Deep|DeepSeek V4: The First Model Custom-Built for Non-NVIDIA Chips, Optimized for Cost
The most important thing about DeepSeek V4 is not just the capability improvements, and not just the longer context support. What stands out to us is that it appears to have been deeply customized for non-NVIDIA AI accelerators and ASICs — most notably Huawei Ascend. V4’s architectural choices are clearly suited to running on non-NVIDIA hardware and cluster environments.
Source: DeepSeek V4 technical report
Historically, large-model architectures have been optimized around the NVIDIA ecosystem by default. CUDA, Triton, FlashAttention, PagedAttention, NVLink, NVSwitch — these are essentially baseline assumptions. Model labs build the model first, then hand it off to the various kernels and inference frameworks within the NVIDIA ecosystem to squeeze out performance.
The problem with other AI chips and ASICs is that they don’t have a software ecosystem as mature as NVIDIA’s. Ascend, for example, relies more heavily on regular, large-block matrix computation — and is more sensitive to complex small-op kernels, frequent data movement, complicated attention paths, and immature kernel optimization. So when you take a heavily NVIDIA-optimized large model and port it directly to something like Ascend, efficiency typically takes a major hit. Being able to run is one thing; running cheaply, stably, and efficiently is another.
This is what makes V4 interesting. It doesn’t look like a simple scale-up of V3/R1 — instead, the architecture appears to make many “non-NVIDIA-hardware-friendly” choices.
