In recent days, markets were unexpectedly shaken by the emergence of DeepSeek, a Chinese AI startup founded by former hedge fund manager, Liang Wenfeng.
The large language model (LLM) the firm offers has challenged behemoth US firms like OpenAI and Google by developing or ‘training’ itself for a fraction of the cost. At just $5.6 million it is 20 times cheaper than its competitors.
JP Morgan has addressed how this feat was achieved: “How did they manage this? Well for starters, they reinvented the whole approach. Traditional AI is like painstakingly writing out every number with 32 decimal places, but DeepSeek thought, "What if we just used 8? It's still good enough!". This slashed memory requirements by 75%. Plus, regular AI reads like a four-year-old – "The... cat... sat..." – while DeepSeek processes full phrases in one go, making it twice as fast with 90% of the accuracy. When you're working with billions of words, this makes a big difference.” (Source: J.P. Morgan)
They also created an 'expert system', where instead of relying on one all-knowing AI, it uses specialised experts that activate only when needed. Traditional models keep all 1.8 trillion parameters running at once, while DeepSeek has 671 billion in total, but only 37 billion are active at any given moment. It’s like having a huge team, but only calling in the experts required for the job at hand (Source: J.P. Morgan).
DeepSeek’s efficiency sent shockwaves through financial markets on Monday, particularly in the semiconductor and AI infrastructure industries. Nvidia, the world’s leading supplier of AI chips, suffered the largest-ever market capitalisation loss in history with circa $600 billion being wiped out in a single day, a drop of 17%. For context, this is just a little less than the total market capitalisation of Italy. The concern? DeepSeek’s ability to train on older, less powerful chips. If AI models can run efficiently on lower-cost hardware, Nvidia’s pricing power could be under threat.
The ripple effects extended beyond Nvidia – Constellation Energy, a major supplier to AI-driven data centres, saw its stock decline 21%, as the market reassessed AI's future energy consumption needs. This is good news for the planet. Traditional AI approaches use an incredible amount of energy with some estimates suggesting that each query on OpenAI’s ChatGPT engine uses about the same amount of energy as running a standard lightbulb for 20 minutes.
Hyperscalers like Amazon, Microsoft, and Google, who have invested heavily in AI data centres, faced pressure, though they remain better insulated due to their diversified revenue streams.
Despite this conundrum, the S&P 500 ended the day just 1.3% down, as sectors like healthcare, industrials and financials showed strength, suggesting a potential broadening of markets beyond the traditional tech giants. We also saw a small rally in bonds, with the 10-year US Treasury yield dipping to 4.5% as investors sought safe havens.
This reflects the fact that whilst in the short-term less demand for chips could be seen as bad for Nvidia and other parts of the AI infrastructure complex, in the long-term lower costs could significantly broaden the market for AI. This could ultimately help everyone.
It could also have positive effects across the whole global economy. If the benefits of AI were only to be enjoyed by those with the deepest pockets because of the vast capital investment needed to power it then we could worry all of the fruits would be confined to the United States and China. If the barriers to entry are lower, the benefits can diffuse across the world economy and smaller developed economies like the UK could arguably gain more benefit. In this sense it has the same effect as anti-trust action which breaks up monopolies – it diffuses profitability across the economy.
Following the major stock market moves OpenAI raised concerns that DeepSeek may have trained its models using unauthorised GPT-4 outputs, a practice known as distillation, which involves querying an existing AI model and using its responses to train a new one – potentially violating OpenAI’s terms of service. DeepSeek denies any wrongdoing, but OpenAI and its partner Microsoft had already blocked their access to OpenAI’s API last year over similar suspicions.
When answering questions about VW car sales in China, ChatGPT, Grok, and Gemini all gave very different answers, while DeepSeek’s response was almost identical to ChatGPT. Formatting is another highly identifiable LLM footprint – when asked to program an impossible graphics function, DeepSeek’s answer was 95% similar to ChatGPT’s but very different from what Co-Pilot, Grok, and Gemini produced (Source: J.P. Morgan).
This brings up critical questions about intellectual property rights in AI – if companies can freely extract knowledge from existing models, closed-source AI firms, like OpenAI, Anthropic, and Cohere, could struggle to monetise their proprietary systems. Meanwhile, the rise of DeepSeek strengthens the case for open-source AI models, which could challenge the dominance of closed-source models.
The biggest question now is whether DeepSeek’s efficiency gains represent a one-time breakthrough or a sustained shift in AI development. For now, DeepSeek has proven that AI discoveries don’t always require hundred billion-dollar budgets. Sometimes, all it takes is a brilliantly engineered approach and a willingness to re-think everything from the ground up. That said, it’s important to recognise that the core strengths of established players remain intact, and while DeepSeek’s entry highlights the potential for innovation outside of traditional models, it’s still too early to gauge the long-term impact.
As always, we are closely monitoring these developments and will respond accordingly, based on a variety of market signals.