In 2019, when the U.S. blacklisted Huawei over national security concerns, few imagined that just six years later, Huawei would deliver a direct challenge to Nvidia — the unchallenged king of AI hardware.
Yet in April 2025, Huawei’s Ascend 910D has forced global chip watchers, investors, and AI builders to take notice.
Historical Rivalry: How Huawei and Nvidia Came to Clash
| Huawei | Nvidia | |
|---|---|---|
| Founded | 1987 | 1993 |
| Founder | Ren Zhengfei | Jensen Huang |
| Original Domain | Telecom equipment | Graphics processors |
| Pivot to AI | 2018 (Ascend Series) | 2012 (CUDA AI) |
- Nvidia dominates AI training chips globally — with over 80% market share as of 2024.
- Huawei, initially a telecom giant, realized post-sanctions that building an independent AI chip industry was critical for China’s technological sovereignty.
Thus began a secretive, heavily funded push into AI semiconductors, resulting in the Ascend chip family.
How Big Are These Players?
| Metric | Huawei | Nvidia |
|---|---|---|
| 2025 Market Cap | ~$160B (down from ~$500B in 2020, due to sanctions) | ~$2.6 Trillion |
| Employees | ~195,000 | ~36,000 |
| R&D Spending | $24 billion | $12.9 billion |
| HQ | Shenzhen, China | Santa Clara, USA |
Notice how despite sanctions, Huawei invests almost 3x Nvidia’s R&D budget — a crucial factor in Ascend 910D’s rapid advancement.
Huawei Ascend 910D vs Nvidia H100: Where Huawei Stands
Globally, however, Huawei faces fierce rivals:
- Nvidia H100 (5nm, TSMC):
Leading in performance, ecosystem (CUDA), and memory tech (HBM3). - Nvidia B100 (Expected 2025, 4nm, TSMC):
Projected to deliver 30–40% higher performance over H100, introducing HBM3e memory. - AMD Instinct MI300X (6nm, TSMC):
Dominates memory bandwidth benchmarks, key for LLM training. - Cerebras and SambaNova:
Offering revolutionary wafer-scale and reconfigurable AI accelerators, although still niche compared to Nvidia.
Thus, while the Ascend 910D is a giant leap for Huawei internally, the global AI race remains crowded and intense.
Huawei’s strategy:
- Focus first on China, Middle East, Russia, and countries less aligned with U.S. trade policies.
- Offer lower TCO (Total Cost of Ownership) solutions than Nvidia/AMD.
- Push MindSpore as an alternative to CUDA, and pre-build LLMs fine-tuned for its chip.
Detailed Technical Comparison: Why Ascend 910D Matters
| Feature | Ascend 910D | Nvidia H100 | AMD MI300X |
|---|---|---|---|
| Peak FP16 | 1.2 PFLOPS | 1.0 PFLOPS | 1.3 PFLOPS |
| Peak INT8 | 2.4 PFLOPS | 2.0 PFLOPS | 2.6 PFLOPS |
| Memory Bandwidth | 800 GB/s (HBM2e) | 3 TB/s (HBM3) | 5.2 TB/s (HBM3) |
| Supported Model Size | 175B parameters | 530B parameters | 500B parameters |
| Process Node | 7nm (SMIC N+2) | 4N (TSMC) | 5nm (TSMC) |
| Energy Efficiency | +12% better than H100 | Baseline | +10% over H100 |
| TDP (Power Draw) | 350W | 700W | 750W |
✅ Strengths:
- Lower power
- Lower cost (~30–40% cheaper)
- Sanctions-proof design
❌ Weaknesses:
- Slower memory access (HBM2e vs HBM3)
- Software ecosystem less mature
How Sanctions Shaped Huawei’s Strategy
Because SMIC (Huawei’s foundry partner) is restricted from buying advanced EUV lithography machines, the Ascend 910D uses 7nm DUV manufacturing — impressive but behind Nvidia’s 4N process (based on TSMC’s 5nm EUV).
Yet Huawei:
- Increased chip area to fit more transistors.
- Focused on low-frequency operation to control heat.
- Optimized matrix multiplication units for LLMs instead of broad AI workloads.
This is a design-for-purpose chip, not a general-purpose competitor yet.
Risks and Limitations for Huawei’s Ascend 910D
Despite its impressive debut, the Ascend 910D faces critical limitations:
- Manufacturing Constraints:
Built at 7nm while rivals move to 5nm and even 4nm processes. Lower transistor density affects long-term competitiveness. - Memory Bottlenecks:
Limited to older versions of HBM memory, affecting ultra-large model training. - Software Ecosystem:
MindSpore adoption is limited compared to Nvidia’s CUDA global dominance. - International Market Access:
U.S. sanctions restrict Huawei from selling freely outside China, Europe remains cautious.
These factors could prevent Ascend 910D from replicating its domestic success internationally.
Why Nvidia Still Dominates Globally
Nvidia remains the undisputed global AI chip leader, for several strategic reasons:
- Full Stack Advantage:
Beyond hardware, Nvidia controls the AI software world via CUDA, TensorRT, and cuDNN libraries. - Strategic Cloud Alliances:
Deep integration with AWS, Microsoft Azure, Google Cloud ensures Nvidia hardware is embedded in the world’s AI pipelines. - Innovation Pace:
Jensen Huang’s leadership pushes Nvidia to aggressive product refresh cycles — every 18–24 months. - Financial Power:
As of April 2025, Nvidia’s $2.2 trillion market cap dwarfs Huawei’s chip division, which operates under strict budgetary and geopolitical constraints.
In simple terms, while Huawei is climbing, Nvidia built the mountain.
Still a long Battlefield
The launch of Ascend 910D underscores Huawei’s growing ability to innovate under pressure.
It strengthens China’s domestic AI landscape and helps Huawei regain pride after years of tech embargoes.
Yet globally, Nvidia’s lead remains decisive — for now.
Over the next five years, the battle for AI chip supremacy will hinge not just on designing better chips, but also on securing access to cutting-edge manufacturing, building trusted ecosystems, and winning developer loyalty.
The real AI chip wars are just beginning.
Discover more from Semiconductors Insight
Subscribe to get the latest posts sent to your email.
