Buy or Rent GPUs? A Cost Comparison for Companies and Universities
Ā·4 min read
A lab needs GPUs for a new research project. An engineering team wants to deploy a model. How much should either set aside? A September purchase by Hubei University of Technology put a 5090 workstation at CNY 63,500. For a few rounds of experiments, that is a substantial upfront commitment.
Renting avoids a large upfront purchase, while buying provides a machine for ongoing work. Start with purchase quotes and rental budgets for the work your team expects to run.
What are universities paying for GPUs?
| Record | Published scope | Unit award price |
|---|---|---|
| Fudan Ā· May 6, 2026 | 5090 32GB card; 15 purchased, installation and five-year warranty requirements | Ā„29,800/card Not a complete computer |
| Hubei University of Technology Ā· September 20, 2026 | One 5090 32GB, 64GB RAM, Ultra 7 270K PLUS workstation | Ā„63,500/system |
| SUSTech Ā· project 2026-00191 | Two RTX PRO 5000 Blackwell 72GB GPUs, dual AMD 9554, 128GB ECC RDIMM, software and services | Ā„224,000/system |
| Shenzhen University Ā· July 13, 2026 | FI82F54F-Z server; GPU and RAM details not itemized in the result notice | Ā„713,990/system |
Verified October 2, 2026. These are historical project awards, not current stock quotes. Delivery, taxes, software and warranty scope depend on the announcements and contracts. The Shenzhen award illustrates purchase scale only and is not used in a matched rental calculation.
Memory also needs fresh pricing. TrendForce's September 30 outlook forecasts another 10ā15% quarterly increase in conventional DRAM contract prices for Q4. That is not a retail DIMM price forecast. Obtain separate quotes for high-capacity ECC RDIMMs rather than using desktop memory budgets.
How much does a year of 5090 rental cost?
Compshare's advertised 5090 rate was CNY 3.32/hour when checked on October 2. Applying that headline rate gives:
| Hours/year | One-year base rental | Three-year base rental |
|---|---|---|
| 500 | Ā„1,660 | Ā„4,980 |
| 2,000 | Ā„6,640 | Ā„19,920 |
| 6,000 | Ā„19,920 | Ā„59,760 |
Source: Compshare, October 2, 2026. Verify additional CPU, RAM, storage, network, paid-image charges and current order terms. Obtain monthly or annual quotes for continuous use. Future rental rates are not forecast here.
For a few hundred hours of experiments each year, rental is easier to budget: 500 hours costs CNY 1,660 in base rental at this rate, without buying a complete machine.
A shared lab running 6,000 hours a year should obtain a purchase quote too: three-year base rental approaches CNY 60,000 at the listed rate. Include power and maintenance on the buying side, and obtain monthly or annual rental quotes for sustained use.
Teams that already own GPUs can keep routine work local and rent extra capacity for deadlines or training campaigns, rather than buying enough hardware for their busiest week.
Why consider Compshare?
Compshare offers usage-based rental and prepared images, making it an option for teams renting by project rather than buying and setting up their own server.
Its service framework agreement includes bank transfers and invoice applications. Have finance approve the contract and invoice requirements before starting the work, not after the bill arrives.
Compshare belongs to UCloud, a STAR Market-listed company (688158), with publicly available corporate filings. Its 2026 interim report identifies the founding team as its controllers; it is not state-controlled.
Protecting company and research data
Restrict access to login and management interfaces when setting up the machine. Compshare's firewall documentation describes port and source-IP controls, with different mechanisms by region.
Data retention also deserves attention. Under the current reclamation rules, usage-based instances and system disks can be reclaimed after seven consecutive days shut down; unpaid resources face reclamation too. Shutdown is not indefinite free storage. Back up important results and check storage charges.
If policy requires on-premises data, specific certifications, isolation or audit evidence, obtain written confirmation for the exact service. An ordinary cloud GPU order does not establish compliance by itself.
Rental spreads spending across intermittent projects. Teams with year-round work can request both purchase and long-term rental quotes. ComputeUnion's GPU rental price comparison provides a starting point for preparing that budget.
Sources Ā· checked October 2, 2026
- Hubei University of Technology: 5090 workstation award, September 20, 2026
- Fudan University: 5090 accelerator purchase, May 6, 2026
- SUSTech: GPU server award, SUSTech-JC-2026-00191
- Shenzhen University: GPU server award, July 13, 2026
- TrendForce: Q4 2026 memory contract-price outlook
- Compshare: public rental prices and product overview
- UCloud: corporate profile
- UCloud 2026 interim report: controlling shareholders
- Compshare: firewall and supported regions
- Compshare: resource reclamation rules
- Compshare: framework agreement, payments and invoices
Purchase prices are historical awards for specified projects, not uniform market quotes on October 2, 2026. The SUSTech announcement does not display a publication date, so its project ID is used. Rental figures are headline prices, not matched full-order quotes. No equipment or paid instance was purchased.