back to article DeepSeek spills Big AI's open secret: Bright people with good ideas can beat billion dollar binges

It would take a heart of stone not to explode with joy at the massive infusion of schadenfreude provided in recent days by the DeepSeek AIpocalypse. Trillion-dollar markdown in tech stocks, slack-jawed panic at tech companies previously "too big to care" suddenly caring a whole lot, and the mainstream media unable to talk …

  1. Irongut Silver badge

    There is mounting evidence that Deepseek spent a lot more on GPUs and operating costs than the ~$5 million they are claiming.

    From what I've seen it is very persuasive evidence that some major slight of hand has happened.

    Plus the entire industry are liars, like their products, so not the world shattering breakthrough that is being claimed.

    1. m4r35n357 Silver badge

      _sleight_ of hand

      1. Jason Bloomberg Silver badge

        But how much more was spent, and does it come anywhere near what other players have spent, claim they need to spend?

        I don't know but it strikes me that the previous players are desperate to assert it couldn't have been done out of self-interest.

        I can't blame them in the face of their house of cards collapsing, their snake oil scam having been undermined, but there appears to be a lack of actual evidence so far.

    2. Anonymous Coward
      Anonymous Coward

      Conventional training costs $100+M

      Given the costs of conventional training, they must have hidden ~$100M from the books. They must have world class accountants there.

      But I see the conspiracy already rising. That small startup ran their training on the Military supercomputers in secret for free.

      DeepSeek is still considered better and faster at inference than the US crop. And they seem to be excellent at cleaning up their output from anything unwanted. Much better than the US crop.

      And as it is an OSS download, world&dog is testing it out. Something some other AI companies never did. Maybe DeepSeek is more confident about their product than these others?

      1. heyrick Silver badge
        Happy

        Re: Conventional training costs $100+M

        Maybe the accounts were done by AI?

    3. Yet Another Anonymous coward Silver badge

      >Deepseek spent a lot more on GPUs and operating costs than the ~$5 million they are claiming

      Deepsek stated that the $5M was only the GPU time spent for the final training.

      How you want to expense the R&D and costs of staff depends on what you are trying to prove.

      It's a little like the stories about India launching a rocket for what Nasa spends on Powerpoint licenses. You get a productivity boost by not having to pay engineers $400K, and you save by not having a lot of legacy infrastructure to pay for - but it's also a cheaper not being the first people to do it.

    4. diodesign (Written by Reg staff) Silver badge

      That $5.5M figure

      To clear everything up: That $5.5M figure is from DeepSeek saying it took the equivalent of 2.78M Nvidia H800 GPU hours to train V3. DeepSeek then multiplied that number by $2/hour to estimate that it would have cost less than $6M to train V3 in the cloud.

      But DeepSeek trained V3 and R1 on its own GPUs plus other expenditures so the true cost was not disclosed by the lab. Just the equiv cloud bill.

      C.

    5. DS999 Silver badge

      Singapore connection

      Nvidia's latest quarterly results showed that 22% of their massive revenue on AI chips comes from Singapore. There's no way the tiny country of Singapore could absorb even a fraction of that. Clearly this is part of efforts to conceal the actual buyer.

    6. David Cotton

      Not at all, they just trained the model on that many hours of gpu usage. You can rent the gou time in the cloud and it'll cost similar.

      It shouldn't be long before western version trained models exist using deep seeks open source models.

  2. heyrick Silver badge

    unless you trust the Chinese state to behave itself

    Why does this point keep getting raised? Yes, it has various amounts of China-specific censorship as to be expected for something Chinese, but the idea that one should be careful about China stealing your information presupposes that the Americans do not, which I do not for one moment believe.

    Honestly the only sensible approach is to not trust any of them with sensitive/personal information.

    Which means the above doesn't really need to be pointed out...

    1. Yet Another Anonymous coward Silver badge

      Re: unless you trust the Chinese state to behave itself

      That's why you shouldn't run Linux - the Finns will steal all your data.

