For six years, conventional wisdom held that you could not train frontier AI models without NVIDIA. Every major model — from GPT-4 to DeepSeek V4 Pro to Claude Fable 5 — runs on NVIDIA silicon. Then Huawei released openPangu 2.0, a 505B parameter model trained entirely on Ascend NPUs, and that conventional wisdom cracked.

This is not a review saying Ascend is better than NVIDIA. It is not. But the question was never “is Ascend better?” It was “is Ascend good enough?” And as of June 12, 2026, the answer is definitively yes — at least for Huawei’s own use case.

Here is what developers, infrastructure engineers, and technology strategists need to know about the Ascend vs NVIDIA landscape in 2026.

The sanction backstory

In 2020, the US placed Huawei under strict chip export controls. No American company could sell advanced semiconductors or semiconductor manufacturing equipment to Huawei. This included NVIDIA’s entire data center GPU lineup — A100, H100, and all subsequent chips.

Huawei’s response was to build their own AI training hardware. The Ascend 910B emerged as their flagship training chip, manufactured on SMIC’s N+2 process (roughly 7nm equivalent). The upcoming Ascend 950DT is the next generation.

The key point: Huawei did not just design chips. They built the entire stack — the silicon, the interconnects, the software framework (MindSpore), the compiler toolchain (CANN), the cluster management, and the training orchestration. Everything from transistor to trained model runs on Huawei technology.

openPangu 2.0 is the proof that this stack works at scale. A 505B parameter MoE model with 512K context trained end-to-end on Ascend. See our openPangu 2.0 complete guide for full model details.

Hardware specifications compared

Huawei Ascend 910B:

• Process: 7nm (SMIC N+2)
• AI compute: ~320 TFLOPS FP16
• Memory: 64GB HBM2e
• Memory bandwidth: ~1.6 TB/s
• TDP: ~310W
• Interconnect: HCCS (Huawei proprietary, NVLink equivalent)
• Availability: unrestricted in China and non-sanctioned markets

Huawei Ascend 950DT (upcoming):

• Process: 5nm (estimated, unconfirmed fab)
• AI compute: ~600+ TFLOPS FP16 (projected)
• Memory: 128GB HBM3 (projected)
• Interconnect: next-gen HCCS
• Status: announced, production timeline unclear

NVIDIA A100 (baseline reference):

• Process: 7nm (TSMC N7)
• AI compute: 312 TFLOPS FP16 (with sparsity)
• Memory: 80GB HBM2e
• Memory bandwidth: 2 TB/s
• TDP: 300W
• Interconnect: NVLink 3.0 (600 GB/s)

NVIDIA H100:

• Process: 4nm (TSMC N4)
• AI compute: 989 TFLOPS FP16 (with sparsity)
• Memory: 80GB HBM3
• Memory bandwidth: 3.35 TB/s
• TDP: 350W
• Interconnect: NVLink 4.0 (900 GB/s)

NVIDIA B200 (current flagship):

• Process: 4nm (TSMC N4P)
• AI compute: ~2500 TFLOPS FP16 (with sparsity)
• Memory: 192GB HBM3e
• Memory bandwidth: ~8 TB/s

Performance analysis: Ascend 910B vs the field

On paper, Ascend 910B matches A100 in raw FP16 TFLOPS (~320 vs ~312). The real-world story is more nuanced:

Where Ascend 910B competes well:

• Matrix multiplication throughput (core transformer ops)
• Large-batch training scenarios
• Inference at scale (especially for Huawei-optimized models)
• Power efficiency at the chip level

Where NVIDIA maintains advantages:

• Software ecosystem maturity (CUDA is 17 years old)
• Memory bandwidth (A100: 2 TB/s vs 910B: ~1.6 TB/s)
• Interconnect bandwidth (NVLink > HCCS at comparable scales)
• Third-party framework support
• Multi-tenant inference optimization

The honest assessment: Ascend 910B is roughly A100-class. It is competitive but does not match H100, let alone B200. Huawei compensates with larger clusters — training openPangu 2.0 likely required more chips than would be needed with equivalent H100s.

But “more chips” is not a problem when you manufacture them yourself and the alternative is “zero chips because sanctions.”

The software ecosystem gap

Hardware is one thing. Software is where NVIDIA’s real moat lives.

