Why is China's LLM DeepSeek R1 Causing Havoc?
If you've been following the news lately, you've likely encountered China's response to ChatGPT, called DeepSeek, on every channel and social media platform. Since the launch of ChatGPT, there have been many competing products—Claude AI, Mistral, Llama, to name a few—which were seen as nearly as good replacements. Why is DeepSeek such a big deal, and why are U.S. company valuations being hit hard by this one Large Language Model?
Here are a few highlights and their sources:
- Nvidia experienced a dramatic plunge—its shares dropped roughly 17–18% in a single day, wiping out nearly $600 billion of market value.reuters.com
- Broadcom, another U.S.-based semiconductor firm, also saw its stock fall by a similar percentage.reuters.com
- Oracle shares tumbled by around 13.8%, reflecting investor concerns over the need for high-cost AI infrastructure.reuters.com
- Microsoft and Alphabet (Google’s parent) were not spared either, with Microsoft dipping by roughly 2–2.5% and Alphabet by about 4–4.2%.reuters.com
This trickled down to power producers who had also secured massive contracts to meet the demand of AI data centers, which consume an exorbitant amount of power.
A while ago, we spoke about the U.S. envisaged Universal Basic Income, where big tech would be levied a special tax redirected at providing every American citizen $2,000 a month. The rationale is that AI will eventually replace jobs, and citizens should be compensated perpetually to live a meaningful life.
Interestingly enough, the U.S. forgot about the rest of the world, who, through making use of products developed in the U.S., would effectively be paying directly to feed their population.
The key to answering why U.S. markets are panicking are:
- Product Originates from China: Due to U.S. sanctions, China has been restricted from purchasing Nvidia's advanced H100 AI GPUs and instead has access to the less powerful H800 models. To mitigate this limitation, Chinese companies have turned to domestically produced alternatives, such as Huawei's Ascend 910C chips, for AI inference tasks. In simple terms, this approach involves training a smaller AI model to replicate the behavior of a larger, more advanced model by having the smaller model learn from the outputs of the larger one.
- DeepSeek R1 Benchmarks: DeepSeek is able to match and, in some cases, outperform OpenAI’s latest model o1. At the time of writing this article, OpenAI just released o3, which is a clear demonstration that the pressure is on.
- DeepSeek Shows Reasoning: This is a huge one, as you are now able to trace the thought process being applied by the Language Model and can pick up mistakes in reasoning, which assists you in improving your questions when prompting LLM.
- DeepSeek Uses Significantly Fewer Resources and is Cheaper: The model was trained in a record-breaking 55 days on just 2,048 Nvidia H800 AI GPUs, with an estimated cost of $5.5 million. This is actually less than 10% of the cost of training ChatGPT.theguardian.com
- Can be Deployed Locally and is Open Source: Strangely enough, we are finding DeepSeek integrated everywhere—Hugging Face (a cloud repository of LLMs), Microsoft Azure Foundry, Perplexity AI, private clouds, and even in home labs.
My thoughts are this is good news for the rest of us. AI is finally being democratized, and I sleep better at night knowing the universe is in balance and we are able to switch channels to the East.
We all one day deserve to earn a universal basic income and I am looking forward to seeing innovation from 2nd and 3rd world countries as they carve a future for their people.
For a more technical breakdown, here are a few resources: