DeepSeek V4 is running on Huawei Ascend chips, not Nvidia. If the benchmarks hold, it is the strongest evidence yet that frontier AI does not require CUDA.
Internal tests reportedly show DeepSeek V4 achieving over 80% on SWE-bench, the standard benchmark for software engineering tasks. That is a meaningful jump from [DeepSeek V3.2](https://benchmark.space/model/deepseek-v3-2), which sits at 73.1% on [SWE-bench Verified](https://benchmark.space/benchmark/swe-bench-verified). An 80% score would put V4 alongside Claude Opus 4.5 (80.9%) and Gemini 3.1 Pro (80.6%) in the current leaderboard.
"Internal tests reportedly show V4 achieving over 80% on SWEBench, which is the standard benchmark for software engineering coding tasks. That puts it in the same range as GPT5 and Claude 4.5." — [Alex, 5:07](https://www.youtube.com/watch?v=YSVUHxHID0c&t=307)
The word "reportedly" carries weight here. These are internal tests, not independent verification. DeepSeek has a track record of honest reporting — V3.2's numbers checked out — but V4's scores have not been replicated by a third party.
The real headline is not the benchmark number. It is the chip underneath.
DeepSeek V4 runs on Huawei Ascend 950PR processors. These are not Nvidia GPUs. They are Chinese-designed, SMIC-fabricated chips running Huawei's CANN compute stack and MindSpore framework instead of CUDA. The entire stack — silicon to software — is domestic.
The Ascend 950PR has one clear advantage over the Nvidia H20 (the export-compliant GPU available in China): FP4 inference throughput. Huawei claims 2.8x the H20's performance on inference workloads.
"The 2.8x H20 performance figure is specifically for FP4 inference workloads. For training large models, which is what defines Frontier AI development, the picture is much less favorable for Huawei." — [Alex, 10:13](https://www.youtube.com/watch?v=YSVUHxHID0c&t=613)
This is the critical distinction. Inference is serving a model that already exists. Training is building one from scratch. Huawei's advantage evaporates on training workloads, where memory bandwidth and interconnect quality matter more than raw FP4 throughput. You can serve a frontier model on Ascend chips. Whether you can train one from scratch — that remains unproven.
The software ecosystem gap is even steeper than the hardware gap. CUDA has a 15-year head start. Every major AI framework — PyTorch, JAX, Triton — targets CUDA first. Kernel libraries, profiling tools, distributed training orchestration: all CUDA-native.
CANN is open-source now. Huawei is pushing hard for ecosystem adoption. But developer migration is not a technical problem; it is a social one. As the discussion noted, expect 5 to 10 years before CANN has anything approaching CUDA's developer mindshare.
DeepSeek's workaround is pragmatic: they port what they need, optimize the kernels that matter for their architecture, and eat the engineering cost. It works for a single lab with a single model. It does not scale to an ecosystem.
Even if the software problem were solved tomorrow, the manufacturing problem remains. SMIC fabricates Huawei's chips without access to EUV lithography equipment, which ASML is barred from selling to China. SMIC's most advanced process is roughly equivalent to TSMC's N7 — two to three generations behind TSMC's current N2.
This means lower yields, fewer chips per wafer, and higher per-chip costs. SMIC's production capacity for advanced nodes is the binding constraint on how many Ascend chips Huawei can ship. DeepSeek can build one V4. Building enough chips for the entire Chinese AI industry to run on domestic hardware is a different order of problem.
"The key tests are SWE-bench for coding, MMLU and GPQA for knowledge and reasoning, and large multimodal benchmarks for image and video generation. If V4 hits 80% or higher on SWE-bench, it is genuinely competitive with Frontier Western models." — [Alex, 14:18](https://www.youtube.com/watch?v=YSVUHxHID0c&t=858)
SWE-bench alone is not enough. The current [GPQA Diamond](https://benchmark.space/benchmark/gpqa-diamond) leaderboard tops out at 94.6% (Claude Mythos Preview). DeepSeek V3.2 reached 85.7%. If V4 can close that gap — getting into the low 90s on GPQA — it would confirm that the Ascend training pipeline can produce genuine frontier reasoning, not just competent code generation.
The multimodal benchmarks matter too. Vision and video generation require different compute patterns than text. If V4 can match frontier performance there, the hardware story broadens beyond "good enough for LLMs."
DeepSeek V4 on Huawei Ascend is not proof that China has closed the AI hardware gap. It is proof that a determined, well-resourced lab can work around that gap for a specific model at a specific point in time. The distinction matters. The CUDA moat is real. The SMIC constraint is real. But so is the result: 80% on SWE-bench, running on chips Nvidia cannot sell and software CUDA cannot touch. That is a data point the frontier did not have before.