TBC Home →
Company profiles›Huawei

Huawei华为

Founded 1987Shenzhenhuawei.comReviewed 2026-08-30

Huawei is the closest thing China has to a full-stack semiconductor demand engine: Ascend AI accelerators, Kunpeng CPUs, networking, cloud, devices, and system integration.

Scale

—no public financials
privatenot exchange-listed

7 tracked chips with sourced pricing/specs

Gap vs frontier

≈1×compute per dollar

Best chip Atlas 300I Duo at 100 TFLOPS/$1k (~$1.4K) vs the 2022–2023 global best AMD Instinct MI300X at 131 — chip-level, same generation, market-priced where available.

Huawei: ≈1× behind the global frontier on FP16-per-dollar (Atlas 300I Duo vs AMD Instinct MI300X, 2022–2023 generation).

The chip stackEach layer depends on the layers below it
1 · ChipsWhat the stack produces
AI accelerators · SubstituteCPU · Close behindConnectivity · Substitute
2 · Design toolsSoftware and IP to design them
3 · ManufacturingPlants that turn designs into silicon
4 · EquipmentMachines inside the plants
5 · MaterialsWhat the plants consume

Company at a glance

Scale

Private, not exchange-listed

Growth prospects

Expanding customer base and product roadmap

CNY 862.1B, according to Huawei's 2024 annual report.CNY 179.7B in 2024, about one-fifth of revenue.

China competition

Competes with Biren, Moore Threads, Cambricon, Enflame in AI accelerators — peers listed at sector level.

Substitute

Global comparables

Global reference set: Nvidia, AMD, Google TPU.

Nvidia for AI accelerator ecosystem ambition, but Huawei's business also resembles a mix of Cisco/Ericsson/Nokia, cloud infrastructure vendors, and v…

Technical gap

≈1× behind the global best on compute-per-dollar (Atlas 300I Duo vs AMD Instinct MI300X).

Obstacles

US export-control exposure

listed · US Entity List (BIS), Foreign Direct Product Rule, and related export-control regimes · Restricts access to advanced logic and memory chips,…

Official filings: Huawei 2024 Annual Report ↗

All sourced facts (9)
  • Founded: Founded in Shenzhen in 1987.
  • Headquarters: Shenzhen.
  • Market cap / valuation: Not applicable (private company)
  • Sanctions / export controls: listed · US Entity List (BIS), Foreign Direct Product Rule, and related export-control regimes · Restricts access to advanced logic and memory chips, leading-edge manufacturing equipment, and EDA software, forcing a domestic substitution and supply-chain orchestration strategy.
  • Ownership / listing: Private, employee-owned company; not exchange listed.
  • 2024 revenue: CNY 862.1B, according to Huawei's 2024 annual report.
  • 2024 net profit: CNY 62.6B.
  • R&D spending: CNY 179.7B in 2024, about one-fifth of revenue.
  • Strategic role: Systems, cloud, telecom, and chip-platform anchor rather than a standalone semiconductor vendor.

Tracked chips: Ascend 910B · Ascend 910C · Ascend 950 · Ascend 950PR · Atlas 300I Duo · Huawei Kunpeng 920 · Huawei Kunpeng 930

About

Sells telecom equipment, cloud and enterprise infrastructure, devices, and systems; the semiconductor layer supports Huawei's own products and domestic-stack substitution.

Shenzhen · Founded 1987

Why it matters

Huawei is China's compute-stack anchor: it designs AI accelerators (Ascend) and server CPUs through HiSilicon, integrates them into Atlas-scale AI clusters and cloud, and is characterized by MERICS as quietly orchestrating the domestic semiconductor supply chain under sanctions. The test that matters is whether Ascend 950 can hold NVIDIA-H20-adjacent inference economics at deployment scale; the binding constraint is access to advanced foundry and packaging, which will confirm or break the substitution thesis.

Where it sits in the stack

Compute

AI accelerators / GPUs

Substitute

CUDA-like software ecosystem, HBM supply, advanced packaging, leading-node access, and large-scale deployment proof.

AI trainingAI inferenceEV / autoConsumer / IoT
Compute

CPU

Close behind

Performance per watt, software ecosystem, server qualification, and dependence on advanced foundry processes.

