DeepSeek-V4-Pro Aims to Undercut Big Tech, Foster China’s Independent AI Hardware Industry

DeepSeek’s latest large language model, DeepSeek-V4-Pro, has been released with a significant price reduction, a move that observers suggest is more than just a competitive pricing strategy. The model’s affordability, coupled with its architectural innovations, appears to be a deliberate step towards fostering an independent Chinese AI hardware industry, less reliant on US-dominated technologies like Nvidia’s GPUs and CUDA.

While DeepSeek-V4-Pro may not match the absolute performance of frontier models from OpenAI, Anthropic, or Google, its capabilities are described as “very decent.” The critical differentiator lies in its cost-effectiveness, with a permanent 75% discount on its promotional pricing making it substantially cheaper for high-token usage tasks. This efficiency is key to its broader strategic goal: enabling China to advance its AI development despite US sanctions that restrict access to leading-edge chips and manufacturing processes.

Key facts

• Input Cost: $0.435 | $5 | $5 | $1.5
• Output Cost: $0.87 | $30 | $25 | $9
• Performance: Good | Exceptional | Exceptional | Very Good
• Strategic Goal: Hardware independence | Market leadership | Market leadership | Market leadership

What has happened

DeepSeek announced last Friday that its 75% promotional price cut for DeepSeek-V4-Pro would be permanent. This pricing makes the model exceptionally affordable, costing $0.435 per million input tokens and $0.87 per million output tokens. This starkly contrasts with competitors like GPT-5.5 ($5 input, $30 output), Anthropic’s Opus 4.7 ($5 input, $25 output), and Google’s Gemini 3.5 Flash ($1.5 input, $9 output), according to data from Artificial Analysis.

While DeepSeek-V4-Pro’s performance is noted as being lower than the top-tier models, Artificial Analysis benchmarks indicate it is competitive. The true value, however, is its accessibility for token-intensive applications, such as agentic workflows, which become economically viable with this model.

A different strategy

DeepSeek’s business model deviates from its competitors. It does not offer subscription plans like ChatGPT Plus or Claude Pro, nor does it focus on voice or image models, or specialized coding agents. Instead, the company releases its model weights openly and shares its technical innovations broadly, even with potential competitors. This approach suggests a long-term vision focused not on dominating the LLM race itself, but on building a robust Chinese AI hardware industry that can operate independently of US influence.

Hardware independence and sanctions

China faces a significant structural challenge in the global AI race due to US sanctions that limit its access to advanced semiconductor manufacturing and high-end chips. The strategy adopted by Chinese AI firms, including DeepSeek, is to develop models that require less computational power to achieve comparable results. This approach aims to circumvent the limitations imposed by the lack of access to cutting-edge hardware.

Efficient architectures

Two key architectural innovations driving this strategy are Mixture of Experts (MoE) and Multi-head Latent Attention (MLA). MoE, an existing architecture, has been adapted by DeepSeek to activate only a subset of the model’s parameters for each query, thereby maintaining precision while reducing computational load. MLA significantly compresses the Key-Value (KV) cache, which stores conversational context, by up to 90%. These techniques collectively reduce the dependency on high-speed High Bandwidth Memory (HBM), a component often manufactured by companies like SK Hynix, Samsung, and Micron, which are part of the US-aligned supply chain.

The importance of KV Cache reduction

The efficiency gains from MLA are substantial. According to analyst GDP on X, DeepSeek-V4-Pro requires only 5.48 GB of HBM for one million tokens. This is a dramatic reduction compared to competitors like Zhipo AI’s GLM 5, which needs 60 GB, and Alibaba’s Qwen 3, requiring 89 GB for the same task. This efficiency allows DeepSeek to offer lower prices and, crucially, enables its models to run on less advanced, domestically produced Chinese memory chips that cannot compete with HBM in raw speed.

Beyond HBM: NAND and SSD integration

These advancements open the door to utilizing NAND flash memory and even Solid State Drives (SSDs) for data processing. This is where Chinese memory manufacturers like YMTC (flash memory) and CXMT (DRAM) become critical players. DeepSeek’s development of Engram, a memory retrieval module for LLMs, further aims to reduce reliance on HBM, positioning these domestic memory solutions as viable alternatives.

Bypassing CUDA’s monopoly

Nvidia’s dominance in the AI market is heavily reinforced by its CUDA parallel computing platform. DeepSeek has proposed an alternative with Tile Kernels, a software layer built using TileLang (a Python variant for AI hardware control). This solution aims to manage advanced AI chips, including GPUs, without direct reliance on Nvidia’s proprietary ecosystem.

Huawei as a strategic ally

Huawei has publicly stated that its new Ascend AI supernodes fully support DeepSeek V4 models. This partnership is pivotal, as it provides DeepSeek with a viable domestic hardware platform, reducing its dependence on Nvidia GPUs and reinforcing Huawei’s position in the AI hardware market. This collaboration could accelerate the adoption of DeepSeek’s efficient architectures by Chinese hardware manufacturers.

Open models to drive hardware adoption

Unlike many US AI companies that maintain proprietary models, DeepSeek, along with other Chinese startups, publishes its model weights openly. This strategy aims to attract developers and users, thereby fostering an ecosystem that encourages hardware manufacturers to adopt and optimize for these efficient architectures. By promoting techniques like MoE and MLA, DeepSeek seeks to establish them as de facto standards, encouraging broad hardware integration and development.

A significant funding round

DeepSeek is reportedly preparing to raise $10 billion in a funding round that would value the company between $45 billion and $50 billion. While still below the valuations of industry giants like OpenAI and Anthropic, this represents a substantial investment for a Chinese startup. The capital is expected to fuel further development and solidify its strategy of building an independent AI hardware industry.

Source: DeepSeek is good, nice, and very cheap. And above all, the weapon to create a Chinese hardware industry independent of Nvidia, Xataka, by Javier Pastor.