AI Infrastructure DatabaseThe chips, power and companies behind AI
Saved
Menu
Home How It's Built Watchlist The Stack
Hardware ChipsCompare Chips Servers & SystemsNetworking MemoryFoundries & Packaging Cooling
Facilities Data CentersPower ProjectsCloud & GPU Compute
Intelligence AI ModelsSoftware AI ProductsBenchmarks Research PapersDatasets Concepts
Ecosystem CompaniesStocks PeopleTimeline TrendsFuture of AI
Learn News Ask the Data AI Agents
About About us Methodology Data & API Disclaimer
Options
DATA & API

Free read-only JSON access to the site's cached data.

Dark mode

Guided View
New to chips, data centers and model jargon? We explain key terms as you browse, in plain English. Same data, with the help built in.

Expert View
You already know the stack. Just the data: clean, fast and compact, with no extra explanations. This is the default view.

Interface language
‹ Back
Home › Chips › Huawei Ascend 950 (950PR / 950DT)

Huawei Ascend 950 (950PR / 950DT)

Chip AI Accelerator Ramping

Huawei's 2026 generation: in-house HBM, FP4 support, 8,192-chip SuperPoDs.

You're looking at a chip record: what this processor is, its key specs, who designs, manufactures and supplies its parts, and where it gets used.

The Ascend 950 series succeeds the 910C. Huawei unveiled it at Huawei Connect 2025 with a claimed 1 PFLOPS FP8 and 2 PFLOPS FP4 per chip and its own HBM. The inference-oriented 950PR launched in March 2026 on the Atlas 350 card; Reuters reported mass production was due to start in April, with a target of about 750,000 chips in 2026, with ByteDance and Alibaba planning orders. The training-oriented 950DT and the 8,192-chip Atlas 950 SuperPoD are slated for Q4 2026. Performance figures are Huawei claims without independent benchmarks.

Specifications & key facts

Each figure is labeled by basis: vendor spec, vendor claim, reported, estimate, or official. "Not publicly disclosed" means exactly that: we never invent values.

Overview

Designer Huawei (HiSilicon)
Variants 950PR (prefill and recommendation inference); 950DT (decode and training)Official Sep 2025, Huawei Connect keynote; Huawei
Status 950PR launched 2026-03-20 (Atlas 350 card); mass production expected from April, full shipments 2H 2026Reported Mar 2026; TrendForce; Reuters via CNBC
Timeline 950PR Q1 2026; 950DT and Atlas 950 SuperPoD Q4 2026; Ascend 960 Q4 2027; Ascend 970 Q4 2028Official Sep 2025, Huawei Connect keynote; Huawei
Price About CNY 50,000 (DDR version) / CNY 70,000 (HBM version) per cardReported 2026-03-27; Reuters via CNBC

Compute

FP8 1 PFLOPSVendor claim Per chip, 950 series; Huawei, Sep 2025
FP4 (MXFP4) 2 PFLOPSVendor claim Per chip, 950 series; Huawei, Sep 2025

Memory

950DT memory 144 GB HiZQ 2.0 HBM, 4 TB/sVendor claim Huawei Connect, Sep 2025
950PR memory HiBL 1.0 (Huawei low-cost HBM)Vendor claim Huawei Connect, Sep 2025

Physical

Interconnect 2 TB/sVendor claim Per chip; 2.5x Ascend 910C; Huawei, Sep 2025
Atlas 350 card (950PR) 1.56 PFLOPS FP4, 112 GB HBM at 1.4 TB/s, 600 WReported Mar 2026; TrendForce citing Huawei and Mydrivers
Atlas 950 SuperPoD Up to 8,192 950DT chips, 8 EFLOPS FP8 / 16 EFLOPS FP4, 1,152 TB memory, 16 PB/s interconnectVendor claim Huawei Connect, Sep 2025; due Q4 2026

Manufacturing

Foundry / process Not publicly disclosed Not disclosed by Huawei; reported to be SMIC 7 nm-class

Customers

Planned buyers ByteDance and Alibaba planning 950PR orders after testingReported 2026-03-27; Reuters via CNBC

How this connects to the rest of the stack

This record's links at a glance: click a group to jump to it.

Designed by

Successor to

Competes with

Sources

TrendForce: Huawei debuts Atlas 350 on Ascend 950PR (Mar 2026) ↗ · CNBC (Reuters): ByteDance, Alibaba plan to order Huawei's new AI chip (Mar 2026) ↗ · Huawei Connect 2025 keynote: Ascend roadmap and SuperPoDs ↗

Record reviewed as of 2026-11-16; see Methodology.