Seek or Swim? Nvidia swims onwards

Nvidia made $40 billion last quarter despite the threat from emergent Deepseek.

Fears of China’s emerging Deepseek AI have failed to take root. Nvidia’s AI chips, utilising its ‘Blackwell’ architecture, have seen a surge of demand as tech companies turn to it for AI training and development.

An announcement by Deepseek in late January spooked the market, prompting a 17% slump erasing nearly $600 billion in market value. It has since managed to somewhat dispel industry fears, reclaiming its place in the $3 trillion club.

In January, Deepseek revealed that it had trained its chatbot by using “less expensive, more primitive chips”, promoting uncertainty amongst investors of the relatively advanced Blackwell architecture. In response, Nvidia CEO Jensen Huang said he held few concerns, adding that “fundamentally software changed”, and predicting that future software will be created with machine-coding chips rather than traditional “hand-coding”. 

Its latest quarterly results are reassuring - upon the fourth financial quarter of 2025, revenues have reflected a 75% year-on-year increase as companies continue to adopt Blackwell infrastructure. Nvidia’s strategy of utilising costly, yet high-performing hardware appears to have achieved enduring success. 

For they are indeed more expensive - Huang has stated that they cost between $30,000-$40,000 per unit. This steep price is justified not only by superior computational power, but also by their cohesion within the Nvidia ecosystem. Nvidia’s industry-standard software stack, including its CUDA developer tools and frameworks such as TensorRT, work seamlessly with the Blackwell architecture. 

Its outmatched computational power has raised concerns of over-reduced profit margins, however. Nvidia’s Blackwell chips’ complexity has led to a decrease in gross profit margin from 76% to 73%. This weakness is exacerbated by Deepseek’s divergent cost-effective strategy, which emphasises “frugal innovation”.

Deepseek has been able to produce high-performance AI models at prices approximately 20 to 40 times cheaper than competitors. It reportedly trained its landmark V3 model for around $6 million - compared to OpenAI which spent $100 million on GPT-4 in 2023. Deepseek’s R1 model production strategy focused on innovative software, circumventing the costs associated with training via high-end hardware.

Liang Wenfeng, founder of Deepseek, has emphasised this research-led approach. As a result, he has garnered the endorsement of President Xi Jinping, who may view the emergent AI as a robust counter to Western dominance of the AI industry. 

Newer competitors are emerging fast, with the Deepseek scare back in January potentially foreshadowing what is to come. Gross profit margin projections for the next quarter have dropped to below 71%, as Nvidia comes under pressure from its pricing strategy and heightened production costs. This is all before they consider the possibility of US Trump-led tariffs and the introduction of further AI regulations.

With these geopolitical threats in mind, Nvidia may have to consider its profit margin weaknesses. Their upcoming performance will define whether the future AI training industry is dominated by hardware-centric, high-performance powerhouses or China’s burgeoning cost-efficient, software-led counter solutions.