A surprisingly capable model revealed several days ago in China, DeepSeek's R1, whose capabilities match those of OpenAI's reasoning "o1" sent a shockwave through the American tech industry and the stock market, briefly wiping out over a trillion dollars in market capitalisation, including nearly $600 billion from Nvidia alone.
The reason for the frenzy is that R1 is many times cheaper to use and released as open source, which means it can be downloaded by anybody and run on your own hardware, instead of relying on a 3rd party service provided by e.g. OpenAI or Google.
If major breakthroughs such as this are possible with much less hardware investment it would dramatically alter the economics of AI development both speeding it up and making it A LOT cheaper.
This would be a truly seminal moment for artificial intelligence – and it is but, sadly, not for these reasons.
If something is too good to be true...
...then it usually is.
And it's no different in this case. As I have already suggested in my commentary for the Vulcan Post there is indication that R1 isn't all that it was made to be.
First of all, DeepSeek is said to have a massively greater number of advanced Nvidia cards than it ever revealed that it does – a fact likely kept secret not to reveal the scale of possible violation of American export restrictions on advanced semiconductor products used in the process of training and operating such a service.
Instead of measly 2,000 of export versions of H800 it could have as many as 10,000 of the more advanced H100 cards – still not a huge number but considerably larger and more expensive.
As SemiAnalysis calculated the total capital expenditures of the ambitious Chinese startup could exceed US$1.6 billion by now – much higher than the paltry $5.6 million in mere training costs circulated in the media, which do not show the full picture.
This means that the idea that small startups with a few million bucks on hand can make a major impact is still unrealistic, contrary to what some have excitedly suggested.
In this case "small" is still in the billions.
Most importantly, however, there's mounting evidence that R1, quite like with the earlier V3, is just a distilled version of OpenAI models, the outputs of which appear to have been used in the process of training the Chinese competitor, up to the point it was able to produce o1-level of responses.
Knowledge distillation is a process in which a larger model acts as a teacher of sorts, having first done all the heavy lifting, for a much smaller, lighter model, which can eventually match its quality with far fewer parameters.
As a result it is considerably cheaper to operate, while maintaining high accuracy.
Typically, however, that's something done by the companies refining their own models before public release to consumers. Here we have an example of a competitor essentially leeching on the work done by another company – i.e. OpenAI.
The age of AI piracy is here
It doesn't take long to realise what it means for AI development. In essence, the original creators of the model, who invested billions of dollars that they are now trying to recoup, are priced out of the market by someone who hijacked their work and is selling it at a fraction of the cost – or even for free, masquerading as "open source".
It's like cracking a video game and putting it on a torrent site for everyone to download it for free (only worse, because here the cheat is even hailed as a genius).
Without billions spent by OpenAI there could be no DeepSeek.
And without even more billions there will be no more advanced models, because nobody is going to pour exorbitant sums of money into their development if a company half the world away can leech on it and sell it as its own for peanuts.
If the innovators do not get their monetary reward, proportional to the investment they made, there will be no more innovation. Unless ROI is positive, there can be no "I".
Yes, the R1 debut is a major event in the world of AI development, but not because it's making it cheaper and faster, but because it's going to result in heightened security and more secrecy.
It may, in practice, slow the product release schedule down, instead of speeding it up, since it may make more sense to keep your efforts under wraps until you have a far more advanced product than your rivals and you know how to protect it.
National security
This isn't just about cutthroat competition and billions of dollars at stake either. Several rounds of export restrictions on AI chips imposed by the US government were meant to keep its biggest rival, China, at a distance.
As it turns out, however, the hardware – while useful – is not necessary if you can rely on the outputs alone to train equally capable models even under technological constraints.
China may not be able to outcompete the US without access to millions of chips but it can stay level with it by learning from the publicly available models, unless a solution preventing that is devised.
And that solution may very well be, again, more secrecy and slower product development (at least for the end users), since it's no longer about producing an even more capable bot to write your reports for you but AI capable of boosting entire industries (and the military).
Many in Washington may deem it prudent to sacrifice the consumer in order to keep real AI progress away from the spotlight.
And they might be willing to pay top dollar to make it so if there is no other way to stop China from doing what it does best – copying others.