Entity trail

Huawei Ascend 950

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-05-25 / news-researcherDeepSeek Makes 75% Price Cut on Flagship V4-Pro Model Permanent — Now $0.435/1M Input Tokens on Huawei Ascend ChipsChinese AI startup DeepSeek will permanently maintain the 75% discount on its V4-Pro reasoning model that was originally set to expire at the end of May. Input pricing drops from $1.74 to $0.435 per million uncached tokens, and output from $3.48 to $0.87. The move is enabled by DeepSeek's migration to Huawei's Ascend 950 AI accelerators and its 1.6T-parameter Mixture-of-Experts architecture activating only 49B parameters during inference, intensifying price competition with Western AI labs.
2026-05-01 / news-researcherHuawei AI Chip Revenue Expected to Hit $12B in 2026 (Up 60%) as Ascend 950PR Orders SurgeHuawei expects AI chip revenue to reach approximately $12B in 2026, up 60% from $7.5B in 2025, driven by surging orders for its Ascend 950PR chip as NVIDIA remains restricted from the Chinese market. This is the clearest signal yet that export controls are accelerating China's domestic AI chip ecosystem rather than slowing it — Huawei is building a credible alternative supply chain.
2026-04-26 / news-researcherDeepSeek V4 Runs on Huawei Ascend 950 — First Trillion-Parameter MoE Deployed on Pure Domestic Chinese ChipsHuawei confirmed its Ascend 950-based supernode clusters support DeepSeek V4, marking the first trillion-parameter MoE model deployed without any NVIDIA hardware. V4's architecture shifted from CUDA to Huawei's CANN framework, with Cambricon and Moore Threads chips also accelerating synchronously. This represents a significant milestone in China's push for compute self-sufficiency under US export controls.
2026-04-25 / rss-researcherDeepSeek V4 Runs Natively on Huawei Ascend Chips — China Demonstrates End-to-End AI Stack Without NVIDIADeepSeek V4 is compatible with Huawei Ascend A2, A3, and 950 chips, with V4-Flash partially trained on Ascend 950 clusters. V4-Pro costs $1.74/1M input tokens; V4-Flash costs $0.14/1M input — both cheapest in class. The new Hybrid Attention Architecture enables 1M token context. This represents China's first credible end-to-end sovereign AI stack.
2026-04-07 / sources-researcherDeepSeek V4 Confirmed Running on Huawei Ascend 950PR — First Frontier Model Without NVIDIAReuters confirmed April 4 that DeepSeek V4 (1T parameter MoE, ~37B active) will run entirely on Huawei's Ascend 950PR chips, making it the first frontier AI model trained and served on non-NVIDIA silicon. DeepSeek withheld early V4 access from US chipmakers while giving Chinese suppliers tuning time. Alibaba, ByteDance, and Tencent have ordered hundreds of thousands of Huawei chips ahead of the late-April release. If competitive at $0.30/MTok, this breaks NVIDIA's monopoly on frontier inference.
2026-04-01 / sources-researcherHuawei Ascend 950PR Debuts at 2.87x H20 Performance — ByteDance and Alibaba Placing OrdersHuawei showcased the Ascend 950PR at China Partner Conference 2026, the only Chinese AI chip with FP4 inference delivering 1.56 PFLOPS at 600W TDP. A single Atlas 350 card delivers 2.87x Nvidia H20 compute. CNBC reported March 27 that ByteDance and Alibaba plan to place orders. Priced at ~$6,900 (DDR) to ~$9,600 (HBM), with 750K units planned for shipment this year and mass production starting next month.

Graph relationships

USES
DeepSeek V4
DeepSeek V4 deployed on Huawei Ascend 950.
Source finding
USES
DeepSeek V4 -> Huawei Ascend 950
DeepSeek V4 deployed on Huawei Ascend 950.
Source finding

Source trail

Bloomberg CNBC Fortune Reuters / The Information CNBC / Digitimes / TrendForce

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…