DeepSeek & Destroy

Analyzing the 1st, 2nd and 3rd order of Cheap + Powerful + Open Source AI

If you have somehow missed it, a previously mostly unknown quant hedge fund in China released DeepSeek, an AI performing as well as, if not better than the most advanced US AIs.

So I interrupt the planned program for a dedicated report on the topic.

It creates a massive set of problems for Big Tech and especially Nvidia, as well as now for-profit OpenAI.

The time could also not have been worse, as it came the day after Trump announced Project Stargate, a plan to spend burn $500B in building massive AI data centers in the US, under the guidance and financing of Oracle, OpenAI and Softbank.

So Many Problems

Too Efficient

Problem #1 for AI, is that DeepSeek requires MUCH less computing power. So little in fact that people are reporting managing to run this AI from high-end desktop computers.

Other research claims to be able to retrain the AI for much less than the already very low $6M DeepSeek claim to have spent. Maybe as low as … $30 (no zero forgotten).

So this creates the #1 first order consequence for all of Big Tech, which have burned HUNDREDS of billions in building computing power that might not be needed after all.

Already, the mere normal depreciation of these assets was going to hurt the profit of these companies. But if they need to write it off entirely, it will REALLY hurt.

Open Source

Now the dastardly Chinese did something even worse than improving AIs. They released the code for FREE, open source. That is problem #2.

This means that despite tentatives to scream it is an evil CCP operation to spy or steal data, it is clear that very soon there will be dozens of “forks” of DeepSeek, retrained to fit specific needs or biases of specific users.

Not only is it now out there for free, but it is also already virtually impossible to destroy or ban efficiently.

This is a death blow to all the closed-source methods used by OpenAI, Google, Meta, as well as Alibaba, ByteDance (TikTok maker), etc.

So Cheap!

Problem #3 is that DeepSeek makes its money by trading, like a good hedge fund. So it is planning to commercialize its AI essentially just at costs, for the sake of demonstrating the superioty of its talent, which will no doubt bring billions in assets to be managed by the hedge fund.

And as these costs are much lower than ANYONE in the industry has managed until now, it utterly destroys the planned business model of all AI companies.

Imagine we’re back in 2017 and the iPhone X was just released. It was selling for $999 and Apple was crushing sales and building a wide moat around its ecosystem.

Now imagine, just days later, another company introduced a phone and platform that was equal in every way if not better and the price was just $30.

That’s what unfolded in the AI space today. China’s DeepSeek released an open-source model that works on par with OpenAI’s latest models but costs a tiny fraction to operate. Moreover, you can even download it and run it free (or at the cost of your electricity) for yourself.

Not Even Sanctioned!

Apparently, DeepSeek trained its AI with older, less performant H100 Nvidia hardware, and not even that many of them.

Some claim that this is not true, and that DeepSeek had access to “illegal” (as in “not approved by the US interferences”) smuggled more advanced chips.

In any case, it already does not matter, because DeepSeek has now tested if they could use Huawei chips instead, and it turns out it is pretty much he same as Nividia chips…

This is a big problem, #4, for Nividia, obviously, as it might lose the entire Chinese market. Which when taking into account “illegal” re-exports and other by-passes (like Chinese companies using data center compute power abroad), might be 30-40% of its sales.

But this is also a massive problem for the American government.

Biden and now Trump have bet the house that hindering China's progress in AI would give the USA the advantage in an upcoming conflict against China, through drone swarms, cyberattacks, broken encryption, etc.

Now it appears that Huawei chips will do the job and that Chinese AI is superior… Oops…

2nd order Consequences

Except if you are a trader that reacted quick enough to grab the -17% in one day on Nvidia stocks, money is usually made from correctly predicting 2nd and 3rd order consequences.

Like betting rightly that remote work will explode due to a pandemic.

So let’s discuss it.

Sanctions Failed? MORE SANCTIONS!

An obvious second-level effect is that the US will panic and triple down on sanctions, in a reflex.

Already, DeepSeek is “banned” in Italy (good luck with that). And many AI big shots are trying to make the case that more needs to be done to break Chinese AI companies.

This will likely fail, and reveal even further the true impotence of the US in blocking China's rise.

AI Generalization