Support NeoGAF

[AtomicStarving]

China's DeepSeek startup roiled artificial intelligence-driven stocks on Monday as Wall Street mulled huge investments by technology industry giants in AI infrastructure as well as demand for Nvidia (NVDA) chips. DeepSeek's apparent advances have raised questions over the computing power needed to develop AI systems, a key driver for AI stocks.

• Nvidia Drops: Shares down 14% due to DeepSeek's less resource-intensive approach.
• Market Reaction: Nasdaq and S&P futures drop significantly.
• Efficiency Over Scale: DeepSeek shows competitive AI with fewer resources, challenging high AI spending.
• Investor Reevaluation: Mixed views on AI investment strategies.
• Geopolitical AI Race: Seen as a pivotal moment in U.S.-China tech competition.
• AI Investment Shift: Potential move towards efficiency in AI development.

DeepSeek Raises Questions​

"DeepSeek clearly doesn't have access to as much compute as U.S. hyperscalers and somehow managed to develop a model that appears highly competitive," said Raymond James analyst Srini Pajjuri in a Sunday report. "The natural question is, how would DeepSeek's emergence impact compute intensity growth and the demand for hardware/semiconductors?"

Startups OpenAI and Anthropic have been battling Google, Meta and others in developing large, multimodal and small language foundation models.

Further, DeepSeek has claimed it spent only $5.6 million over two months to develop its latest AI model, noted economist Ed Yardeni in a report published on Sunday.

Source: https://www.investors.com/news/tech...ial-intelligence-capital-spending/?src=A00220

 

[Hookshot]

Buy the dip

 

[ChuckeRearmed]

Very interesting, though probably can buy Nvidia on dip now.

Though for USA, instead of trying to "save the humanity" and "DEI" and "finding nazis", they should have invested more into the progress "whatever it takes". Unlike Boston Dynamics who instead of mass producing robots opted for "we don't want to hurt humans", China wanted to produce robots as effective as possible. Plus instead of trying to save the enviroment they wanted to get as much energy as possible. Without Musk USA would be in the same hole as Europe too - no electric cars, no proper rocket launches etc - especially looking at the success of Blue Origin or Rivian.

Hope it is a wake up call.

 

[mitch1971]

This reminds me of game devs who make games for 300 million bemoaning the costs, and along comes a dev and makes an equivelent game for 100 mil and shows it can be done.

 

[PSYGN]

Hah nice

 

[EviLore]

 

[ChuckeRearmed]

It is pretty clear that there was no magic involved.

 

[ManaByte]

 

[SJRB]

Typical market overreaction to something they don't fully understand.

People who think deepseek actually trained their model in their basement on a spare PS3 and a chromebook, lol. Training models is not some occult alchemy process, it's hard science and the notion that some guerilla group somehow did it with an order of magnitude less resources is naive at best.

 

[ManaByte]

 

[ManaByte]

 

[ChuckeRearmed]

Training models is not some occult alchemy process, it's hard science and the notion that some guerilla group somehow did it with an order of magnitude less resources is naive at best.

Feel the AGI ~

 

[Trogdor1123]

Assuming this is true, it could have a lot of negative impacts. First of which is that one of ais best “features” was that it did cost so much to use. There was a barrier to getting stuff up and running which forced collaboration and a bit more openness. If this is true, that could be gone and the implications are troublesome.

 

[AJUMP23]

The limits of China ai. The China virus.

 

[Sakura]

Typical market overreaction to something they don't fully understand.

People who think deepseek actually trained their model in their basement on a spare PS3 and a chromebook, lol. Training models is not some occult alchemy process, it's hard science and the notion that some guerilla group somehow did it with an order of magnitude less resources is naive at best.

I thought the big deal was that it was much cheaper to use than Open AI's stuff because it is more efficient, not that they had a magic training process.

 

[DrFigs]

If it is open source, couldn’t someone make a version without Chinese censorship? Isn’t that the advantage of open source AI? I am assuming I am misunderstanding something.

 

[lachesis]

I would love to see the rap dis battle between US AI and China AI....

(DeepSeek probably would error out as soon as Tienanmen is mentioned, as if it heard of "yo mama".)

 

[GymWolf]

And nvidia lost 400 bil against this thing?

 

[StreetsofBeige]

Lucky for me I only got one tech stock (down -10% today). But the rest of my portfolio is actually green. So only down less than 1%. Lucky.

 

[Sgt.Asher]

And nvidia lost 400 bil against this thing?

Investors are dumb panicky animals with no sense of forsight.

 

[DeepEnigma]

Investors are dumb panicky animals with no sense of forsight.

It's just the classic sell high, and rebuy more shares low from the massive hedge funds.

 

[Dr.Guru of Peru]

I mean, isn't it all open source? I'm fairly certain OpenAI/Meta/big AI labs are all going to be verifying this.

And nvidia lost 400 bil against this thing?

