DeepSeek-V4 Could Change Global AI Model Race

DeepSeek's newly launched AI model shows how the Chinese startup's focus on open source and lower costs could help China advance in the geopolitical AI race, especially as the vendor begins to break its reliance on U.S. AI chip giant Nvidia.

DeepSeek released DeepSeek-V4 in preview on April 24. The open-source model can process longer prompts than previous DeepSeek models. It is also the startup's most significant release since its R1 launch in January 2025, which shocked the AI market with its reasoning capability and low price.

DeepSeek-V4 comes in two versions: V4-Pro and V4-Flash. V4-Pro is larger, with 1.6 trillion total parameters, while V4-Flash is a high-speed version with 284 billion parameters. Both versions support context lengths of up to 1 million tokens, making them comparable in long-context scenarios to Google Gemini and Anthropic’s Claude. The models are designed for long-horizon reasoning, coding and agentic workflows.

The Cost Factor

DeepSeek is positioning V4 as a lower-cost alternative to leading frontier models.

• V4-Pro: $1.74 per million input tokens and $3.48 per million output tokens
• V4-Flash: $0.14 per million input tokens and $0.28 per million output tokens

By comparison:

• Gemini 3.1 Pro: $2 per million input tokens and $12 per million output tokens
• GPT 5.5: $5 per million input tokens and $30 per million output tokens
• Claude Opus 4.7: $5 per million input tokens and $25 per million output tokens

"The token pricing is a third of the frontier labs' pricing," said Kashyap Kompella, CEO of RPA2AI Research. "That kind of pricing can change buying behavior."

Kompella added that while DeepSeek’s models may trail frontier offerings from OpenAI, Google or Anthropic by three to six months, the performance gap may be less important if the cost gap is substantial.

"Enterprises do not always need the absolute best model for all use cases," Kompella said. "They need good enough performance, predictable cost and control. DeepSeek is forcing Western frontier labs to innovate on cost, not only on model capabilities."

Moving Away from Nvidia

The V4 series also marks a significant shift in DeepSeek’s hardware strategy. The models are optimized for inference on Chinese chipmaker Huawei's Ascend supernode. Earlier models such as V3 were trained on Nvidia H800 chips. V4-Flash is reported to be partially trained on Huawei hardware, while V4-Pro still relied on Nvidia due to its extensive compute requirements.

The partnership benefits both companies, according to Lian Jye Su, an analyst at Omdia, a division of Informa TechTarget. Huawei has struggled to gain traction with its own Pangu series of models.

"Being able to support DeepSeek now natively, showcasing pretty reasonable performance, is a very advanced achievement," Su said. "It does allow China to gain a bit more respect from other vendors."

Su added that China-friendly countries may become more open to adopting other Chinese products, given that China appears to be overcoming restrictions and limitations compared to 12 months ago. However, he noted that Huawei still lags behind Nvidia and the broader global AI ecosystem in chip fabrication and software development.

The Commitment to Open Source

DeepSeek’s adherence to open source is also significant geopolitically. While U.S. vendors such as Meta and OpenAI previously embraced open source, they have largely moved away from it. Open source could serve as DeepSeek’s primary differentiator.

"Open source helps them attract developers, build trust and create an ecosystem," Kompella said, adding that it is also advantageous in the Chinese market.

Other Chinese vendors, including Alibaba, Kimi, Qwen and Minimax, also offer open-weight models. Su described open source as fundamentally a commercial strategy but said it is embedded in DeepSeek’s approach rather than being a temporary tactic before shifting to closed models.

Open source also gives China an opening in price-sensitive markets outside the U.S.

"It also lets China influence global AI markets without needing to own every application layer," Kompella said.

DeepSeek’s gradual move away from Nvidia aligns with the Chinese government’s broader aim of reducing dependence on Western vendors. However, Western enterprises remain cautious.

"There are now significant reservations in the Western camp about adopting any open source solution from China," Su said. "There is a large pushback from large enterprises, especially those in critical industries, not choosing or avoiding Chinese models entirely, mainly because of scrutiny from governments."

Still, for enterprises already using Huawei products, V4 models may warrant consideration. For others, DeepSeek’s combination of low cost and open-source strategy could be influential in shaping the next phase of the global AI competition.

"The global AI race is about who can deliver intelligence at scale, at low cost, on a sovereign technology stack," Kompella said.