Come on, let’s put another piece in the AI machine by talking about DeepSeek again. At the time of its launch, the figure of $6 million in development costs was put forward… But we’re going to come back on that, because the cost is much higher than that.
DeepSeek cost more than $6 million!
No, the AI didn’t just cost $6 million to develop… In fact, it turns out that the Chinese AI cost a lot more. However, marketing has decided to highlight the low cost, which is linked to the cost of GPUs for the pre-training phase. As SemiAnalysis points out, choosing to highlight this cost is like “pointing to a specific item on a list of components and considering it as the total cost”.
In reality, the total cost of DeepSeek V3 would easily exceed one billion dollars. It is estimated that the hardware cost already exceeds $500 million. Then there are the costs of research, staff salaries, development and so on. All in all, we’re a long way from the $6 million we’re talking about.
Nevertheless, we must give credit where credit is due, as the AI is proving impressive in spite of everything. In terms of performance, V3 of the Chinese AI outperforms GPT-4o. But AI is a constantly evolving sector and these advances in performance were ultimately expected. So it comes as no surprise that V3 will outperform GPT-4o, which will be released in May 2024… And the next version of GPT will be even more powerful, and so on.
SemiAnalysis points out that this is the aim of AI research: to increase the number of calculations performed with the same hardware capacity. The aim is to get as much ‘intelligence’ as possible for every dollar spent. It is estimated that every year that passes, algorithmic progress increases by a factor of 4. So, to offer an AI equivalent to GPT-3 in terms of quality, inference costs have been divided by 1200. We can now have AI running locally on a laptop where previously we needed a supercomputer.
Finally, while Chinese AI may have caused a sensation with its $6 million figure, the reality is much more complex, but it does illustrate a certain dynamic where performance is increasing while costs are decreasing. What’s more, this model is putting pressure on the artificial intelligence market, since when it came out, the share prices of major companies plummeted, with NVIDIA’s stock market valuation falling by $589 billion. Be that as it may, we can look forward to a bright future for AI.