MornAI (Chenjianyun): renting A100 and RTX 4090 GPU capacity by the hour

MornAI is a GPU rental platform positioned as ready-to-use access to high-end graphics cards. It lists A100, RTX 4090 and RTX 3090 class GPUs, offers three delivery forms - cloud host, cloud container and bare-metal machine - and bills by the hour, day, week or month, with discounts on weekly and monthly terms. For teams running deep learning, large-model training or scientific computing, it solves a familiar problem: you no longer have to buy cards, rack them, find a data centre and staff the operations before the first experiment runs.


The hidden cost of building your own GPU server is easy to underestimate. Procurement takes time, a chosen model is hard to change afterwards, idle machines keep depreciating, power and cooling are fixed monthly items, and driver, CUDA and framework versions consume engineer hours before any model trains. Renting compute does not automatically lower the unit price; its value is turning a capital asset into an expense line you can switch on and off, which matters most when project demand fluctuates. MornAI's product line follows that logic: cloud containers for short jobs, cloud hosts when you need full control of the operating system, and bare-metal machines when exclusive performance is the requirement.

On price, the vendor describes GPU rates as 10 to 30 percent of what traditional public clouds charge, and mentions a programme that lets idle resources earn money. Those are vendor claims, and real cost depends heavily on the card model, rental duration and current stock, so it is worth calculating against your own workload hours rather than applying the discount figure directly.

Core capabilities

GPU capacity on demand

A100, RTX 4090 and RTX 3090 class cards cover deep learning training, large-model fine-tuning, scientific computing and rendering workloads, with GPU count configurable alongside CPU, memory and disk. The source material lists card families but does not give memory size, interconnect type or per-card throughput, so confirm exact specifications on the compute marketplace page before ordering.

Preconfigured environments

System images ship with deep learning, large language model and scientific computing stacks, so a new instance does not start from a blank CUDA install. For researchers and small teams without dedicated operations staff, the hours saved here often matter more than the money saved.

Three delivery forms

Cloud hosts suit workloads that need full operating system control, cloud containers suit short, reproducible jobs, and bare-metal machines target training and rendering that demand exclusivity and stable performance. Each carries a different operations burden and cost structure, and a project can move between them as it matures.

Elastic billing and scheduling

Billing granularity runs from hourly to daily, weekly and monthly, with discounts on the longer terms, and the platform describes second-level scheduling so resources can be adjusted as demand changes. For research workloads that run hard for a stretch and then sit idle, hourly billing is usually easier to control than a monthly commitment.

Managed infrastructure and support

The platform handles underlying hardware operations and advertises 24/7 expert support, plus one-to-one guidance for academic work. Specific support channels, response times and the scope of that guidance are not detailed in the source material.

Getting started

1. Register and sign in at the official site, mornai.cn.
2. Open the compute marketplace and choose cloud host, cloud container or bare-metal machine.
3. Pick a system image, for example one preloaded with deep learning, large language model or scientific computing tooling.
4. Configure hardware: select GPU model and count, then set CPU, memory and disk capacity.
5. Choose a billing term - hourly, daily, weekly or monthly - with discounts on weekly and monthly.
6. Create the instance, review it in the console and log in with one click to start working.

Pricing and billing

• Billing granularity covers hourly, daily, weekly and monthly terms, with larger discounts on longer rentals, which lets a project budget follow its own cadence rather than the calendar month.
• The vendor describes GPU pricing as 10 to 30 percent of traditional public cloud rates, a relative claim that does not name the comparison providers or matching card models.
• The source material does not disclose per-card unit prices, minimum rental periods, deposits or refund rules; actual quotes need to be checked on the compute marketplace page.
• An idle-resource monetisation programme is mentioned, but eligibility, settlement mechanics and supported card models are not described in the source.

Who it fits

• AI startups whose compute demand swings between training pushes and quiet product phases, and who want to avoid a one-off hardware purchase.
• Universities and research groups with sporadic, bursty demand and short concentrated experiments.
• Enterprise R&D teams running large-scale training or complex rendering who need elastic scheduling and lower idle cost.
• Individual developers and designers using 3090 or 4090 class cards for lighter inference and rendering, at a lower entry point than buying a workstation.

Limitations and things to check

• The source lists card families only, without memory size, per-card throughput, GPU interconnect such as NVLink, or network bandwidth, all of which must be verified in the console before committing to multi-GPU training.
• Availability of each delivery form depends on current stock, and popular cards may queue during peak demand, so project schedules should include buffer time.
• The 10 to 30 percent figure is a vendor comparison that does not name the reference clouds or card models, so cost planning should use your own workload hours and specifications.
• Data security, compliance certifications, data residency and backup policy are not covered in the source; for sensitive or regulated data, confirm with the vendor before uploading anything.
• Bare-metal rental involves delivery and exclusivity confirmation, and provisioning lead time plus exit process are not described, so short projects are better validated on cloud hosts or containers first.
• The advertised 24/7 support and one-to-one guidance come without stated response times or service boundaries, so critical projects should not depend entirely on external support for troubleshooting.

Frequently asked questions

How is this different from buying GPUs and building a server?

Buying cards is a capital purchase followed by ongoing costs for space, power, cooling and operations, and the chosen model is hard to change. MornAI rents GPU capacity by the hour, day, week or month across cloud hosts, containers and bare-metal machines, so you can stop renting when a project ends or demand drops. The difference is flexibility in the cost structure rather than a guaranteed lower unit price.

Which GPUs are available?

The source material lists A100, RTX 4090 and RTX 3090. These cover large-model training, deep learning and rendering workloads, but memory capacity and interconnect configuration should be confirmed on the instance configuration page of the compute marketplace.

Do I have to install the software stack myself?

System images come with deep learning, large language model and scientific computing environments preinstalled, so a new instance usually does not need a from-scratch CUDA and framework setup. If your project pins specific versions, check the version combination inside the image first.

How does billing work?

Billing runs hourly, daily, weekly or monthly with discounts on the longer terms, and the vendor describes GPU rates as 10 to 30 percent of traditional public cloud. Per-card prices, minimum rental periods and deposit rules are not disclosed in the source, so treat the live quote on the marketplace page as authoritative.