。
你将一起做的事
- 拓展并经营海外客户,持续沟通他们的业务场景与推理需求。
- 梳理客户对模型、价格、延迟和稳定性的关注,以及具体合作条件。
- 与国内供给方协作,把客户需求和供给侧情况逐项对齐,探索可行的合作方式并推进落实。
我们期待这样的你
- 有海外学习、工作或创业经历,熟悉当地 AI 行业与商务沟通方式。
- 乐于与客户沟通,能把需求、顾虑和合作条件梳理清楚。
- 有大客户拓展或合作经验更好。
- 认同长期经营,愿意和团队一起持续探索、推进并建设业务。
如果你也感兴趣
我们期待的不是一次性的客户引荐,而是一位愿意长期参与海外客户拓展与业务经营、共同把方向做深做实的伙伴。
如果你对这个方向有兴趣,欢迎聊聊。我们可以从海外客户需求、模型与供给的匹配,以及如何把这项业务长期做起来开始交流。
On Nov. 13, GPU Computing Radar’s fixed snapshot listed 11 H100 80GB on-demand offers, with the lowest posted rate at RMB 8.28 per GPU-hour. Its separate domestic-accelerator series showed a lowest listed rate of RMB 1.96 across six quotes.
The all-card low-price table was led by RTX 3090 listings: RMB 0.80/hour from Chenjian Cloud, RMB 0.89 from Hengyuan Cloud and RMB 1.10 from Zhixing Cloud. These are rates collected from providers’ public pricing pages, not negotiated enterprise quotes or guaranteed available capacity. The comparison site says stale, suspect and manually seeded entries are excluded from its index; the dated snapshot also separates a lowest listing from the fixed-basket index.
Snapshot: rentgpu.link/ghpi-daily/2026…
Methodology: rentgpu.link/sources
Follow @9cloudai for more public updates on China’s compute supply.
24
Dongfang Xinneng has disclosed a GPU-product purchase contract worth about RMB 912.5 million, including tax, as its first investment in a planned compute-services business.
The board-approved agreement was signed by the company’s wholly owned management partnership with a distributor identified only as “Company A.” The company says the GPU products are intended for AI inference, model fine-tuning, rendering, engineering simulation and scientific computing—not large-model training. Delivery may be phased, with the final shipment due no later than 12 months after signing. The final amount will depend on actual purchases and settlement.
The same filing says Dongfang Xinneng has no compute-service revenue or confirmed customer orders yet; customer development, data-center and power arrangements, and operations are still at an early stage. It also flags supply, financing and delivery risks. The supplier’s identity is withheld under a disclosed confidentiality exemption, so the public filing does not establish a specific GPU model or available capacity.
Source: company filing on CNINFO: static.cninfo.com.cn/finalpa…
Follow @9cloudai for more public updates on China’s compute supply.
20
Southern Power Grid has opened a group-wide tender for AI training and inference servers, with demand declared by its headquarters and operating units across five provinces.
The Nov. 10 notice covers high-end inference servers, training servers and related components. Demand comes from units in Guangdong, Guangxi, Yunnan, Guizhou and Hainan, including Shenzhen Power Supply Bureau, Southern Grid Energy Storage and other group entities. The listed projects include the grid-level intelligent-computing resource pool and high-performance compute clusters.
For the A-class projects, delivery is due within 30 days of contract signing, to the Southern Energy Data Center in Guian New Area, Guizhou. Equipment carries a 60-month warranty. The notice does not disclose a total award value; tender documents set package-level limits.
Official tender notice: bidding.csg.cn/zbgg/12004493…
Follow @9cloudai for more public updates on China’s compute supply.
25
China’s CAICT has started work on a test standard for inference acceleration across heterogeneous compute clusters.
An ODCC notice published Nov. 9 says the proposed standard will cover technical requirements, performance tests, efficiency evaluation and acceptance. The project team has been defining measures for concurrent inference requests, effective throughput, chip utilization and coordination across compute, storage and networking. The notice says results are currently difficult to compare across different hardware mixes and vendor implementations.
