GPU servers for universities and research labs, on fixed budgets

GPU servers for universities and research labs work best when the cost is as predictable as the academic calendar. We rent dedicated NVIDIA and AMD GPUs by the month, at a flat price billed in advance, so a lab can match GPU spend to a grant line or a semester and stop paying when the work ends. Each server is assigned to your institution only, with root access, so you decide how students and researchers share it.

Budgeting GPU servers for universities and research labs

Monthly billing maps cleanly onto academic budgets. A 15-week semester spans four billing months, so one H100 for a course costs 4 × $2,190 = $8,760. A research group sharing an HGX H100 node for six months pays 6 × $16,900 = $101,400 before any term pricing, which is worth requesting for periods of 6–12 months or for several servers. For the mechanics, see how billing works.

• The price is fixed for each paid month, and any change for future months comes with at least 30 days' notice.
• There is no setup fee and no egress fee within the included bandwidth: 20 TB a month on single-GPU plans and 50 TB on 8-GPU servers.
• Prices are in USD and exclude VAT and other taxes, which are added where applicable.
• Plans renew automatically, with an email reminder before each renewal. Cancel before the renewal that follows the end of term and you stop paying; see renewals and cancellation.

GPU servers for teaching

For a course, size the GPU to the largest exercise, then decide how students share it:

• Intro deep learning with CNNs and small transformers runs well on an RTX 5090 with 32 GB or an L40S with 48 GB.
• LLM courses: a QLoRA exercise on a 7B–8B model needs roughly 10–20 GB including activations, so several students can share one 96 GB RTX PRO 6000 if jobs are scheduled. A 70B QLoRA project needs about 40 GB for the base model, which suits one H100 per team.
• Sharing: create Linux accounts, give each student a container, and assign GPUs with CUDA_VISIBLE_DEVICES or a scheduler. Jupyter and VS Code both work over SSH, as described in our remote development guide.

Whole GPUs are never sliced or shared with other customers, so the only contention is the one you manage among your own students.

Shared lab servers for research groups

An HGX H100 node gives a group eight 80 GB GPUs on NVLink, 2 TB of DDR5 and eight 3.84 TB NVMe drives. Installing Slurm on the node lets jobs queue fairly and request one to eight GPUs each. Groups training larger models can step up to an HGX H200 node with 1.13 TB of HBM3e. For simulations and FP64 workloads, our page on GPU servers for research and HPC compares the H100, H200 and MI355X.

You keep all rights to your data, and we access it only to operate or repair the service at your request, or when required by law. Local storage is not backed up unless agreed in writing, so keep copies on institutional storage, and check with your data office before uploading regulated or human-subject data. When a server is returned, its disks are securely erased.

An example course plan

Suppose a 40-student course runs weekly QLoRA labs on 8B models, then ends with team projects on a 70B model. At roughly 10–20 GB per job, one 96 GB RTX PRO 6000 holds about four to nine jobs at once, each running slower while they share the card, so two of them carry a lab section working in pairs with a simple queue. For the final month, ten teams of four each get one H100, where the 4-bit 70B base takes about 40 GB.

• Two RTX PRO 6000 servers for the four-month term: 2 × 4 × $1,390 = $11,120.
• Ten H100 servers for the final month only: 10 × $2,190 = $21,900, with volume pricing worth requesting at that quantity.

Cancel each server before its next renewal and the spend stops with the course. Because the price per month is known in advance, this plan can go into a budget request before the semester starts.

Procurement, invoices and payment

Ordering follows four steps: you send a request, we confirm the configuration, price, location and delivery date in writing, you pay through a payment link, and credentials arrive by email. The written confirmation lists those details before anything is paid, which helps when a purchase needs internal approval. If your department pays by card, Visa, Mastercard and American Express are accepted, and crypto is accepted at the same USD price. Every payment comes with the same invoice, whatever the method; see invoices and VAT. For large orders we may ask for verification, such as institution details and the identity of the person ordering, and export control and sanctions rules apply to every order.

When a monthly server is the right fit

If your university's HPC center offers allocations and its queue fits your deadlines, use it. A monthly server makes sense when queue waits block a paper or a thesis, when you need root access or a GPU your cluster lacks, or when a course needs dedicated GPUs on fixed dates. For jobs that run only a few hours a month, hourly rental elsewhere will cost less.

To get started, send your course or project needs through the contact form a few weeks before term. We confirm the delivery date in writing before you pay, and if we miss it you can cancel for a full refund.