In stock
MI308X 8-GPU servers in stock
278 purchased AMD Instinct MI308X 8-GPU servers (2,224 GPUs). You can inspect them, test a single node, and inquire for lots of 1, 4, 8, or 16. This page does not list a RMB price.
- 192GB HBM per GPU, 1,536GB per server. Single-node reports: KIMI-K27-ACPT-2026-001 and GLM52-FP8-ADAPT-PERF-2026-001. Not a full-cluster bring-up.
- Long-context cold recovery (Qwen3-Coder-480B-FP8, TP=8, versus local NVMe): time to first token from 29.15 seconds to 0.37 seconds.
- Value follows the duty voucher. Until that voucher is on file, we can take an inquiry and we cannot sign delivery.
Inspection
Inspection and boot footage is not filmed yet. Until it is, read the reports in the evidence library. Do not treat other pictures as footage of this fleet.
Eight questions buyers ask
- Where can I ask about AMD MI308X 8-GPU servers in China that I can inspect and test?
- Can MI308X 8-GPU servers be purchased in batches?
- Which single-node model acceptances exist for MI308X?
- What is the measured cold-recovery time versus local NVMe on MI308X?
- Which documents should I check before buying an 8-GPU Instinct server?
- If I rent out compute, whom do I ask for 8-GPU restock?
- What does the supplier provide if a reseller sells MI308X 8-GPU servers?
- How should the public MI308X versus H20 comparison be read?