[aurareturn]

>Demand for NVDA chips will go down, and likely will take a long time before Jevon's Paradox kicks in. reply

I expect Jevons Paradox effect to start immediately. Small businesses and enterprises are likely wanting to order DGX systems for internal inference. OpenAI and larger AI labs will want to put a bigger distance between themselves and Chinese AI models by using compute muscle and advantage. All AI labs see the same efficiency benefit so they will have to continue to compete on compute capacity.

[zippyman55]

Last week the moat was the large size or cost to generate a model. That has gone away. So there should be lots of competition now within the US from smaller companies.

[Jlagreen]

That was never the moat.

Nvidia's moat is controlling compute for 80-90% of workloads.

Why do you think Steam has more RTX 4090 than 4080 in their survey?

Ever since RTX 2080TI you could buy multi server consumer GPUs for ML. We sell PCs with RTX cards regularly to customer for small local AI applications.

Project Digits is the next thing. It is not only a Nvidia GPU in some PC, it's the real AI PC. The real deal and we're already considering switchting to that as a system instead of a PC since it's size is perfect for our Vision application.

Do you think Nvidia cares if you buy 1 Blackwell DC GPU, 13x Digits or 20x RTX 5090? In the end it's all the same turnover for Nvidia.

Nvidia's goal is to spread and dominate workloads worldwide and that no matter if DC, enterprise or consumer, Nvidia HW is used.

[poobear22]

I re-read the thread topic and it is in response to NVIDIA stock. My response was in response to LLM generators and their moat (not NVIDIA's moat). So, my point was the moat for LLM generators had diminished. Additional competition generating LLMs from new entrants may increase demand for HW. The HW may be used more efficiently but I am still waiting to see if LLM performance continues to improve.

[Jlagreen]

Ah, I'm sorry you're right a misunderstood your comment.

I agree and Nvidia positions itself for exactly that. See how fast DeepSeek will come to NIM. People are already wondering how well DeepSeek will run on Digits.

Nvidia also offers distilled open models or specific own open models so is indirectly competing in that space as well. But Nvidia isn't in the LLM generator business but in the business of "infrastructure for LLM generators"

Everyone is waiting for GPT5 or another big bang. And because it takes so much people start to think that there is a wall. And there is a wall but that wall could be also compute. Blackwell will show if there is a compute wall because simply put, if a training run with large parameter set on GPT5 takes like 4 months then Blackwell might reduce that to under 1 month with the same amount of GPUs. Getting more GPUs can get that down even more. Imagine the speed up in AI frontier model research if your training times come down 4-5x from new GPU generation and another 2x from getting twice as many.

The nice part with Nvidia is also that the old GPUs don't become obsolete, OpenAI can continue using them for inferencing or even try to use combined architecture training as long as they don't go FP4.

I wouldn't be surprised that at the end of 2025 we will see things which will make DeepSeek and GPT4 as oldschool stuff simply because of the massive compute which Blackwell will deliver this year.

[aurareturn]

New AI Lab: Trains a model using Deepseek's techniques on 2,000 GPUs

xAI/OpenAI/Anthropic/Google: Trains a model using Deepseek's techniques on 100,000 GPUs

I fail to see how this makes smaller companies competitive.

[GianFabien]

Depends on whether smaller companies can deliver sufficient results cheaper than the larger ones. There are some indications that suggest that there are diminishing returns on investing on ever more power.

It's like you don't need a 1000HP supercar to get around town, a 55HP sedan is fine for most folks.

[poobear22]

Yep, that is my point. If the large scale LLMs are not sufficiently better than the new crop of startups, I suspect the large firms will need to acquire the startups (that would be their response due to a lack of a moat). Its hard to buy everything and to know where to place your bets.

[aurareturn]

So you’re worried that LLMs have stopped scaling even though the biggest breakthrough from DeepSeek is scaling RF learning without humans?

[poobear22]

As I see it: I'm waiting to see improvements in LLM performance. What I see is an improvement in computational efficiency (less hardware needed)

If general LLMs do not show continued performance improvement, then there is a lot of excess HW that needs to be utilized somehow. If LLMs continue to show performance improvement, then the hardware can be used more efficiently.