DeepSeek, a prominent Chinese AI startup, has unveiled its V4 family of large language models, a move that directly tests Nvidia's formidable presence in the artificial intelligence hardware market. The launch of DeepSeek V4 coincided with Nvidia's market capitalization crossing the $5 trillion mark, setting the stage for a high-stakes competition.
Huawei Partnership Bolsters China's Domestic AI Stack
A key differentiator for the V4 models is their optimization for Huawei Technologies' Ascend chips, rather than Nvidia GPUs, which were used for earlier DeepSeek iterations like V3 and R1. This strategic alignment reinforces China's concerted effort to integrate its AI software with domestically produced hardware, reducing reliance on foreign technology.
Nvidia CEO Jensen Huang previously voiced concerns about such a development, stating on the Dwarkesh Podcast that if future AI models were optimized on a non-American tech stack, it could lead to China becoming superior in AI globally. Huawei has announced full support for DeepSeek's V4 models across its Ascend chip range, solidifying this domestic partnership.
V4-Pro and V4-Flash: Performance and Cost Efficiency
The new family includes V4-Pro, a massive 1.6 trillion-parameter model designed for complex coding and agentic tasks, and V4-Flash, a smaller variant optimized for speed and cost-efficiency. DeepSeek claims V4-Pro achieves near frontier-level performance while activating only 49 billion parameters per token, significantly lowering compute costs.
This cost disruption is evident in its API pricing: V4-Pro costs $1.74 per million input tokens and $3.48 per million output tokens, which DeepSeek states is approximately 50 times cheaper than models like Claude Opus. V4-Flash offers even more aggressive pricing, starting as low as $0.14 per million input tokens. This strategy echoes DeepSeek's earlier R1 model, reportedly trained for just $6 million, far below industry norms.
Benchmarking Against Industry Leaders
DeepSeek asserts that V4-Pro competes favorably with top closed-source models such as GPT-5.4 and Gemini 3.1, outperforming several open-source alternatives across coding, math, and STEM benchmarks. The V4 models also introduce an expanded 1 million token context window, a substantial increase from the previous 128,000 tokens, enabling them to process significantly larger datasets.
In long-context scenarios, V4-Pro reportedly uses only 27% of the computing power required by its predecessor, with V4-Flash cutting that further to 10%.
Geopolitical Implications and Nvidia's Challenge
This move comes amidst reports of policy pressure within China to increase the adoption of local chips, including sourcing quotas. For Nvidia, the challenge extends beyond a single model; its market dominance is built on a comprehensive, integrated software ecosystem alongside its powerful GPUs. Shifting to Huawei's Ascend chips necessitates extensive code rewriting, tool rebuilding, and performance validation at scale, barriers that have historically protected Nvidia's lead.
However, if companies like DeepSeek can consistently demonstrate comparable performance at a significantly reduced cost on alternative hardware, Nvidia's competitive moat could begin to erode. DeepSeek's previous R1 launch showcased the potential for rapid shifts in market sentiment, and V4, with its strategic pairing of low-cost models and domestic chips, aims to further localize China's AI infrastructure.