Tech Stocks Brace for a DeepSeek Haircut, NVIDIA Down 12% in Pre-market Trading

[Albatros39]

Worth pointing out that the models that people are running on M4, Mac Minis and gaming GPUs have very little to do with the actually capable one DeepSeek is operating on its website. That is a completely different, much larger model requiring at least 700GB of VRAM.

What the market is concerned about is that such a capable model could be trained with ~5M USD worth of compute (excluding GPU costs). That doesn't mean though that putting more compute on it won't improve the results...

But also some people are running the 685B model.
It gets 6 tokens/s with 1 Epyc CPU, using the 12-channel SP5 platform, with a maximum memory bandwidth 576 GB/s, per CPU.
The DDR4 platform gets <2 tokens/s, but with a few RTX3090 you can get it to usable speed.
You still need an expensive machine, but with used parts, its doable for an individual.

 

[csendesmark]

I liked the 8B Llama version with the full F16
DevQuasar/deepseek-ai.DeepSeek-R1-Distill-Llama-8B-GGUF

DevQuasar/deepseek-ai.DeepSeek-R1-Distill-Llama-8B-GGUF · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

Was fast and did solve a lot of problems what was a real struggle for Phi-4 and Llama 3.3.

But the Qwen 32B Q4_K_M was really slow for some reason.

 

[Solid State Brain]

Sure but - to play devils advocate , 'we reuse portions of the SFT dataset of DeepSeek-V3' will read as "we bypassed alot of the heavy computational expense by using output from other models".

This just literally means they reused portions of the same SFT (supervised finetuning) data they had previously used for DeepSeek V3 to make it a useful chat model.

The people that sank trillions of dollars into open ai, and said other models, thats what they will see. Add to that a general distrust of AI announcements from Chinese firms/labs and voila.

I think most of that will probably be infrastructure and R&D. The material training costs are lower, although DeepSeek showed that a fairly good model could be trained at much lower costs than commonly thought—partially due to the architectural choices employed for their own model, partially due to the reasoning capabilities they gave it.

 

[mb194dc]

They clearly stated that they're using H800's for their model:

Yes, but on a tiny scale, it only cost $5m... Prices for renting GPU time could also come down even more.

Simon Willison on llm-pricing

16 items tagged ‘llm-pricing’. Posts about the pricing of various LLMs. See also my pricing calculator.

simonwillison.net

 

[ty_ger]

X/Twitter has yet to make money for a year as well... Money isn't the biggest thing to chase here. The real prize is power and information dominance.

The bank is trying to sell their bad twitter debt. Get it while you can, lol. The CEO of the bank said that him and Elon worked out their differences, realized he is a genius, and now you can buy that genious' debt at premium pricing. lol ... because the debt of the genius is worth owning; just not for the bank.

 

[katzi]

*Smugface*

I love that I was right about this.

BURN BABY BURN

 

[TheEndIsNear]

It's so insane how many people don't care about people only money. China should be isolated right now but ya know slave labor and easy cash makes rich people happy and that's what this world seems to care about the most.

 

[Vya Domus]

China should be isolated right now but ya know slave labor and easy cash makes rich people happy

China should be isolated for those things then and not because they made some cheap ML model.

 

[R0H1T]

I love that I was right about this.

Right about what? 10% won't even tickle Nvidia's underbelly

 

[dragontamer5788]

The bank is trying to sell their bad twitter debt. Get it while you can, lol. The CEO of the bank said that him and Elon worked out their differences, realized he is a genius, and now you can buy that genious' debt at premium pricing. lol ... because the debt of the genius is worth owning; just not for the bank.

The banks have always known the debt was bad.

The minute Elon badmouthed the deal in May 2023, the Banks knew they were eff'd. When the Twitter purchase finally completed in August there was still Elon suing to try to stop the Twitter deal. Even Elon didn't want to take over Twitter.

By November the court cases were done and Elon was solidly in control of the company. And now the debt (which was negotiated and signed and created back in April) was seen as worthless to the public.

Banks are now trying to sell the debt before it gets worse for them. Better to lose half (or more) on the debt than hope for Twitter to rebound.

 

[R0H1T]

Well, X is another mouthpiece for joint(pseudo?) Prez so holding onto the debt may not be all that bad.

 

[dragontamer5788]

Well, X is another mouthpiece for joint(pseudo?) Prez so holding onto the debt may not be all that bad.

I mean if you want it, it will be available to bond markets soon. But I'd consider it quite speculative.

 

[csendesmark]

Right about what? 10% won't even tickle Nvidia's underbelly

Nvidia's market cap is $3.49T
-13% loss means billion Dollars erased

 

[dragontamer5788]

Nvidia's market cap is $3.49T
-13% loss means billion Dollars erased

It's an insane valuation anyway.

But yeah, -13% means that NVidia lost about 4 or 5 Intel's this morning.

 

[csendesmark]

It's an insane valuation anyway.

But yeah, -13% means that NVidia lost about 4 or 5 Intel's this morning.

AI and Crypto bubble is going hard now.
And both will remain important, but yes - both are overvalued as of now.

 

[LittleBro]

Ooooops, not that I care for so called "AI" stuff but if those training costs are real, Nvidia has more than an overheating problem on the table.

 

[R0H1T]

Nvidia's market cap is $3.49T
-13% loss means billion Dollars erased

And it affects Nvidia how? Market cap doesn't really do anything to their real business, although once the (AI) gold rush ends, there will almost certainly be a mega crash all across the globe!

 

[Daven]

Tech stocks...

NVDA down almost 14%
TSMC down almost 13%
AVGO down almost 15%
AMD down almost 6%
Intel down less than 0.5%

Really shows what companies are viewed as invested in AI and by how much. Intel not so much or at all.

 

[csendesmark]

And it affects Nvidia how? Market cap doesn't really do anything to their real business, although once the (AI) gold rush ends, there will almost certainly be a mega crash all across the globe!

LOL
It effects those who hold the shares.
This is why you see headlines like this:

 

[phanbuey]

Tech stocks...

NVDA down almost 14%
TSMC down almost 13%
AVGO down almost 15%
AMD down almost 6%
Intel down less than 0.5%

Really shows what companies are viewed as invested in AI and by how much. Intel not so much or at all.

intel is also a foundry though, and most of their up and down is foundry related. You're right tho, no one expects that chip division to do anything with ai

 

[csendesmark]

It is already -15% of the company...

S&P 500 Map

Stock screener for investors and traders, financial visualizations.

finviz.com

-15.43%
-15.57%
-15.71% falling like a rock

 

[kondamin]

A lot of tech is down and I seriously doubt it's because of some language model but more because of macro economic forces like Japan upping their interest rates while the EU and UK are going to lower them And the guy who we can't mention also wants to do that (swap trade troubles we had this summer come to mind)

On the bright side, the Euro seems to be recovering a bit ground.

 

[LittleBro]

That's the beauty of stock trading, isn't it? Someone says they can do same thing better with fraction of costs and this happens.

 

[L0stS0ul]

People

sees the decline in value on the stock market today, but not the value of the company from three or five years ago? Value or rather valuation on the stock in 2010-2015, 2015-2020, and 2020-2025?

 

[mechtech]

But how deep is the seek???