Last year, DeepSeek-R1 forced every major AI lab to rethink their pricing strategies. Now the Chinese research outfit is back with DeepSeek-V4 - and this release carries a detail that goes well beyond benchmark scores: the models run on Huawei's Ascend AI chips for inference (the compute work that happens when a model answers your question).
That combination - open weights, low cost, non-Nvidia hardware - is a deliberate signal. The U.S. has spent the last two years restricting advanced chip exports to China. Huawei's Ascend chips exist partly because of those restrictions. DeepSeek using them for production inference suggests China's domestic chip industry is further along than many Western analysts have assumed.
What DeepSeek-V4 Actually Is
The V4 series continues DeepSeek's pattern of releasing capable open-weight models - meaning anyone can download the model weights and run them on their own infrastructure, without paying per-query fees. That matters a lot for businesses handling sensitive data or wanting predictable costs.
The "low-cost" positioning is central to DeepSeek's strategy. Previous releases used a Mixture of Experts (MoE) architecture - a design where only a fraction of the model's parameters activate for any given query, which keeps the cost of running each query much lower than a similarly capable dense model. The V4 release appears to continue that approach.
For teams already testing open models as alternatives to ChatGPT, DeepSeek-V4 will go on the evaluation list immediately. Previous DeepSeek releases have been genuinely competitive on coding, reasoning, and multilingual tasks.
The Huawei Chip Connection
Nvidia's H100 and H200 GPUs dominate AI inference workloads globally. U.S. export controls mean Chinese companies can't legally import them. Huawei's Ascend chips have been positioned as the domestic alternative, and early assessments placed them meaningfully behind Nvidia in raw performance.
If DeepSeek can run competitive models on Ascend hardware at production scale, that's evidence the gap is closing. It also means China's full AI pipeline - training, model architecture, deployment - is becoming self-sufficient in ways that make export controls less effective over time.
For Practitioners
The immediate practical question is whether DeepSeek-V4 performs well enough to use. Benchmark positioning will become clearer in the weeks after release. Organizations already running open models on-premises will likely test it; those considering a switch from cloud-hosted APIs should do the same.
The geopolitical layer adds a separate consideration. Open-weight models from Chinese labs have faced regulatory scrutiny in some jurisdictions - Italy blocked DeepSeek briefly in early 2025 over data privacy concerns. Companies in regulated industries or with government contracts should factor that uncertainty into any evaluation.
What's consistent across every DeepSeek release: capable open models at low prices, with shrinking dependence on Western hardware. That pattern puts steady downward pressure on the economics of the entire AI tools market, and V4 doesn't look like an exception.