DeepSeek vs OpenAI
DeepSeek’s R1 model has snuck up on the AI establishment and emerged as a serious challenger to OpenAI’s dominance; achieving comparable performance at a fraction of the cost.
At 671 billion parameters the model is a significant technical achievement, but while OpenAI relies on massive computing power, DeepSeek has focused on efficiency through innovative architecture, including multi-head latent attention (MLA).
Silicon Valley startup co-founder Anthony Poo put it succinctly:
OpenAI’s model is the best in performance, but we also don’t want to pay for capacities we don’t need.
His company switched from Anthropic’s Claude to DeepSeek after tests showed similar performance at around a quarter of the cost.
Why DeepSeek Has Gained Ground
DeepSeek has managed to turn lemons into lemonade.
Restricted from accessing NVIDIA’s most advanced H100 chips due to U.S. export controls; the company optimised its operations with less powerful H800 chips, which transfer data at half the rate.
Their training process demonstrated remarkable efficiency: using less-powerful NVIDIA H800s, they trained on 14.8 trillion at a cost of $5.58 million – a lot less than the $100 million – $1 billion range cited by competitors. The model uses chain-of-thought (CoT) reasoning, which breaks down requests into sequential thought processes to reduce errors and hallucinations.
The business was founded by hedge fund manager Liang Wenfeng in 2023 and funded through his quantitative fund High Flyer, and brings valuable algorithmic trading experience to AI development.
Consequences for NVIDIA
DeepSeek’s achievements could significantly affect NVIDIA’s business model, which has relied heavily on selling high-end AI chips at premium prices.
Ironically, this stems from U.S. trade policies aimed at restricting Chinese access to advanced AI technology. When the U.S. imposed export controls preventing NVIDIA from selling its most powerful H100 chips to China, it inadvertently created conditions that forced Chinese companies to innovate with less powerful hardware.
Instead of having access to H100 chips, which can transfer data at 600 gigabits per second, Chinese companies like DeepSeek had to work with the export-approved H800s, which operate at half that speed.
This limitation led to breakthrough innovations in resource management and training efficiency. Sometimes, optimisation can sometimes trump raw computing power.
AI Stocks Tumble as DeepSeek Takes the lead
The impact rippled through markets, with Dutch chip equipment maker ASML falling over 10%, Siemens Energy plunging 21%, and AI behemoths like Nvidia, Microsoft, and Meta seeing stock declines.
Fiona Cincotta, Senior Market Analyst at City Index said:
This idea of a low-cost Chinese version hasn’t necessarily been forefront, so it’s taken the market a little bit by surprise.
So, if you suddenly get this low-cost AI model, then that’s going to raise concerns over the profits of rivals, particularly given the amount that they’ve already invested in more expensive AI infrastructure.
Future Market Impact
DeepSeek’s breakthrough extends beyond cost savings. Their open-source approach, including smaller distilled versions ranging from 1.5 billion to 70 billion parameters, could help to democratise AI development globally.
The model is available on Hugging Face under the MIT license, making it accessible to developers worldwide.
This shift could trigger several market changes:
- Increased competition in the AI model space as companies realise high performance is achievable with lower infrastructure investments
- Pressure on existing AI providers to reduce prices or improve efficiency
- A potential shift in AI development strategy from raw computing power to algorithmic efficiency
- Greater global participation in AI development, particularly from regions with limited access to cutting-edge hardware
Microsoft CEO Satya Nadella’s comment at the World Economic Forum underscores the significance: “We should take the developments out of China very, very seriously.”
As the AI landscape evolves, DeepSeek’s success demonstrates that innovation often thrives under constraints, potentially reshaping the future of AI development toward more efficient, accessible solutions.