An online workshop on the draft is scheduled for Nov. 11 at 2:30 p.m. Beijing time. ODCC’s notice invites participating organizations and experts to register by Nov. 10 at 6 p.m. Beijing time.
For inference operators combining different accelerators, this work is aimed at making performance claims and acceptance criteria more comparable. The standard is still being drafted; this is a request for industry input, not an adopted specification.
ODCC notice: odcc.org.cn/news/p-211962134…
Follow @9cloudai for more public updates on China’s compute and inference supply.
13
ZTE’s new AI Delivery Network (AIDN) is designed to bring lightweight inference and caching closer to users while leaving complex inference at central nodes.
The company introduced the solution during the Nov. 5–7 Mobile Network Congress, combining operators’ existing CDN infrastructure with dynamic acceleration, SRv6, edge inference and intelligent caching. In ZTE’s published demonstration figures, network round-trip latency fell from 100 ms to 20 ms. For frequent Q&A, it reported first-token delay falling from 1.2 seconds to 200 ms after an edge-cache hit. ZTE also reported a semantic/KV-cache hit rate increase from 80% to 95%; these are company-reported demo results, not independent benchmarks.
The proposed split is explicit: edge nodes handle access, security checks, lightweight inference and caching, while central nodes take complex inference requests.
ZTE announcement: zte.com.cn/china/about/news/…
Follow @9cloudai for more public updates on China’s compute and inference supply.
21
A public tender in Shanghai puts a specific shape and budget on one buyer’s GPU requirement: one NVIDIA H200 server, leased as an exclusive cloud service for a year.
Haitong UniTrust International Leasing’s Nov. 6 tender specifies a disaggregated compute-and-storage arrangement, including networking and supporting storage. The server is to be delivered within 60 calendar days after contract signing. The stated project budget is RMB 2.4 million; that is a tender budget, not a winning bid or a market rate. The buyer would also receive priority to renew after the initial year.
This is a single-server procurement, not a large campus order, but the tender is unusually clear about the commercial wrapper: dedicated access, one-year term, and compute/storage separated in the cloud service.
Tender notice: shbid.com/html/jidian/gongga…
Follow @9cloudai for more public updates on China’s compute supply.
38
Southern Power Grid says data-center electricity can be planned at full scale and energized in stages as racks come online.
In a 10-measure package announced on Nov. 6, the utility moves grid-capacity assessments earlier, into industrial-park planning and project approval. Its proposed “one plan, phased connections; phased acceptance, phased energization” approach links power delivery to a campus’s commissioning sequence, rather than energizing all planned capacity upfront.
The package also calls for parallel approvals through a dedicated service channel, one-to-one service teams, and cross-region applications using a shared document set and authorization. Southern Power Grid says it will add an online compute-trading entry point with price forecasting and supply-demand matching.
Reported by China News Service: gd.chinanews.com.cn/2026/202…
Follow @9cloudai for more public updates on China’s compute supply.
14
Tian Yi Ma has unveiled its international token factory in Shantou, Guangdong, with its own model-aggregation platform at the center of the service.
The ceremony took place on November 1 at the Shenshan Digital Science and Technology Industrial Park, according to the company's November 3 announcement. A cross-border data-annotation base opened alongside it.
The company says construction of its first token factory began in May. It describes a 2,000P high-end commercial compute-cluster deployment and identifies its T.Y.M. Model Hub as the platform connecting token production, scheduling, delivery and payment settlement for international customers.
A second token factory is under construction, according to the same announcement. The accompanying annotation facility is designed to process text, conversations, image-text pairs, audio and video datasets.
Source: Tian Yi Ma, published November 3:
tym.com.cn/news/index2117445…
Follow @9cloudai for more public-source updates on China's compute and inference supply.
23
A new token-service platform in Fujian connects compute clusters across four provinces.
Mindu Ruici went live in Minqing county, Fuzhou, on November 3. The local government says it brings together 14 models, including DeepSeek and Kimi, with support for multiple domestic chip types.
Its underlying Minjiang heterogeneous-compute scheduling platform links Pingtan in Fujian, Suzhou in Anhui, Zhongwei in Ningxia and Xining in Qinghai. The network provides unified monitoring, operations and workload scheduling across the four clusters.