What does the censorship have to do with anything? No one's getting excited because of the publicly released API. If this is the paradigm shift it seems to be, Meta, MS, Google, Apple, etc. may no longer need to spend billions on nvidia chips with huge profit margins if they can get their own LLMs operating on lesser hardware or on less hardware.

 

[EviLore]

I mean, isn't it all open source? I'm fairly certain OpenAI/Meta/big AI labs are all going to be verifying this.

The models (weights) are open source. Basically, we only have the end result. There are no training logs or details other than their claim that it was trained on $6M of compute.

 

[sono]

If they have developed more efficient training that is good
nVidia are too powerful

 

[Dr.Guru of Peru]

The models (weights) are open source. Basically, we only have the end result. There are no training logs or details other than their claim that it was trained on $6M of compute.

Why would they trust the training logs though? They would have to replicate the model and evaluate for themselves its training and inference efficiency.

 

[kevboard]

them being hardware limited makes sense, that's how you are forced to make your software more efficient.

 

[EviLore]

Why would they trust the training logs though? They would have to replicate the model and evaluate for themselves its training and inference efficiency.

You can’t just replicate the model from scratch. They didn’t open source the entire training pipeline, just the resulting weights.

 

[AtomicStarving]

Somebody's not happy

 

[IDKFA]

Further, DeepSeek has claimed it spent only $5.6 million over two months to develop its latest AI model

I doubt this very much. I'm not an expert on developing LLMs, but AFAIK, training large models over weeks or months can run into tens of millions of dollars. How the hell did this cost so little and get knocked up in two months?

 

[Robot Carnival]

so, had they actually show proof to support their claim on the amount of money and time they invested in this? because I tend to not trust any numbers provided by companies from mainland China without some verifications. and even then I'll be doing it with a large grain of salt.

 

[Gp1]

Just make the game market great again Nvidia!

Plz!

 

[Chaplain]

 

[AtomicStarving]

I doubt this very much. I'm not an expert on developing LLMs, but AFAIK, training large models over weeks or months can run into tens of millions of dollars. How the hell did this cost so little and get knocked up in two months?

They did it. It's open source. There's absolutely nothing to hide or doubt. It's all about code engineering, and doing things more efficiently with less energy.

GitHub - deepseek-ai/DeepSeek-R1

Contribute to deepseek-ai/DeepSeek-R1 development by creating an account on GitHub.

github.com

 

[Go_Ly_Dow]

 

[SJRB]

This 3 minute summary is all you need to understand the situation.

 

[AtomicStarving]

Trump was clearly informed about this, by announcing massive investment in AI on the same day. A lot of things are going to change, I guess...

Beijing puts spotlight on China’s new face of AI, DeepSeek’s Liang Wenfeng

Liang’s participation at Monday’s meeting with China’s premier reflects start-up DeepSeek’s growing prominence in the AI industry.

www.scmp.com

 

[vpance]

Anyone try to run it locally yet? R1 version

 

[Gp1]

What does the censorship have to do with anything? No one's getting excited because of the publicly released API. If this is the paradigm shift it seems to be, Meta, MS, Google, Apple, etc. may no longer need to spend billions on nvidia chips with huge profit margins if they can get their own LLMs operating on lesser hardware or on less hardware.

Or, since they already invested on the hardware, we are going to see a huge leap in a short time.

 

[Tams]

I bought some Nvidia shares. I've been waiting for them to dip.

The whole $5.5 million development thing seems to be the cost of running it (electricity and staff).

It seems they used 10,000 H100s, which is $300-$400 in GPUs alone, not accounting for any costs required to get so many through the grey market.

 

[ManaByte]

 

[DeepEnigma]

Artificial, indeed.

 

[Come On Tars]

And nvidia lost 400 bil against this thing?

buying the dip might be a good idea lol

 

[Mistake]

These "investors" aren't too bright, especially given after some chinese companies were playing by different rules on the exchange and were rightly called out for it.

Buy the dip

 

[HoodWinked]

My speculation is that the Parent Hedge Fund company "High-Flyer" heavily shorted Nvidia and other tech stocks. Released this Deepseek model and exaggerated its low cost to tank the markets. Now that the trade has paid off they can cut off access to Deepseek and claim "malicious attacks".

 

[GymWolf]

buying the dip might be a good idea lol

Not really a nachos eater but thanks anyway.

 

[Hudo]

Good. Fuck Nvidia. And fuck their investors, too.

 

[Gp1]

Meta is reportedly scrambling multiple ‘war rooms’ of engineers to figure out how DeepSeek’s AI is beating everyone else at a fraction of the price

The engineers will study how DeepSeek accomplished its tech feat and how Meta can benefit.

fortune.com

 

[winjer]

 

[Kadve]

Reminds me on asking Cortana on how to pirate windows. Don't recall but was something like "I'm not allowed to answer that".

 

[RCX]