[彻夜难眠●不敢相信] Nvidia, Super Micro Computer all 惨重大跌.... whats happening to AI theme?

selemuse

Senior Member
Joined
Dec 7, 2011
Messages
609
Reaction score
439
I asked your question to deepseek, R1 and Search switched on:


The recent plunge in NVIDIA's stock price, attributed to the rise of Chinese AI startup DeepSeek, reflects investor concerns about structural shifts in the AI industry. While NVIDIA specializes in hardware and DeepSeek in software, their relationship is more competitive than complementary due to **DeepSeek's disruptive efficiency innovations**, which reduce reliance on high-end NVIDIA chips. Here's a detailed analysis:

---

### 1. **DeepSeek’s Efficiency Reduces Demand for High-End NVIDIA Chips**
DeepSeek’s AI models, such as **R1** and **V3**, achieve performance comparable to OpenAI’s models but with **30–45x lower computational costs** and **3–5% of the operational expenses**. Key innovations include:
- **Mixed-Precision Training (FP8):** Native 8-bit floating-point training reduces memory usage, enabling training on smaller GPU clusters. For example, DeepSeek-V3 was trained on 2,000 NVIDIA H800 chips, while similar U.S. models required "tens of thousands" of advanced H100 chips.
- **Reinforcement Learning (RL) and Mixture-of-Experts (MoE):** These techniques allow selective activation of model parameters, cutting computational overhead. RL enables autonomous reasoning without costly supervised datasets.
- **Distillation:** Smaller models inherit capabilities from larger ones, reducing inference costs. This challenges the industry’s reliance on brute-force scaling with expensive hardware.

These advancements suggest that AI developers can achieve high performance with fewer, less advanced chips—directly threatening NVIDIA’s revenue from premium GPUs like the H100.

---

### 2. **Geopolitical and Market Dynamics Amplify the Threat**
- **U.S. Chip Export Restrictions:** NVIDIA’s inability to sell H100/H200 chips to China forced companies like DeepSeek to optimize for older H800/A800 GPUs. Ironically, this scarcity-driven innovation has made Chinese models **more efficient** and **less dependent on cutting-edge hardware**.
- **Partnership with AMD:** DeepSeek collaborates with AMD, NVIDIA’s rival, to optimize models for AMD Instinct GPUs and ROCm software. This diversification reduces reliance on NVIDIA’s ecosystem.
- **Open-Source Models:** DeepSeek’s open-source strategy (e.g., R1 on Hugging Face) democratizes access to efficient AI, pressuring U.S. firms to lower prices. Hugging Face’s "Open R1" project further accelerates this trend.

---

### 3. **Investor Anxiety Over NVIDIA’s Growth Trajectory**
NVIDIA’s stock surge in recent years was fueled by the perception that AI progress depends on its hardware. DeepSeek’s breakthroughs challenge this narrative:
- **Lower Training Costs:** DeepSeek trained V3 for $5.6 million vs. $100M–$1B for U.S. rivals, proving that expensive hardware isn’t mandatory for state-of-the-art models.
- **Market Sentiment Shift:** Analysts warn that if DeepSeek’s methods are widely adopted, demand for NVIDIA’s high-margin chips could stagnate. The Nasdaq Composite fell 2% on January 27 amid these fears.
- **Long-Term Risks:** Microsoft, Google, and Amazon are developing custom AI chips, further eroding NVIDIA’s dominance.

---

### 4. **Why DeepSeek and NVIDIA Aren’t Symbiotic**
While DeepSeek uses NVIDIA’s H800 chips, their innovations reduce the **quantity** and **quality** of chips needed:
- **Fewer Chips Per Model:** DeepSeek’s optimized training pipelines require smaller GPU clusters.
- **Shift to Cheaper Hardware:** Efficient models can run on older or less powerful chips, reducing demand for NVIDIA’s latest products.
- **AMD Partnership:** DeepSeek’s collaboration with AMD signals a potential shift away from NVIDIA in the long term.

---

### 5. **Broader Implications for the AI Industry**
- **Price Wars:** DeepSeek’s low-cost API ($0.55/million tokens vs. OpenAI’s $15) pressures U.S. firms to slash prices, squeezing margins.
- **Sustainability Focus:** Energy-efficient AI models align with global sustainability goals, making DeepSeek’s methods attractive.
- **Geopolitical Reckoning:** U.S. export controls intended to stifle China’s AI growth have backfired, spurring innovation. As Microsoft CEO Satya Nadella noted, "We should take developments out of China very seriously".

---

### Conclusion
NVIDIA’s stock decline reflects fears that DeepSeek’s software-driven efficiency could decouple AI progress from reliance on advanced hardware. While NVIDIA remains critical to AI infrastructure, DeepSeek’s rise highlights a paradigm shift where **algorithmic ingenuity compensates for hardware limitations**. Investors now question whether NVIDIA’s growth can sustain if the industry prioritizes cost and efficiency over raw computational power.

