What is going on with Deepseek?

The release of this model by China has really shook the AI world. Many believe it is bursting the "AI bubble". It sent stocks like Nvidida flying south, with the largest one day eradication in history.

Is this the end for Big Tech? It is a view many are proclaiming. The days of tens of billions being spent on GPUs are over.

To this I say, hold your horses. It is not as clear cut as it seems.

For this reason, let us dive into it.

https://img.republicworld.com/all_images/deepseek-1737964174034-16_9.webp

Source

Is Deepseek For Real?

The hysteria stems from the claim that Deepseek was trained for $6 million. This is a drop in the bucket compared to what the likes of Meta, Google, and xAI spend. Does this mean their approach is cooked?

Not at all.

To start, it is highly unlikely that Deepseek was trained for $6 million. Perhaps that was the case for the last version. However, there is a super cluster of Nvidia chips being used. Of course, this cannot be stated as fact since there is a US embargo against that.

However, 20% of Nvidia's sales are to Singapore. Do you think that 20% of all the GPUs Nvidia creates are in that country? If you do, I have a bridge to sell you.

It is not a tightly kept secret that a tremendous black market exists for these chips. I would not be surprised if Deepseek used Hopper chips. That said, it is reported that there were A200s involved.

For this reason, I call horsehockey on the $6 million assertion.

Does this mean the model is not superior? I am not going to state that. From what the testers are saying, it is performing high levels against the benchmarks. From a model standpoint, there is certainly validity.

We also do see the capability to run it on smaller machines. This is a positive. The question is whether this is totally surprising?

It is not.

Technology Improves And Advances

Actually, there was going to be a day where innovation led to models requiring less compute. We cannot have all the inference only on massive data centes.

If we simply look at the progression of ChatGPT, the earlier models required a great deal more compute to operate than today. In fact, if we look at the power of a smartphone compared to a 1980s personal computer, we can see the difference. At that time, images were not even possible for processing. Today, even the cheapest device can store and process photos.

Deepseek made headlines due to the fact it is open source. This is a positive in my view since I think this is the path AI needs to take. It will likely spread Deepseek far and wide.

Of course, when it comes to AI, if you do not like where things are, wait about 3 weeks. It is probably that someone else will be the "talk of the town" by the end of February. Also, when it comes to open source, at some point Meta will have another release, garnering the headlines.

Billions Lost

As for the destruction of market cap, anyone who is basing their buying of stock on LLMs is foolish to begin with.

I have long stated these are going to be commodities. Generative AI is general purpose technology (that is the GPT). This means it is akin to electricity.

Does electricity have value? Sure. Utility companies are worth a pretty penny. However, how does that compare to technology companies, which require large amounts of electricity to run their operations. Utilities might be worth billions whereas Big Tech is in the trillions.

The future of AI economic productivity is not coming from LLMs. Rather, it will be what is built upon them and where they are incorporated. Is the electricity at one utility company different from another? Obviously not.

That said, the use of electirivity and what is generates in output will vary among customers.

In other words, how are these models utilized and what do people (developers, entrepreneurs) do with them?

That is where the real money is.

Deepseek is another step in the process. We are in the very early stages. Those companies with larges amount of compute, if there is a new structure for training, will only see their total output increase.

Posted Using INLEO