Just as Nvidia crossed the $5 trillion market capitalisation mark, China’s DeepSeek has dropped a new set of large language models that could once again test the chip giant’s dominance, this time with a deeper shift away from its hardware ecosystem.

DeepSeek’s V4 family, launched the same day Nvidia closed at a $5 trillion market cap, is the company’s most ambitious release since its R1 model rattled global markets in early 2025, wiping nearly $600 billion off Nvidia’s valuation in the aftermath.

On a recent podcast, Nvidia CEO Jensen Huang said that if Chinese AI startup DeepSeek optimised its new models on chips from Huawei Technologies, it would be “a horrible outcome” for the US.

If “future AI models are optimised in a very different way than the American tech stack”, and as “AI diffuses out into the rest of the world” with Chinese standards and technology, China “will become superior to” the US, Huang had said on the Dwarkesh Podcast.

Must read: China narrows AI gap with US to near zero, Stanford report shows

Cheaper, faster and not Nvidia-powered

At the centre of the release is V4-Pro, a 1.6 trillion-parameter model designed for coding and complex agentic tasks, alongside a smaller V4-Flash variant built for speed and cost efficiency.

But the bigger shift lies under the hood.

For the first time, DeepSeek has optimised its flagship model for domestic chips from Huawei Technologies, rather than Nvidia GPUs, marking a significant step in China’s efforts to reduce reliance on US technology.

Earlier DeepSeek models, including V3 and R1, were trained on Nvidia hardware. V4, in contrast, is positioned as a test case for whether China’s AI ecosystem can stand on its own chip stack.

Frontier performance at a fraction of the cost

DeepSeek is once again leaning heavily on cost disruption. V4-Pro activates only 49 billion parameters per token despite its 1.6 trillion total, allowing it to deliver near frontier-level performance at significantly lower compute costs. The company claims this enables output comparable to leading models but at a fraction of the price.

Its API pricing underscores that pitch. V4-Pro costs $1.74 per million input tokens and $3.48 per million output tokens, roughly 50 times cheaper than models like Claude Opus. The V4-Flash model is even more aggressive, with pricing as low as $0.14 per million input tokens.

This follows the template set by DeepSeek’s earlier R1 model, which it claimed was trained for just $6 million in two months, far below industry norms. By comparison, Meta Platforms reportedly spent $60 million on Llama, while OpenAI has invested billions into its models.

Must read: DeepSeek returns after year-long silence, researcher warns of AI’s threat to human jobs

Benchmarking with the best

On performance, DeepSeek says V4-Pro competes with top closed-source models such as GPT-5.4 and Gemini 3.1, while outperforming several open-source alternatives on coding, math, and STEM benchmarks.

It also introduces a 1 million token context window, up from 128,000 tokens in its previous flagship, allowing it to process significantly larger datasets in a single pass.

In long-context scenarios, V4-Pro reportedly uses just 27% of the computing power required by its predecessor, while V4-Flash cuts that further to 10%.

Huawei partnership sharpens geopolitical edge

Shortly after the launch, Huawei announced full support for DeepSeek’s V4 models across its Ascend chips, reinforcing the alignment between China’s AI software and domestic hardware stack.

The move comes amid reported policy pressure in China to increase the use of local chips, including sourcing quotas and requirements to pair foreign hardware with domestic alternatives.

For Nvidia, the concern is less about a single model and more about the larger trend.

Its dominance has been built not just on GPUs, but on a tightly integrated software ecosystem. Shifting to Huawei’s Ascend chips requires rewriting code, rebuilding tools and proving performance at scale, barriers that have so far protected Nvidia’s lead.

But if players like DeepSeek can demonstrate comparable performance at significantly lower cost on alternative hardware, that moat could begin to narrow.

DeepSeek’s R1 launch last year showed how quickly sentiment can shift. The model’s efficiency and performance triggered a wave of open-source releases across China and challenged assumptions around the cost of building frontier AI.

V4 builds on that momentum, but with a more strategic twist. By pairing low-cost models with domestic chips, DeepSeek is not just competing on performance; it is aligning with a broader push to localise the AI stack.