Exactly this is what people believe. But guess what, Deepseek is just the outlier able to provide more innovative approach because of the export hardware constraint by the US. But don't anyone feel that Deepseek can even do more with more powerful Nivida GPU, creating thing that far beyond current generative AI, like Adanced General Intelligence (AGI).

Very soon, Nvidia's stock will go again, making a fool of everyone that didn't see the rationality.

People really need to read the article on how Deepseek develop its model critically, rather than just take at news level.
 

Humster

Arch-Supremacy Member
Joined
Jul 15, 2000
Messages
23,095
Reaction score
2,474
running on lower resources and can do it at free.
so no need super powerful gpu liao.
fundamentals all changed liao.
could be the Netscape moment when browsers are suddenly free.
Reading about it now in other thread.
 

congster2

Arch-Supremacy Member
Joined
Feb 15, 2007
Messages
12,108
Reaction score
4,807
Ppl doing Sell the News.
Calm down for a moment.
Right now ppl still not fully understand how Deepseek actually works, i.e, the machinery behind( is it running on air? What cards? Daisy chain?)
They published a technical paper. At the high level they used reinforcement training, probably using other LLMs, instead of bottom up sup learning. So the experts out there know what is under the hood.
 

Ahgirlgirl

Banned
Joined
Jan 17, 2024
Messages
580
Reaction score
222
China will reverse engineer Nvidia H100 chips. That’s how dangerous China copycat is. Trump must take quick action to stop it. He must also look at which countries are indirectly letting China to get hold of the chips. Sanctions must be imposed on these countries.
the company will be Mvidia?
 

chaiscool

Arch-Supremacy Member
Joined
Mar 25, 2018
Messages
13,904
Reaction score
4,984
Ppl doing Sell the News.
Calm down for a moment.
Right now ppl still not fully understand how Deepseek actually works, i.e, the machinery behind( is it running on air? What cards? Daisy chain?)
Ain't that the point of playing stocks? If you fully understand then no need speculate liao. Most ppl buy cuz they speculate that it would be a great value in the future.
 

Jirachi

Great Supremacy Member
Joined
Jan 17, 2010
Messages
54,290
Reaction score
2,850
Exactly this is what people believe. But guess what, Deepseek is just the outlier able to provide more innovative approach because of the export hardware constraint by the US. But don't anyone feel that Deepseek can even do more with more powerful Nivida GPU, creating thing that far beyond current generative AI, like Adanced General Intelligence (AGI).

Very soon, Nvidia's stock will go again, making a fool of everyone that didn't see the rationality.

People really need to read the article on how Deepseek develop its model critically, rather than just take at news level.
The DeepSeek thing is likely overblown.

There are a lot of LLMs on Ollama with FP8 quantisation and comparable parameters size. I really think this is a China operative.
 

AlmightyOnes

Arch-Supremacy Member
Joined
Sep 16, 2010
Messages
24,855
Reaction score
13,301
rn6gwpi3bgu11.gif
 

Humster

Arch-Supremacy Member
Joined
Jul 15, 2000
Messages
23,095
Reaction score
2,474
Finally can read what makes DeepSeek tick.

Exactly this is what people believe. But guess what, Deepseek is just the outlier able to provide more innovative approach because of the export hardware constraint by the US. But don't anyone feel that Deepseek can even do more with more powerful Nivida GPU, creating thing that far beyond current generative AI, like Adanced General Intelligence (AGI).

Very soon, Nvidia's stock will go again, making a fool of everyone that didn't see the rationality.

People really need to read the article on how Deepseek develop its model critically, rather than just take at news level.

It depends how scable it is I guess.
 

selemuse

Senior Member
Joined
Dec 7, 2011
Messages
609
Reaction score
439
Exactly this is what people believe. But guess what, Deepseek is just the outlier able to provide more innovative approach because of the export hardware constraint by the US. But don't anyone feel that Deepseek can even do more with more powerful Nivida GPU, creating thing that far beyond current generative AI, like Adanced General Intelligence (AGI).

Very soon, Nvidia's stock will go again, making a fool of everyone that didn't see the rationality.

DeepSeek shows you don't need as much capital investment as you think
They also open sourced their code as others can see how they did it
First of all, it is the beginning of generative AI advancement, and there is still way to go, Chat Generative AI is just one approach, and there are other usage beside chat. Deepseek deal with Chat because this is what it currently do, who know they able to advance further into other areas with more Nvidia GPU.
 

Clarence NMT

Supremacy Member
Joined
Dec 28, 2012
Messages
8,397
Reaction score
2,788
Recently got edmw bash china thread mention nvidia will not drop, think he lose money liao.

Btw nvidia can say still stable, cos they not doing mainly ai alalso. Many think need gpu.

Openai maybe more impact, cos ppl free they charge money
 
Important Forum Advisory Note
This forum is moderated by volunteer moderators who will react only to members' feedback on posts. Moderators are not employees or representatives of HWZ. Forum members and moderators are responsible for their own posts.

Please refer to our Community Guidelines and Standards, Terms of Service and Member T&Cs for more information.
Top