AI trainingAI inferenceRoboticsEV / auto
Compute

Connectivity / networking chips

Substitute

High-speed SerDes, Ethernet switching, optical interconnect, software stack, and global standards ecosystems.

AI trainingAI inferenceRoboticsEV / auto

Key semiconductor products

  • Kunpeng general-purpose CPU platform for servers and cloud infrastructure.
  • Huawei Cloud infrastructure and AI services, including the Pangu model ecosystem.
  • Carrier network equipment, enterprise networking/storage, devices, digital power, and intelligent automotive solutions.
  • Ascend AI accelerators
  • Atlas SuperPods
  • Kunpeng server CPUs
  • HiSilicon SoCs

Chip pricing

Ecosystem

Investors / affiliates

  • HiSilicon

Competitors

  • NVIDIA

Customers / end markets

  • South Korean AI chip customers
  • Carrier networks, cloud and enterprise customers, government and state-owned enterprise procurement, financial services, manufacturing, devices, and automotive OEM partnerships
  • The most important semiconductor question is not just whether Ascend or Kunpeng chips exist, but whether Huawei can pull memory, packaging, networking, OS, compiler, and application layers into deployable systems

Constraints

  • External foundry and packaging partners

Global peer set

  • Nvidia for AI accelerator ecosystem ambition, but Huawei's business also resembles a mix of Cisco/Ericsson/Nokia, cloud infrastructure vendors, and vertically integrated device companies
  • Intel, AMD, Arm, and cloud-provider custom-silicon programs are useful comparables for the CPU/platform side, but none maps perfectly

Affiliates

  • HiSilicon

Competitive position

  • Huawei is not cleanly comparable to a pure chip designer. Its leverage comes from owning end markets - telecom, enterprise, cloud, devices, and auto - that can absorb domestic silicon.
  • For the tracker, Huawei is best treated as a demand aggregator and platform coordinator: it can turn otherwise narrow domestic component progress into systems customers actually buy.
  • Ascend 950 is positioned against NVIDIA H20 on inference performance and cost; frontier training and node access remain constrained by sanctions.
  • Fabless dependency on external foundry and advanced packaging, compounded by export controls on tools and equipment.
  • Global comparables: Nvidia for AI accelerator ecosystem ambition, but Huawei's business also resembles a mix of Cisco/Ericsson/Nokia, cloud infrastructure vendors, and vertically integrated device companies.
  • Global comparables: Intel, AMD, Arm, and cloud-provider custom-silicon programs are useful comparables for the CPU/platform side, but none maps perfectly.
  • Advanced foundry access, high-bandwidth memory, advanced packaging, EDA/IP constraints, and export-control exposure remain the key semiconductor-layer constraints.
  • The system-level risk is software: domestic compute hardware only matters if compilers, frameworks, libraries, model tooling, and customers converge around it.

Milestones and signals

  • 1987—Huawei was founded.
  • 2026 coverage—Other: Atlas SuperPod deployments contain 8 to 192 Ascend 950 accelerators
  • Listing—Private, employee-owned company; not exchange listed.

What to watch

  • Ascend deployments in Huawei Cloud and state/enterprise AI infrastructure.
  • Kunpeng server adoption beyond policy procurement.
  • signals of domestic HBM, interconnect, packaging, and accelerator software maturity around Huawei systems.
  • Whether automotive and telecom demand keeps funding the semiconductor stack despite export controls.

Questions

Last reviewed 2026-06-28

Answered questions

Can Ascend and Kunpeng scale through real cloud and enterprise workloads rather than one-off procurement wins?

Named exposure so far: Carrier networks, cloud and enterprise customers, government and state-owned enterprise procurement, financial services, manufacturing, devices, and automotive OEM partnerships, The most important semiconductor question is not just whether Ascend or Kunpeng chips exist, but whether Huawei can pull memory, packaging, networking, OS, compiler, and application layers into deployable systems, and South Korean AI chip customers.

Open questions

  • How much of the surrounding stack - memory, packaging, networking, tools, and software - can Huawei pull into a domestic ecosystem?
  • Where should tracker entries separate Huawei parent-company demand from specific chip-platform claims?