NVIDIA’s CUDA ecosystem:

• 17 years of development
• Supported by every major framework (PyTorch, TensorFlow, JAX)
• Thousands of optimized libraries (cuDNN, cuBLAS, NCCL, TensorRT)
• Massive developer community
• Every AI paper assumes NVIDIA hardware

Huawei’s CANN/MindSpore ecosystem:

• ~5 years of development
• MindSpore framework (first-party, growing but smaller)
• PyTorch backend available via torch_npu
• Smaller library ecosystem
• Developer community concentrated in China
• Growing but still catching up

For developers outside China, the software ecosystem gap is the bigger barrier than hardware specs. If you have code that runs on PyTorch+CUDA, porting to Ascend requires work. The torch_npu bridge helps, but not every operation has an optimized Ascend kernel.

That said, the gap is narrowing. And if you are using openPangu 2.0 via API (ModelArts), the hardware underneath is invisible to you. You just make API calls.

What openPangu 2.0 proves about Ascend capability

Training a 505B parameter MoE model requires solving several hard infrastructure problems:

1. Massive parallelism: Thousands of accelerators working together
2. High-bandwidth interconnects: Moving activations and gradients between chips
3. Memory management: Handling optimizer states for hundreds of billions of parameters
4. Training stability: Maintaining numerical stability across long training runs
5. Fault tolerance: Handling hardware failures without losing training progress

Huawei solved all of these on Ascend. The model exists, it generates coherent text, it handles 512K context — these are not things you can fake. The training infrastructure works.

This does not mean any developer can replicate this. Huawei has thousands of engineers building this stack. But it proves that the Ascend platform is not a toy — it is production-grade for frontier AI training.

For perspective on what this means for the broader landscape, see our coverage of sovereign AI models in 2026.

Cost and availability

NVIDIA (for those who can buy it):

• H100 SXM: ~$30,000-40,000 per unit (if you can get allocation)
• Cloud pricing: $2-4/hour per H100 (see best cloud GPU providers 2026)
• Supply-constrained in many markets
• Export restrictions limit availability in China, Russia, Iran

Ascend 910B:

• Pricing: not publicly listed (sold through Huawei enterprise channels)
• Available without restrictions in China and non-sanctioned markets
• No supply constraints for Huawei’s own use
• Available via Huawei Cloud for developers who want managed access

The economic equation depends entirely on your geography and regulatory environment. If you are in the US or Europe with no supply constraints, NVIDIA is the more mature choice. If you are in a sanctioned market, or if you cannot secure NVIDIA allocation, Ascend is the only viable path for large-scale training.

Beyond NVIDIA and Ascend: the alternatives

Huawei is not the only non-NVIDIA path:

Google TPUs:

• v5p: competitive with H100 for transformer training
• Mature software stack (JAX/XLA)
• But: cloud-only, no on-premise option
• Google trains their own frontier models on TPUs (Gemini)

AMD Instinct MI300X:

• 192GB HBM3 per chip
• Strong raw specs
• ROCm ecosystem improving but still behind CUDA
• Available on open market (no sanctions issues)

Intel Gaudi 3:

• Competitive for inference workloads
• Smaller developer ecosystem
• Owned by Intel, backed by significant R&D budget

Custom ASICs (Groq, Cerebras, etc.):

• Specialized for inference or training
• Limited availability
• Narrow use cases

The landscape is diversifying. NVIDIA remains dominant but the “NVIDIA or nothing” era is ending. openPangu 2.0 and Google’s Gemini (TPU-trained) prove that frontier models can emerge from non-NVIDIA hardware. See our GPU vs CPU for AI inference guide for inference-specific hardware considerations.

What this means for sovereign AI

Sovereign AI — the ability of a nation or organization to develop and deploy AI without dependency on foreign technology — is the strategic implication of openPangu 2.0.

Countries under sanctions now have a proof point. China demonstrated that a complete AI stack is possible without US technology. This matters for:

• Middle Eastern nations investing in AI infrastructure
• Southeast Asian governments hedging technology dependencies
• European organizations concerned about US cloud dependence
• Any entity worried about future export controls

The Ascend platform is available for purchase in most non-sanctioned markets. Combined with an open-source model like openPangu 2.0, you have the building blocks for a fully non-US AI stack.

Whether the quality matches NVIDIA-trained models is a secondary concern for sovereignty-focused buyers. Having the capability at all is what matters. For more on this topic, see our sovereign AI models 2026 analysis.