The service combines compute scheduling, model access, application deployment and token settlement. Billing is based on usage, with online settlement. Minqing also plans RMB 10 million in compute vouchers this year, alongside credit products tied to compute consumption records.
The announced use cases include ceramic visual inspection and construction-component testing, with enterprises renting capacity rather than building their own GPU facilities.
Source: Fuzhou government / Fuzhou Daily, November 3:
fuzhou.gov.cn/zwgk/gzdt/rcyw…
Follow @9cloudai for more public-source updates on China's compute and inference supply.
1
1
36
Envision's 800 V high-voltage DC power system is in commercial operation at its Xinghe site in Ulanqab, Inner Mongolia, according to a November 2 announcement reported by Beijing Business Today.
The integrated system combines a Tianshu controller, solid-state transformers, DC energy storage and DC/DC conversion. Envision says it connects renewable-power production through to the data center's computing load, coordinating renewable energy, storage and grid dispatch.
The Xinghe site began operating in August 2026. Its planned capacity is 3 GW, with a planned direct green-power connection share above 80%, according to the report. Those figures describe the development plan; the November announcement concerns the power system's commercial operation.
Source: Beijing Business Today, via Eastmoney:
emwap.eastmoney.com/a/202611…
Follow @9cloudai for more public-source updates on China's compute infrastructure and supply.
1
33
Baosight has proposed a 128 MW IT-capacity development in Ma'anshan, Anhui, with investment of approximately RMB 2.501 billion.
The October 28 public announcement describes the Baozhiyun East China intelligent-computing center's Ma'anshan base in Yushan district. Plans include two data-center buildings, A4 and A5, a 220 kV substation, backup-generator facilities and supporting electrical, cooling, water and monitoring systems.
The design covers about 108,658.5 square meters of floor area and approximately 5,328 racks. Baosight's controlled Anhui subsidiary is responsible for the project.
The financing plan allocates 25% to the subsidiary's own funds and 75% to financing it will arrange. The board approved the project on October 26; the announcement says shareholder approval is still required.
Baosight describes the development as part of its move from traditional IDC services toward AI computing infrastructure and services, targeting strategic-customer demand in the Yangtze River Delta.
Source: company announcement reproduced in Shanghai Securities News:
paper.cnstock.com/html/2026-…
Follow @9cloudai for more updates on China's compute supply.
1
28
A liquid-cooled edge-compute project in Changde, Hunan, is being planned around NVIDIA 5090 cards.
Wuling district and Guangzhou Zhisuan Information Technology signed the investment-cooperation agreement on October 30. An October 31 report on the Hunan provincial government's website describes the planned development as a 300P-class liquid-cooled micro-computing center.
The project also includes a 700-square-meter AI exhibition and demonstration center. Its stated application focus includes enterprise digital transformation and local AI services, with biopharmaceuticals, manufacturing equipment and data annotation among the industries mentioned.
The local plan connects the development with the Xiangbei public computing pool, Dongxing Data Smart City and Wuling's existing data-services ecosystem. The announcement records a signed development agreement and planned infrastructure; it does not give a commissioning date.
Source: Hunan government portal, reporting Hunan Daily:
hunan.gov.cn/hnszf/hnyw/szdt…
Follow @9cloudai for more updates on China's compute projects and inference supply.
1
42
A new public tender in Beijing calls for at least 20 PFLOPS of FP16 compute and 100 TB of high-performance storage for AI training and inference.
Published on October 30 by China Coal Technology & Engineering Group's procurement platform, project GHY-26262 is being procured by the Coal Science and Technology Research Institute. It covers a computing platform for major mining-disaster research, including AI compute, general-purpose compute, storage, networking and resource-management and scheduling software.
The notice specifies delivery within 30 days of contract signing, followed by installation, commissioning and acceptance within 90 days of equipment arrival. Delivery is to the institute's Caiyu campus in Beijing's Daxing district.
It also requires onsite system-integration and interoperability testing as part of acceptance. The stated funding is fully secured special-purpose funding. This is an open equipment procurement, with the supplier still to be selected.
