AI compute · GPU rental
GPU compute rental
Rent whole servers or time-shared GPUs, matched to the job: H-series for training, professional cards for inference, consumer cards for experiments. GPUs, power, networking and remote management go on one work order. No need to buy a room full of machines first and then look for somewhere to plug them in.
GPUs available now
These are the series we can deliver. Configuration, lead time and cabinet power are confirmed per work order. Price on request. Please don't back-calculate a whole-server cost from public cloud hourly rates.
Nvidia H200 / H100 / H20
For training clusters with long context, multi-node parallelism and GPUs running at full power. Delivered in Guiyang by default, with InfiniBand and high-power cabinets.
- H200 · 141GB HBM3e, long-context inference and large-scale training
- H100 · 80GB HBM3, Transformer Engine
- H20: ask us; subject to stock and export rules
- China Telecom bandwidth from 50M · 1 public IP
PRO 6000 / 6000D
Single-GPU LLM inference and LoRA fine-tuning. If the memory is enough, you don't need to start with H-series GPUs.
- 96GB / 84GB GDDR7 ECC professional cards
- For inference, fine-tuning and local models behind TokenHub
- Can run AI Inference Software on the same card
- For training peaks, we suggest queuing on H-series
GeForce 5090 / 4090
Cost-effective inference, benchmarking, development and testing. They can be combined in multi-GPU cabinets, but are not our default for production training.
- For demos, comparison experiments and short fine-tuning runs
- Multi-GPU builds that suit high-power cabinets
- Remaining 3080 Ti / 3090 Ti stock available on request
- ECC, drivers and licensing spelled out in the work order
Nvidia A40 / A16
Professional graphics, cloud desktops and mid-sized inference. Suited to splitting into multiple instances, not to large-scale pretraining.
- A40 for professional graphics and mid-sized inference
- A16 for virtualized partitioning and multi-user desktops
- Can be placed in Guangzhou or Beijing, close to your users
AMD MI325X / MI300X
A second route for large-memory training. Can be evaluated alongside Nvidia clusters in a mixed deployment. Drivers and frameworks need separate acceptance testing.
- For memory-sensitive teams that can work with the ROCm stack
- Lead time, interconnect and mixed-deployment plan confirmed per order
- Not assumed to be interchangeable with CUDA workloads
AI Compute Appliance / DGX Spark
Validate in your office or at the edge first, then move to high-power cabinets in Guiyang. Data and trial units can stay on your premises for now.
- For confidential material, low concurrency and getting the workflow running first
- Add GPUs after validation instead of renting a full cabinet on day one
- AI Inference Software preinstalled; connects to TokenHub
Three ways to rent
Start small to validate, then choose time-sharing, dedicated GPUs or full cabinets based on results.
Time-shared rental
Use GPUs by time slot. Good for benchmarking, demos, short LoRA runs and comparing models. Measure quality, latency and cost first, then decide on a long-term lease. The billing period is set in the work order.
Dedicated GPU / monthly server rental
An exclusive whole server or dedicated GPUs, for steady inference or shared development machines. Environments and data are kept long term. Leases start at one month and can be renewed, as specified in the contract.
Full-cabinet training clusters
H-series multi-node, multi-GPU servers + InfiniBand + 15 / 20 / 40kW high-power cabinets in Guiyang. Contracts are based on cluster size and project duration. For pretraining and large-scale fine-tuning.
Have your own GPU servers and only need power and rack space? See Cabinet Rental or Server Colocation. Data can't leave your office? See on-premises deployment with the AI Compute Appliance.
Start from the workload, not the GPU model
When you ask for a quote, please tell us: training or inference, model size, whether you need multiple nodes, estimated power draw, whether data can leave your premises, and where your users are. Without these, no “lowest-price H100” will match what is actually deployed.
- Large-scale training, long context, multi-node parallelism
- Inference, LoRA and benchmarking that need predictable memory and latency
- Data must stay in the data center, or training and user access are split across two locations
- Your current GPUs are underfed and need high-power cabinets, not 2–5kW low-density racks
Pick GPUs by workload, not by maxing out the spec sheet
| Workload | Typical GPUs | Power / interconnect | Preferred site | Rental model |
|---|---|---|---|---|
| Large-scale training / long context | H200 / H100 / H20 | High-power cabinet + IB | Guiyang | Whole server; time-sharing negotiable |
| Single-GPU inference / LoRA | PRO 6000 / 6000D | Assessed per GPU power draw | Guiyang, or a nearby inference site | Whole server or dedicated GPU |
| Benchmarks, demos, short experiments | 5090 / 4090 | Multi-GPU cabinets possible | Guiyang | Usually time-shared |
| Graphics, cloud desktops, multi-instance | A40 / A16 | Colocation-grade power is enough | Guangzhou / Beijing | Whole server |
| Large-memory alternative / mixed-deployment evaluation | MI325X / MI300X | Assessed per cluster | Guiyang | Ask for lead time |
| Validate in the office first | AI Compute Appliance / DGX Spark | Office power | On premises, then move to Guiyang | Purchase or trial unit |
A selection guide, not a delivery commitment to every customer. Lead time, interconnect and time-sharing availability depend on stock.