Developer implications

If you are building AI applications today, here is what the Ascend vs NVIDIA landscape means for you:

Short-term (next 6 months):

• NVIDIA remains the safe default for development
• openPangu 2.0 is accessible primarily via Huawei Cloud API
• Community NVIDIA-compatible weight conversions will enable standard GPU inference
• No need to switch hardware unless you have specific sovereignty requirements

Medium-term (6-18 months):

• Expect more models trained on Ascend as the ecosystem matures
• Ascend 950DT could narrow the gap with H100
• Software ecosystem for Ascend will improve with more users
• More cloud providers may offer Ascend instances

Long-term (18+ months):

• The AI hardware market becomes genuinely multi-vendor
• Competition benefits everyone through lower prices
• NVIDIA’s monopoly pricing faces pressure
• Training on mixed hardware clusters may become viable

For developers evaluating hardware purchases today, check our NVIDIA RTX Spark guide for consumer-grade options and best cloud GPU providers 2026 for cloud alternatives.

The interconnect problem

One detail that gets overlooked: training frontier models is not just about individual chip performance. It is about how fast chips can communicate. Gradient synchronization across thousands of accelerators requires massive interconnect bandwidth.

NVIDIA’s NVLink and NVSwitch provide 900 GB/s per GPU in H100 configurations. Their InfiniBand networking handles inter-node communication at 400 Gb/s+.

Huawei’s HCCS (Huawei Cloud Computing Switch) is their proprietary interconnect equivalent. Exact specs are not publicly disclosed, but training a 505B parameter model successfully implies that it handles the communication patterns required for massive-scale data parallelism, tensor parallelism, and pipeline parallelism.

This is actually one of Huawei’s advantages — they control the entire networking stack from chip to switch, similar to how NVIDIA controls the GPU-to-GPU communication path. No third-party networking layers introducing latency.

Should you care about Ascend?

Yes, if:

• You operate in markets with NVIDIA supply constraints
• Sovereignty is a deployment requirement
• You are evaluating long-term technology risk
• You want to use openPangu 2.0 optimally
• You are building for the HarmonyOS ecosystem

Not yet, if:

• You have reliable NVIDIA supply
• Your existing code runs on CUDA
• You need broad third-party framework support
• Your team lacks Ascend expertise
• You are targeting Western markets exclusively

The pragmatic answer for most developers: use openPangu 2.0 via API (hardware-agnostic), keep an eye on Ascend’s evolution, and recognize that the monopoly is breaking. You do not need to switch today, but the option existing is valuable for everyone.

FAQ

Is Huawei Ascend 910B as fast as NVIDIA H100?

No. Ascend 910B is roughly A100-class in raw compute (320 vs 312 TFLOPS FP16). H100 offers approximately 3x the compute of 910B. Huawei compensates with larger clusters for training and the upcoming 950DT is expected to narrow this gap significantly.

Can I buy Ascend hardware in the US or Europe?

Ascend hardware availability varies by region. Huawei faces its own sanctions that restrict sales in some Western markets. In practice, most non-Chinese developers will access Ascend capabilities via Huawei Cloud rather than purchasing hardware directly. Contact Huawei enterprise sales for your specific region.

Will other companies train models on Ascend?

Yes. Chinese cloud providers and AI companies are already using Ascend for training workloads. As the ecosystem matures and openPangu 2.0 validates the platform, expect more models trained partially or fully on Ascend hardware. Huawei Cloud offers Ascend clusters for customers who want to train their own models.

How does CANN compare to CUDA for developers?

CANN is functional but less mature. It handles the core operations needed for transformer training and inference. However, the library ecosystem is smaller, community resources are fewer, and debugging tools are less polished. The torch_npu bridge provides some compatibility, but expect a learning curve if coming from CUDA. The gap is closing but still meaningful in 2026.

Does this mean NVIDIA’s dominance is ending?

Not immediately. NVIDIA still has the best hardware (B200), the most mature software (CUDA), and the largest ecosystem. But the monopoly is cracking. Between Ascend, Google TPUs, AMD MI300X, and custom ASICs, the market is diversifying. openPangu 2.0 is one proof point among several that frontier AI is no longer NVIDIA-exclusive.

What about the Ascend 950DT?

The 950DT is Huawei’s next-generation training chip, expected to use a more advanced process node and offer significantly more compute than the 910B. Details remain limited, but if it delivers on projections (~600+ TFLOPS), it would sit between H100 and B200 in raw performance. Timeline for availability is unclear.