Almost entire LinkedIn has written their opinions about DeepSeek 🙂 So, instead of writing a full blog, I will stick to analyzing some of the myths of “the sky is falling” opinions that I saw in the last few days.
They came out of the left field and surprised everyone.
No, they didn’t. While DeepSeek-R1 is the new announcement, they have been at it for a while. Developing such massive AI models doesn’t happen in a vacuum overnight. They released DeepSeek coder way back in November 2023 (open source model for coding), DeepSeek LLM also in November 2023 (67B parameter LLM), DeepSeek v2 in May 2024, DeepSeek v3 and DeepSeek-Coder-V2 in December 2024, DeepSeek R1 preview in November 2024, DeepSeek-R1 and DeepSeek-R1-Zero about a week ago. What shocked the world was the DeepSeek-R1 Capabilities and performance chart that they just released. They claim to beat every other LLM in the market including the mighty GPT-4o. Besides these new models were released a week ago, and the world is suddenly waking up to it today – a week later? Is it possible they decided to create disruptive messaging after the new administration has been sworn in?
They did it all by themselves.
Not true. They have strategic relationships with NVIDIA, and AMD, among others. The combination of cheaper AMD Instinct GPUs and open-source ROCM software allowed them to keep the costs very low. While their claim of “only $6 million” needs to be validated, this could be a killer to many AI providers from the network all the way to the app/software level. If this is proven, almost everyone will take a hit. Chips, networking, data center components (cooling, wiring, fiber optics, security, etc.), software providers, and everyone’s pricing will be scrutinized and open-source alternatives will be heavily considered. AI supply chain known to use open-source software for a while now. However, if it is proven that the entire AI software supply chain can be done cheaply using open-source software, many startups will take a hit. VCs will stop writing blank checks to start-ups that have Generative AI on their pitch deck and demand proof for their sky-high valuation.
AI supply chain stocks are doomed.
In the short term, absolutely. To begin with, many of these companies are way overvalued, especially with multiple AI good news in the last few months. They were bound to have a pullback regardless. This event triggered it. The market always overreacts, especially to negative news. I don’t think the entire AI supply chain will be affected. Maybe portions of it such as Chips, Software, etc. might get affected. Even if you build cheaper and more efficient data centers, there will be demand for networking, and data center components with almost every country announcing massive data centers. The other parts of the ecosystem will remain intact. As there is not much competition and they can control the price.
NVIDIA as a growth stock is done.
I doubt that. Jensen has been an amazing innovator and storyteller. He transformed industries by using his GPUs which were invented for graphics and gaming purposes initially. He spurred growth in crypto, and AI among others. He will figure out a way. On a short-term basis, AMD could pose some threat if their Instinct GPUs can prove the capability. BTW, the Blackwell chips are sold out. They are taking orders for one year out now. But if AMD can truly deliver that performance there will be a massive shift in chip power. A lot needs to be proven with many companies to come out and suggest they were able to train their LLMs using AMD chips. There are also reports that DeepSeek used just 2,048 Nvidia H800s and $5.6mn to train a model with 671bn parameters, a fraction of what OpenAI and Google spent to train comparably sized models. I see this as more damaging to NVIDIA than anything else.
Their model and usage are open source.
Yes, but…. The actual pre-trained models, source code, model weights, parameters, architecture, training scripts, and training methods themselves are fully open. All are licensed under the MIT license model and anyone can modify or distill the models or use them freely for commercial purposes. But the real question here is transparency. The data set on how it was collected, what it was trained on, what is the bias removal process, what kind of censorship is done on the outputs, etc., is more important. Granted OpenAI is also opaque and doesn’t disclose many of these details. When choosing a model transparency, the model creation process, and auditability should be more important than just the cost of usage. Right now, it is hard to trust any AI models produced by a Chinese company. Alibaba has GWEN out for a while, but it hasn’t gained any meaningful traction in the Western world. TikTok’s parent company ByteDance also released their Doubao-1.5 pro and its performance matches OpenAI’s non-reasoning GPT-4o model on third-party benchmarks and costs only 1/50th of OpenAI’s model.
US dominance as an AI leader will end.
Most probably not. However, this has proven to other countries that they, too, can produce AI models with less US company dependency. This should accelerate companies like India, which soon has the manpower, knowledge, capital, and potential data center capabilities, to take such initiatives to reduce dependency on US companies. While the US is a costly option, most countries are still hesitant to use Chinese-based companies. For example, though Alibaba QWEN models performed comparable to the advanced US models, their adoption has been very slow.
It will cost a lot less to build LLMs and LAMs.
Maybe. Their claim of just $6 million is still not validated. Assuming it is true, this will have a major impact on the AI supply chain. There will be a major squeeze on prices as consumers will start to demand better pricing showing this as an example. We really don’t know how much the Chinese government or agencies funded this project under the covers. It will be hard to estimate until then.
Open-Source vs. Closed-Source AI: A Growing Tension
Until now, except for Llama (which is not fully open source) and Allen Institute, many open-source LLMs have not performed as well as closed-source LLMs. DeepSeek uproots this. To some extent, this can shift the power more toward open-source models.
Overall, we are having a knee-jerk reaction right now. But it has proven a point that you can train your own models at a fraction of the cost that the major players are doing. The bigger players are already feeling the pressure to make money to cover the costs. OpenAI increased the price to $200/month from $20/month, and others (Google, and Microsoft) are starting to embed their options into their offerings and forcing their entire base to start using it instead of opt-in mode.
But this brings up an interesting question. If DeepSeek, somehow navigated the semiconductor restrictions that the US has imposed on China, cutting the country off from the most powerful AI chips, then how will this game play out in the future? Were they able to circumvent export controls or found a way to work with weaker chips?
Regardless of what comes out of this, when choosing an AI model, consider the following:
License: Understand the terms and conditions of different open-source licenses is essential for compliance and to avoid potential legal issues.
Data Privacy: Ensure that the use of LLMs complies with data privacy regulations is crucial, especially when handling sensitive information.
Copyright: Address copyright concerns related to the training data and the generated output of LLMs is important to avoid infringement.
