QQQ (Nasdaq) - (Anyone who shorted the TQQQ has love this)
NVDIA
BELOW IS IS MY UNDERSTANDING — BUT I MAY NOT BE CORRECT, DON’T TAKE IT FOR GOLD.
The hype is, that DeepSeek is a Chinese artificial intelligence company, which, to me, appears to be positioning itself as an open source alternative to OpenAI and most other US based AI Products. It seems to me, DeepSeek attempts to differentiate itself from the US AI’s by keeping up with their models, while requiring much less GCP/Infrastructure to use the product.
This could explain NVIDIA tanking this morning because I think it’s rumored that it rivals OpenAI in performance, but can run locally, and doesn’t require as many GPUs/Infrastructure. Therefore, lowers the demand for hardware.
I’m not asking because I’m looking for market advice, I’m more interested if this really is a revolutionary AI out of my interest in tech.
With that said — does anyone know about this thing? It really shook things up this morning.
From what I’ve gathered and seen. It’s more innovative than revolutionary. The model they’ve created, from all accounts, is pretty good and is very efficient in comparison to what the western counterparts have done. But it’s not like they created this out of thin air as other models like Ollama are open source.
IMO, most large US corporations have over leveraged into AI/ML and are probably spending way too much on it. On top of that sanctions may have also encouraged the devs to innovate with their model to improve efficiency. This will probably just be the start. Nvidia is tanking as this could hurt demand as this model is open source and it’s not really going to stop Meta, Google, or OpenAI from implementing their efficiencies or even innovating on top of them. However, I don’t think it will have any immediate effects on demand and it will be hard to tell if it will even impact long term demand.
I expect Nvidia at least to recover. I think Deepseek did still train this thing on datacentre GPUs, it’s just they’re not supposed to have those so that part is being kept quiet.
I think that the model works great for what little I have tried it. Well, I don’t have the big boy model just DeepSeek-R1-Distill-Qwen-14B-GGUF.
They claim that even my 14B version shoud beat LLMs like Gemini, Claude and some GPT versions and to me these results seem very believable. It has performed well, but I have used it very little, so can’t be too sure yet. I’m carefully optimistic that maybe there’s actually some basis for this hype for once.
Another angle to consider is china is not known for its culture of honesty. Unless there is some sort of proof/verification of how cheap this is, I would take their claims with a grain of salt.
But yeah like many of you already posting I kinda hope they AREN’T bullshitting and this helps burst the stupid AI bubble. Although the downside of that is if it becomes much cheaper, that might encourage more absolutely asinine adoption of “AI” in places it doesn’t belong. Or maybe it removes the investor hype aura and companies care about shoehorning it into everything less. Who knows.
This led investors who do not understand machine learning to assume NVIDIA and other ML players won’t sell as much.
The reality is, people will simply do more with whatever they can afford.
They won’t buy less, just get more done.
This also let a whole bunch of people that believed Nvidia stocks were over valued to abandon ship for a reason that investors won’t end their relationships over.
3 Trillion dollar market caps makes Nvidia’s (P/E) ratio 124.35.
The big thing is the (alleged?) efficiency. If deepseek could train this model for $6m it means that much less hardware and infrastructure is needed than previously thought to develop and deploy LLMs.
obviously bad for nvidia and chipmakers (ASML, TSMC, …), since possibly 10-100x of their hardware is ‘needed’.
The big cloud providers and other AI companies like OpenAI might have massively overspent trying to brute force solutions when there was an innovative solution found by DeepSeek.
If China could catch up so quickly they might soon overtake the US.
Trump can’t even block/tariff it to protect US companies since it’s open weights (no politics, just a technical obsevation).
But there’s upsides too.
OpenAI is still making a loss even on their $200 subscription. This development shows AI might become useful in a cost effective way much sooner than presumed.
As far as I can tell (have not looked in too much depth), DeepSeek is more open than most US companies: papers with reasonable detail show the methods used and weights are open (1, 2). It should therefore not be too difficult for others to catch up. If the model really is so cheap to train OpenAI/Google/Meta could have reproduced it over the weekend, theoretically.
Anyone monetising AI (Microsoft mostly, but also Apple, Google & co.) stands to achieve profitability and scale much easier.
On the whole I’d categorise it as ‘good’ for the software people but ‘bad’ for the hardware people.
The big question mark is the claims for efficiency, cost to develop, and cost of inference. Chinese companies have been hoarding as many GPUs as they could for a long time, and nvidia even catered to them with custom models undercutting export bans. And they might have obtained a lot via middle men. No one really knows what they have and those that do won’t talk to stop the flow or get sanctioned. It would not be the first time either China subsidizes a sector to suffocate international competition (solar, steel, …)
EDIT: all that said, personally I’m not too convinced AI is that interesting for humanity. It drives the cost of mediocre output to zero, but won’t mean much for actual novel or creative work IMO…
I do not believe that will have people cancelling 95% of their GPU orders, just increasing their capacities and model scales.
If you can now train a frontier LLM on 50,000 GPU’s instead of 1,000,000 hyperscalers can now afford to simultaneously train multiple frontier LLM’s simultaneously with existing hardware, perpetual research loops, additional regression testing, scale up deployment, etc.
I’m not sure what the model training cost was. I don’t know if we should believe that $6m figure and I’m not sure if 'open’AI disclosed this for o1/o3. I was referring to API costs, i.e. inference.
Maybe. I’m in disbelief how much energy and money they were willing to spend last week. I will be in even more disbelief if they still want to scale out to nuclear reactor levels of spending this week, but maybe it’s too late to change course now.
Did wall street just randomly notice it or something finally? Odd timing. It was in my google news feed at least a month ago. I didnt see what the big deal was when I read the article. I figured most likely it was “typical China” claiming a cost of $6mil by hiding the actual cost by putting the money spent in another category, or getting the state to buy all the things they need and then claim some randomly low cost cause it didnt cost the company anything or some BS like that. Even if it was true I figured it still wouldn’t matter, as companies will just use anything more efficient to train larger models, not stop buying GPUs. But now all of a sudden wallstreet is in a panic? Just goes to show how filled with morons the investors are when it comes to anything tech related.
I need to look further into this, but it’s mostly an unsurprising development.
The AI hype train is mostly all sorts of weird, coupled with markets going Maximum Speculation. I don’t see this factually devaluing any of the hardware manufacturers unless I’ve been grossly underestimating precisely how hard we’ve been pushing the “you need 1.21 gigawatts worth of Blackwell GPUs to do anything at all” thought successfully.
really move an entire market with all its complexities?
I think the people are seeing the dips and trying to pin it on something while the truth something more simple like @Phence put it
All those companies are probably overpriced so this is just giving the market correction (to a degree) that they need.
Big money carries a lot of inertia from retail-tier traders, and they probably are shaking the tree and/or moving some money around.
I expect it to bounce back (minus panicked traders who sold after seeing red) unless companies start putting out earnings calls where they just claim “yeah we lost money we doneski”
