GPU Hourly Rental
Elastic compute, billed by the hour
弹性算力，按需调用
Rent GPU capacity on demand, monthly, or annually, covering domestic and mainstream multi-generation GPUs. Scale smoothly from a single card for validation to a thousand-card training cluster — and never pay for idle compute.
↑85%
GPU 利用率提升
30%+
综合成本降低
<30s
故障自动切换
Core capabilities
What GPU 卡时租赁 solves
The problem with building your own cluster is buying for the peak and using it at the average. Hourly rental turns a fixed cost into a variable one, while cross-architecture scheduling pushes utilization from 30–50% to 85%.
多型号算力池
覆盖国产与主流多代际 GPU，按训练、推理、渲染等负载类型匹配最优卡型。
弹性扩缩容
按小时计费、随用随停，支持从单卡验证到千卡集群训练的平滑扩缩，避免算力闲置。
统一调度平台
异构芯片统一纳管与调度，同一套作业描述可在不同芯片架构上运行，避免厂商锁定。
高性能存储与网络
配套分布式存储与高速互联网络，保障大模型训练的数据吞吐与通信效率。
- 国产与主流多代际 GPU 多型号可选
- 跨芯片架构统一调度，GPU 利用率提升 85%
- 按小时计费，弹性扩缩不闲置
Instance types
Three billing models for three workload shapes
Deciding whether your workload is short-term, always-on, or peak-driven — and billing accordingly — is the single most effective cost lever.
On-demand instances
Billed hourly
- Best for
- Model validation, short experiments, bursty inference
- Notes
- Start and stop at will, delivered within minutes. Suited to teams still validating, who should not prepay for an uncertain timeline.
Reserved instances
Monthly or annual
- Best for
- Steady training jobs, always-on inference services
- Notes
- Reserved capacity with tiered discounts over longer terms. Suited to predictable production workloads with stability requirements.
Cluster instances
Quoted per project
- Best for
- Large-scale training, CAE / CFD simulation
- Notes
- Multi-node fabric with distributed storage and high-speed interconnect, scaled elastically for the project and released when it ends.
Specifications
What surrounds the compute matters just as much
For large-model training and simulation, the bottleneck is often storage throughput, interconnect bandwidth, and data isolation policy rather than the chip itself.
- Accelerators
- Domestic and mainstream multi-generation GPUs
- Minimum rental
- 1 hour
- Delivery
- Minutes for on-demand instances; agreed schedule for clusters
- Storage & network
- Distributed storage with high-speed interconnect
- Data isolation
- Isolated instance networks with independently mounted volumes
- Data erasure
- Wiped according to policy on release, with customer-managed keys supported
Need a specific card type or a mixed deployment? Tell us in the enquiry and we will provide an actionable compute list with a measured baseline.
How to work with us
Four steps from first contact to live
A standard commercial process. Technical material and integration documentation are provided on request once we start working together.
- 01
Confirm workload characteristics
Tell us the framework, model size, memory, and interconnect requirements, and we match the best card type and topology.
- 02
Provision compute
On-demand instances are delivered within minutes; clusters are networked and mounted on the agreed schedule.
- 03
Submit jobs
Mainstream training frameworks and container images are supported, and the same job description runs across chip architectures.
- 04
Scale and release
Scale with your business rhythm, release instances when a job finishes, and storage is wiped according to policy.
FAQ
What buyers ask most
最小起租粒度是多少？
能否指定芯片型号？
数据安全如何保障？
Explore the other product lines
The three business lines combine freely and share one account, one metering system, and one bill.
Data basisPerformance and cost figures on this page come from production statistics on the DengCloud platform and have been reviewed with our product and engineering teams.
Get a compute plan and quote
Tell us your workload characteristics and timeline, and our team will recommend card types with a cost estimate.
Explore platform capabilitiesBusiness response, Mon–Fri 9:00–18:00 (CST)