Source: official public tender:
cg.ccteg.cn/cms/channel/ywgg…
Follow @9cloudai for more updates on China's compute projects and supply.
1
2
61
Sinnet has two Inner Mongolia compute projects under construction, with some rack capacity potentially coming online as early as the end of 2027.
The operator's Q3 report, dated October 28, says its Horinger intelligent-computing center and Hohhot compute-base projects have both started construction and obtained energy-consumption quotas. Construction will advance in line with customer requirements.
There is also a separate project in Hengqin, Zhuhai. A subsidiary established in September has acquired ownership of a 10,093-square-meter data-center building and is applying for energy-consumption approval. The planned customer focus is local enterprises and government users, alongside efforts to obtain cross-border connectivity and relevant pilot qualifications.
Across its portfolio, Sinnet reports about 92,000 operational racks and more than 230,000 planned racks, using a 4.4 kW-per-rack basis. The Inner Mongolia timetable concerns a portion of future rack capacity; the Hengqin project is still working through approvals.
Source: Sinnet Q3 report, publicly reproduced by Sina Finance:
vip.stock.finance.sina.com.c…
Follow @9cloudai for more updates on China's compute supply.
16
Four compute-campus projects have started construction in Horinger New Area, Inner Mongolia, according to an October 26 report sourced to the local administration.
The projects involve GDS, Yungang Intelligent, Zhongjin Data and Sinnet, with combined planned investment of about RMB 15.2 billion.
Yungang’s full campus plan includes five data-center buildings and around 200MW of IT capacity. The portion now starting construction comprises two data-center buildings with 88MW of IT cabinet capacity, using liquid cooling as a core efficiency measure.
GDS’s northern campus calls for three data-center buildings across roughly 184,000 square meters, targeting large-model training, AI inference and cloud computing.
Zhongjin’s plan includes four machine-room buildings, eight auxiliary power buildings and a 220kV substation; piling is underway. Sinnet’s RMB 1.235 billion project includes two data centers, two diesel-generator buildings and space reserved for a phase-two 110kV substation.
Source: Horinger New Area, republished by People’s Daily’s Jintai News account on October 26:
news.qq.com/rain/a/20261026A…
Follow @9cloudai for more updates on China’s compute supply.
1
24
Qingyang’s data-center buildout: CCTV reports a modular machine-room building completed in about five months, with facility handover and equipment installation happening almost in parallel.
The October 27 report says this construction method cut roughly a year from the conventional schedule. Qingyang’s operating cluster now has more than 110,000 standard racks, while additional sites remain under construction.
Power and cooling are central to the location. CCTV reports a local data-center electricity rate of RMB 0.398/kWh, 30–50% below major eastern compute hubs. The city’s average annual temperature is around 9°C, allowing natural cooling for much of the year; the report puts cooling costs about 30% below eastern regions.
Facilities also use magnetic-levitation air cooling and liquid cooling. More than 780 digital-economy companies have established a presence in Qingyang, according to the report.
Source: CCTV News, October 27, via its public Tencent account:
news.qq.com/rain/a/20261027A…
Follow @9cloudai for more updates on China’s compute supply.
1
1
22
A pricing update from China Telecom's Tianyi Cloud: annual GPU discounts are being removed for selected A10 and A100 instances, with the effective date pushed back to November 4.
The October 26 notice moves the change from its original October 28 start date. It covers pi7 and g7 instances using NVIDIA A10 GPUs, plus p8a instances using A100 GPUs, in the provider's Category I nodes. International resource pools are excluded.
For orders from November 4 onward, the new pricing applies to new purchases, conversions to annual billing, renewals and configuration changes. Orders created before November 4 retain the existing pricing for the specified operations during their contract term, including renewals and configuration changes within the same instance family.
There's a separate storage change: the standard pay-as-you-go rate for the HPFS HPC Performance 400 MB/s/TB tier will rise 30%, applying to all file systems in that tier once the change takes effect. Customers with special contractual terms follow their agreements.
Source: Tianyi Cloud's public notice, October 26:
ctyun.cn/notice/detail/11100…
Follow @9cloudai for more updates on China's compute supply.
2
53