Cabinets, network and remote access, delivered together
The GPU is one line on the list. At full load, power and cooling hit first; across servers, interconnect; when things fail, whether remote hands can fix it.
High-power cabinets
Guiyang offers 15 / 20 / 40kW high-power cabinets, limited by power rather than rack units. Traditional 2–5kW cabinets make multi-GPU servers throttle, or eat into the cooling of neighboring cabinets.
See Guiyang data centerInfiniBand and 100G
IB switching and 100G links are configured separately for training or inference. Gradient sync should not run over patched-together public internet links.
Bare-metal VNC remote access
Reinstall, reboot and rescue remotely. No site visit needed unless it's a hardware fault.
Ready-to-use environment
Free one-time help installing the OS and common training and inference environments. During hardware faults, we can arrange a standby server.
Training needs power; serving needs network
Our sites in mainland China split the work. Guiyang runs GPUs at full power; Guangzhou and Beijing handle user access and inference APIs; Dongguan joins when public entry points need protection. GPUs go where the workload is, not “GPUs in every city.”
Guiyang Data Center
H-series, multi-GPU clusters, full-power training and inference
View site Inference / accessGuangzhou Data Center
Inference APIs, users in South China, graphics GPUs
View site Proximity / complianceBeijing Data Center
Government and enterprise customers in North China, inference with data residency
View site Egress / protectionDongguan Data Center
Added to the plan when public entry points need DDoS-protected egress
View siteSuggested way to start
- 01
Define the workload
Training, inference or benchmarking; model size; single or multi-node; whether data can leave your premises; which city your users are in.
- 02
Validate small
Start with time-shared GPUs, an appliance or TokenHub credits. Check latency, stability and cost.
- 03
Scale the power
Once results are stable, move to high-power cabinets or add dedicated GPUs. Separate training from inference, and put a firewall in front of public entry points.
- 04
Long-term lease and hybrid
Sign for whole servers when you need sustained usage. Existing public cloud accounts can be connected to our data center GPUs on one network.
FAQ
Can I rent just GPUs, without a cabinet?
Yes. Ask for a quote for whole servers or time-shared use. H200 / H100 / H20, professional cards and 4090 clusters are mainly delivered in Guiyang, and power and networking are still covered in the same work order. If you have your own equipment to rack, see Cabinet Rental or Server Colocation.
Can I put an H100 training cluster in Guangzhou, Dongguan or Beijing?
Standard cabinets are negotiable, but power is assessed on colocation terms, not designed for training. We still recommend Guiyang for fully loaded multi-GPU H100 / H200 clusters. When inference moves back to city sites, request power based on actual draw.
Why is all pricing on request?
GPU model, lead time, time-sharing, IB / 100G, public bandwidth and power form one package, not a shelf price. We don't publish fixed unit prices, so the price of one batch of stock isn't taken as a promise to every customer. Please include your workload and estimated power draw when you ask for a quote.
Time-shared or whole server: how do I choose?
For benchmarks, demos and short LoRA runs, start with time-sharing. For steady inference, team use or large-scale training, sign for a whole server or dedicated GPUs. The appliance is good for validating in your office first.
Who installs the system environment? What if something fails?
We help install the OS and common training and inference environments once, free of charge. Remote access is via bare-metal VNC for reinstall, reboot and rescue. Non-hardware issues don't need a site visit. During hardware faults, we can arrange a standby server.
Can 4090 / 5090 GPUs be used for production training?
They work for experiment clusters and cost-effective inference, but are not our default for large-scale pretraining. ECC, drivers, licensing and cooling need to be confirmed separately. For long context and multi-node parallelism, we still recommend H-series + IB.
How is GPU compute rental different from the AI Compute Appliance and AI Inference Software?
GPU compute rental means using GPUs in our data centers, by whole server or time-shared. The AI Compute Appliance is a machine you buy and place in your own server room or office. AI Inference Software and TokenHub are the software layer that runs on these GPUs, and can be added to either option.
Related
Send us your workload and GPU needs
Tell us: training or inference, model size, how many GPUs, estimated power draw, and whether you need IB. We'll propose a plan based on stock and sites, rather than quoting an “all-in-one package” up front.
- leo@hs.com.cn
- 24/7 technical support