Later this week, we will be breaking down "What are the costs of integrating AI into an FMCG manufacturing company" and we will add the link to this blog post here.
Alexandr Wang, founder and CEO of Scale AI, brings an important perspective to this discussion, let's take some of Alexandr Wang’s comments on the situation. Scale AI, is an American AI company focussing on creating labelled datasets which are necessary for creating high quality AI models.
1: What is DeepSeek?
DeepSeek creates open source LLM’s with the latest being V3. As it is open source you can see it for yourself on Github: https://github.com/deepseek-ai/DeepSeek-V3.
And it has people talking too…

Whilst its performance is very competitive compared to the best public OpenAI models, there are some possible issues early users have raised:
Censorship: The developers might have deliberately restricted responses to sensitive topics, reflecting the political landscape of the country.
Dataset bias: If the training data excluded diverse perspectives on sensitive topics, the model might lack the ability to provide answers on sensitive topics.
Safety filters: Pre-set rules might prevent the model from addressing politically sensitive or controversial questions.
2: Why were NVIDIA shares down 16% yesterday?
The big news was that DeepSeek claims to have only spent $5,576,000 training the LLM using 2048 H800 GPUs which is a fraction of the cost OpenAI spent in its latest model. This would mean that the compute power needed to train a large LLM is roughly 11x less. A great breakthrough in the development of AI but this would reduce the number of NVIDIA GPU’s you need.
“During the pre-training stage, training DeepSeek-V3 on each trillion tokens requires only 180K H800 GPU hours, i.e., 3.7 days on our cluster with 2048 H800 GPUs. Consequently, our pre-training stage is completed in less than two months and costs 2664K GPU hours. Combined with 119K GPU hours for the context length extension and 5K GPU hours for post-training, DeepSeek-V3 costs only 2.788M GPU hours for its full training. Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576M.” View the research pdf here
3: Alexandr Wang
The H800 is a modified version of Nvidia's H100 GPU, designed for export to China. The H800 has a chip-to-chip data transfer rate of about 300 GBps, roughly half that of the H100's 600 GBps which are available in places like the United States of America.
Alexandr Wang had this to say in an interview with CNBC “My understanding is that DeepSeek has about 50,000 H100s, which they can't talk about, obviously, because it is against the export controls that the United States has put in place.” Alexandr Wang and Elon musk both seem to agree that you would be right to be sceptical about the number of GPU’s DeepSeek claim to have and use. This doesn’t help the integrity of the claims.
4: Should I migrate everything to DeepSeek V3?
Not exactly, over the last 2 years OpenAI has gained alot of experience with ChatGPT. They have been able to scale a viral web app and API with minimal downtime. With DeepSeek’s user numbers increasing rapidly its less likely to be as stable as OpenAI.
Pricing:
Model | Input Tokens | Output Tokens |
DeepSeek V3 | $0.27/million tokens | $1.10/million tokens |
OpenAI gpt-4o: | $2.5/million tokens | $5/million tokens |
But if you are using LLM’s at a large scale with non-sensitive data in a device which cannot run the model locally you could achieve almost a 75% cost reduction in API usage costs.
If an LLM being open source is necessary, the big main competitor would still be META’s LLama 3.1 405 Billion parameter model compared to DeepSeek V3 671 Billion parameter model.

If your AI implementation in your FMCG manufacturing business requires Computer Vision, reinforcement learning, simulation or anything that is not an LLM, the current DeepSeek models will not help as they are only LLMs. If you need a low cost LLM API this could definitely beat OpenAI’s cheapest models but there are security concerns.
What this change has shown is that talented AI engineers from across the globe can outperform OpenAI and that knowledge will likely transfer into better ML, Computer Vision, Simulation models which will consume less computer power for a better model that you can use in your FMCG manufacturing company.
If you’re ready to assess your company’s readiness for Digital Transformation with AI and ML, our team can help. We believe that the success behind technology projects is down to people not the technology itself.
Contact us to get started with our Smart Industry Readiness Index and take the first step toward transforming your operations.