---
format: "aidr-story-markdown/v1"
id: "8a50208d6ea2f2c6b326174c8fc8233d0f0dbb35c7c176df5daa58e0e4e1ac1d"
canonical_url: "https://aidr.today/8a50208d?lang=en"
title: "Huawei Ascend 950DT Hits 86% of Nvidia H200 Inference Speed in Leaked Report"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-18T10:22:21.000Z"
category: "Chips"
topics: ["nvidia","inference"]
source_urls: ["https://marketbrief.now/ai/huawei-ascend-950dt-hits-86percent-of-nvidia-h200-inference-speed-in-lea-9a718211","https://huggingnews.com/ai/huawei-ascend-950dt-hits-86percent-of-nvidia-h200-inference-speed-in-lea-9a718211"]
summary: "Huawei's Ascend 950DT AI chip achieved 86% of the inference throughput of Nvidia's H200, according to a leaked October test report from China Mobile. In tests using a 64-card supernode, the 950DT processed 1,951 tokens per second per card on the GLM-5.2 model and 1,451 tokens per second on DeepSeek-V4-Pro. This performance is roughly four times faster than the previous Ascend 910C model. The performance gap narrows significantly in inference but remains wide for training. The 950DT's FP16 single-card compute power is 425 TFLOPS, which is about 43% of the H200 SXM's 990 TFLOPS theoretical non-sparse FP16 capacity. The report noted that while the 950DT is competitive in inference economics, large-scale training still depends on factors like communication efficiency and software ecosystems."
---

# Huawei Ascend 950DT Hits 86% of Nvidia H200 Inference Speed in Leaked Report

> [Open the canonical story](<https://aidr.today/8a50208d?lang=en>)

**Published:** 2026-11-18T10:22:21.000Z
**Category:** Chips
**Topics:** nvidia, inference

## Summary

Huawei's Ascend 950DT AI chip achieved 86% of the inference throughput of Nvidia's H200, according to a leaked October test report from China Mobile\. In tests using a 64\-card supernode, the 950DT processed 1,951 tokens per second per card on the GLM\-5\.2 model and 1,451 tokens per second on DeepSeek\-V4\-Pro\. This performance is roughly four times faster than the previous Ascend 910C model\. The performance gap narrows significantly in inference but remains wide for training\. The 950DT's FP16 single\-card compute power is 425 TFLOPS, which is about 43% of the H200 SXM's 990 TFLOPS theoretical non\-sparse FP16 capacity\. The report noted that while the 950DT is competitive in inference economics, large\-scale training still depends on factors like communication efficiency and software ecosystems\.

## Sources

- [Story source](<https://marketbrief.now/ai/huawei-ascend-950dt-hits-86percent-of-nvidia-h200-inference-speed-in-lea-9a718211>)
- [Story source](<https://huggingnews.com/ai/huawei-ascend-950dt-hits-86percent-of-nvidia-h200-inference-speed-in-lea-9a718211>)

