Specifications
Sourced specs
| Spec | Value | Source |
|---|---|---|
| FP16 compute | 345 TFLOPS | via techinsights.com ↗2026-05-07 |
| Memory capacity | 96 GB | via spectrum.ieee.org ↗2025 |
| Process node | 7 nm | via flopper.io ↗2026-08 |
| Foundry | Samsung Foundry (reported) | via ft.com ↗2025-04-03 |
| Interconnect | 600 GB/s (NVLink-compatible or PCIe 4.0) card-to-card (8-card Baie AI-in-one machine) | via cloud.baidu.com ↗2025-02 |
| Single-server LLM capability | 8 P800 cards run full DeepSeek-V3/R1 671B (8-bit inference) | via cloud.baidu.com ↗2025-02 |
| Deployment scale | 30,000-card P800 cluster lit April 2025 | via reuters.com ↗2025-04-25 |
| 2024 unit volume (China AI accelerator cards) | 69,279 cards shipped, No.2 domestic vendor | via moomoo.com ↗2025-07-09 |
| External order win | Over 1 billion yuan (~$139M) of orders via China Mobile first-tier suppliers | via reuters.com ↗2025-08-22 |
Pricing
Price history, newest first
Step chart of per-unit observations (system and order totals stay in the table). The line is drawn in USD at each observation year's average FX rate so mixed-currency histories read as one trend; hover any point for the price exactly as reported.
| Date | Price | Type | Source | Confidence |
|---|---|---|---|---|
| 2026-08-27 | ~$13.2K from ¥96,000 at 2026 avg 7.25 CNY/USD per cardUndated aggregator listing observed 2026-08-27, described as an official marked price for the card. | Reseller listing | via Taobao e-commerce aggregator listing ↗ | low |
| 2025-07 | ~$13.9K from ¥100,000 at 2025 avg 7.19 CNY/USD per cardPerson familiar with the matter: P800 pricing is close to Nvidia's China-market H20, 'comparable to the H20, or even somewhat higher depending on the promotion strategy'. | Reported contract | via Tencent Tech (via Moomoo) ↗ | medium |
vendor-official or multiple independent reports · single major report or several secondary · estimate or lone secondary source
Prices in USD, converted at each observation year's average exchange rate; the original figure and rate print under every converted price. How prices are sourced and graded ↗
Context
Notes
- Memory type is contested: Flopper.io's teardown-based analysis says GDDR6 and explicitly rejects '96 GB HBM3' marketing claims, while Chinese coverage (e.g., CSDN) says 96 GB HBM3; TDP has not been disclosed in any located source.
- Architecture naming conflicts: Flopper.io labels P800 as third-generation 'XPU-P', while Tencent Tech describes it as using the second-generation XPU-R architecture.
- Roadmap: Reuters (2025-12-05) and CNBC (2025-11-28) report successors M100 (inference, early 2026) and M300 (training + inference, early 2027); CNBC describes a five-year Kunlun roadmap.
- Also ships as a P800 OAM module compatible with mainstream OAM 2.0 platforms (e.g., Guoxin 'Haiqing' servers) per Kunlunxin supernode announcement (https://www.kunlunxin.com/news/4541.html).
- Kunlunxin's IPO materials acknowledge adverse restrictions on advanced-node wafer foundry and HBM supply due to relevant policies, per Tencent Tech (2025-07).
- Production timeline per Tencent Tech: chip returned from fab in H1 2024; FT (2025-04) describes the P800 as 'released in February' (2025).
Where it sits
In the stack
All specs and prices link to their original sources, dated as observed. Figures are market observations, not offers; treat each as of its date.