Home Fleet Models Rental FDE On-prem About

GL Compute

Efficient scheduling · Stable operation · Intelligent acceleration

AI video from $0.01/s · 2K at $0.028/s · 18,848 GPUs we own
01Our fleet

The campus at a glance

2,3568-GPU servers
18,848GPUs
1 PBof GPU memory (about 1,006 TB)

8 × NVIDIA RTX 5090

Video and image generation, mid-size models

2,000servers
16,000GPU
  • GPU8 × RTX 5090 32 GB
  • CPU2 × Intel Xeon Gold 6530 (32C/64T)
  • RAM512 GB DDR5
  • Disk960 GB + 3.84 TB SSD
  • Network2 × 25 Gb SFP28

8 × AMD MI308X

Large language model inference

228servers
1,824GPU
  • GPU8 × MI308X 192 GB
  • CPU2 × AMD EPYC 9654 (96C)
  • RAM512 GB DDR5-5600
  • Disk2 × 480 GB + 2 × 7.68 TB SSD
  • Network4 × 400 G + 2 × 200 G + 2 × 25 Gb

8 × NVIDIA H200

Training, fine-tuning, day-one model launches

128servers
1,024GPU
  • GPU8 × H200 141 GB
  • VRAM1.1 TB per server
  • LinkNVLink
  • UseWhole-server long-term rental
  • ConfigBuilt to order
02Models and prices

12 leading open models (custom model installs supported)Tokens at 50% of market price

Chinese flagships plus the open models the world uses, behind one API key. OpenAI-compatible.

Struck-through prices are public market reference prices, November 2026. Text models are priced per 1M tokens (input / output). MiniMax H3 video is not currently offered in the US, EU, UK or South Korea.

03GPU rental

Whole servers by the month50% of market price

Delivered as a whole machine with remote access. Run whatever you like.

8 × RTX 5090

Video and image generation, mid-size model inference

  • GPU8 × 32 GB
  • CPU2 × Xeon Gold 6530
  • RAM512 GB DDR5
  • Disk960 GB + 3.84 TB SSD
  • LAN2 × 25 Gb
  • UplinkSized to order
$2,800$1,400per server / monthAsk to rent

8 × AMD MI308X

1.5 TB of GPU memory per server, for very large language models

  • GPU8 × 192 GB
  • CPU2 × EPYC 9654
  • RAM512 GB DDR5-5600
  • Disk2 × 480 GB + 2 × 7.68 TB
  • LAN4 × 400 G + 2 × 200 G
  • UplinkSized to order
$16,000$8,000per server / monthAsk to rent

8 × NVIDIA H200

Training and fine-tuning, the most widely accepted hardware

  • GPU8 × 141 GB
  • LinkNVLink
  • ConfigBuilt to order
  • TermFrom 1 month
  • LANBuilt to order
  • UplinkSized to order
$21,200$10,600per server / monthAsk to rent

Struck-through prices are reference monthly rates for comparable servers abroad, November 2026. Long terms get a further discount.

04FDE service

Hand the repetitive work to agents

Our engineers work inside your company, map your workflows, deploy AI on your own machines and wire it into the systems you already run. Many agents on many machines, working together around the clock.

10%of the headcount
200%of the output (target)
24hnon-stop operation
Workflow mapping On-premise AI ERP CRM Finance Customer service Auto editing Image generation Video generation
05On-prem AI servers

Installed, tuned and ready when it arrives

Your data never leaves the building. We install and test the system, drivers, inference stack, knowledge base and workflow tools before shipping. Plug it in and start chatting, coding, and generating images and video.

OS, CUDA and drivers pre-installed Inference stack and web UI Knowledge base and workflow tools Works with mainstream coding agents Graphical remote management (BMC) Tower or 4U rack chassis

Reference prices before tax, including installation and testing, excluding shipping. USD figures are approximate at 7.1 CNY per USD.

06About us

GL Compute · Ulanqab campus

Inside the data hall

Our data halls are in the Ulanqab compute cluster in Inner Mongolia. The average temperature is 4.3 °C, so outside air cools the servers for nearly ten months a year, and the region has abundant, low-cost wind and solar power. Those two facts are why we can charge half the market price.

Two dedicated fibre routes run to Beijing with a round trip of about 4.2 ms. Beijing is about 90 minutes away by high-speed rail.

4.3℃average annual temperature
< 1.2PUE (cluster design level)
~10months of free cooling a year
4.2 msround trip to Beijing
Emailsales@example.com
WhatsApp+00 000 0000 0000 (to be added)