      1. Philo T Farnsworth Silver badge

        Finnish Line.

        Let them.

        I'd trust LInus Torvalds with my data to a far greater extent than the nest of vipers that's currently got their fangs in it.

        1. Yankee Doodle Doofus Bronze badge

          Re: Finnish Line.

          Likewise, but only because he does not seem to want my data. It's kinda like politicians, if they want the job, they are probably the wrong person to do it.

          1. Yet Another Anonymous coward Silver badge

            Re: Finnish Line.

            I mean you could always download the model and run it locally.

            But, like Open Source, that probably funds drug smuggling and terrorism somehow

      2. John Smith 19 Gold badge
        Coat

        "the Finns will steal all your data."

        You mean it vanish into "Finn air"?

        Sorry. Poor impulse control.

  3. Throatwarbler Mangrove Silver badge
    Angel

    In further news

    Today's Penny Arcade captures what many people are feeling.

  4. Nightkiller

    The problem is we are making less of them.

  5. captain veg Silver badge

    intelligence

    AI is a fun research field. Practically, what's it for?

    There's at least 8 billion instances of real human intelligence wandering around, the vast majority of which are powered by utterly tiny amounts of money and energy compared with your average LLM, while being capable of amazing feats of reasoning and imagination completely beyond our most advanced silicon.

    What's the point?

    -A.

    1. Yankee Doodle Doofus Bronze badge

      Re: intelligence

      Well, the last human intelligence that I posed 50 unrelated questions to in quick succession decided they had better things to do than to put up with me...

    2. Michael Strorm Silver badge

      Re: intelligence

      > What's the point?

      Those instances of real humans have the inconvenient habit of using that computational power for *their* own benefit *and* they the audacity to expect you to pay *them* if you want to use it.

      Companies would very much like not to have to do that.

      1. captain veg Silver badge

        Re: intelligence

        You think those data centres pay for themselves?

        -A.

        1. Michael Strorm Silver badge

          Re: intelligence

          I'm quite sure they're thinking- or hoping- that even with those overheads it'll still work out much cheaper than having to pay people regardless.

          Whether it actually will is, of course, another question.

    3. amanfromMars 1 Silver badge

      Re: a Virtual Utility for Almighty IntelAIgents ‽ All urSystems belong to Us ‽

      AI is a fun research field. Practically, what's it for?...... captain veg

      It's for that and/or those with supernatural otherworldly extraterrestrial abilities and facilities venturing into extremely rich territory and rewarding enterprises requiring more than just the hopeless and helpless leadership of those exercising the wannabe GOD complex in the midst of at least 8 billion instances of real human intelligence wandering around, the vast majority of which are powered by utterly tiny amounts of money and energy compared with your average LLM, while being capable of amazing feats of reasoning and imagination completely beyond our most advanced silicon without a clue about what to do about everything revolving and crashing around them .

      It's for future building leaderships and SMARTR* natives that don't wander around pondering on matters beyond their practical and remote virtual command and control, muttering to themselves such questions as ... What's the point?

      * .... SMARTR Mentoring Analysis Reporting Titanic Research/SMARTR Mentoring Analysis Researching TitanICQ Reports

  6. Anonymous Coward
    Anonymous Coward

    In other news...

    ...new algorithms and hardware reduce costs. Price of tech falls. News at 11.

  7. Anonymous Coward
    Anonymous Coward

    Have to confess

    I have pondered a time or two what impact essentially hot housing China would have by blocking their access to the good toys and it looks like this might be the start of it.

  8. Will Godfrey Silver badge
    Happy

    My face hurts

    That's because I haven't stopped grinning since the news about Deep Seek first surfaced.

  9. Ian Johnston Silver badge

    I'm finding it hard to get over-excited about a cheaper way of producing useless shite, though I suppose it will destroy the planet a little less before the bubble bursts, which is nice.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like