Cantor Fitzgerald On NVIDIA: The DeepSeek “Announcement Is Actually Very Bullish With AGI Seemingly Closer To Reality And Jevons Paradox Almost Certainly Leading To The AI Industry Wanting More Compute, Not Less”

This is not investment advice. The author has no position in any of the stocks mentioned. Wccftech.com has a disclosure and ethics policy.

With NVIDIA shedding nearly half a trillion dollars in market capitalization today as analysts and investors alike begin to question the demand paradigm for hyperscalers, especially given the purportedly phenomenal efficiency gains demonstrated by DeepSeek's R1 AI model, Wall Street analysts are coming out in droves with broadly reassuring takes on the GPU giant's prospects.

As we've noted in a previous post, China's DeepSeek recently shocked the global tech industry by training its R1 model at a cost of just around $6 million, which is roughly 1/50th the cost of comparable LLMs from the US and the EU. The model, in some respects, is quite superior to OpenAI's o1 model. What's more, the R1's operating costs are just 3 percent of what OpenAI typically charges for compute-intensive outputs.

ALRIGHT, HERE’S MY QUICK, TECH-FLAVORED RUNDOWN OF DEEPSEEK, WHY IT’S SO COST-EFFICIENT:

1) Big Picture on Cost: Traditional AI labs (OpenAI, Anthropic) blow through $100M+ in compute to train something like GPT-4. DeepSeek reportedly did a similarly capable model for just $6… https://t.co/etCMxlWJdH

— Wall St Engine (@wallstengine) January 27, 2025

DeepSeek was able to achieve these phenomenal efficiency gains by implementing a few novel ideas:

1. Used 8-bit floats to curtail memory use by around 75 percent.
2. Able to process multiple tokens simultaneously.
3. Only a fraction of total parameters are active at any given time.
4. Reinforcement learning, which utilizes a rules-based reward system, used to teach the model to "think through" a given problem step-by-step.

This brings us to the crux of the matter. At first glance, DeepSeek's R1 model appears to be a dark cloud for NVIDIA, questioning the need for hundreds of thousands of top-notch GPUs when an effective model can be trained with just 2000 H800s, as was the case with R1. However, a Cantor Fitzgerald analyst disagrees.

Cantor Fitzgerald: DeepSeek V3 is Actually Very Bullish for Compute and $NVDA:

"Following release of DeepSeek's V3 LLM, there has been great angst as to the impact for compute demand, and therefore, fears of peak spending on GPUs. We think this view is farthest from the truth…

— Wall St Engine (@wallstengine) January 27, 2025

Cantor Fitzgerald's investment note concedes at the outset that DeepSeek's R1 model has generated "great angst as to the impact for compute demand, and therefore, fears of peak spending on GPUs."

However, the investment bank thinks this view is "farthest from the truth."

We think this view is farthest from the truth and that the announcement is actually very bullish with AGI seemingly closer to reality and Jevons Paradox almost certainly leading to the AI industry wanting more compute, not less.

Accordingly, the investment bank declares that it would be "buyers of NVIDIA shares on any potential weakness."

For the benefit of those who might not be aware, Jevons paradox posits that increased efficiency in utilizing a resource can lead to a greater consumption of that resource. Cantor Fitzgerald has applied the same reasoning towards DeepSeek's R1 model and the democratization of the AI moat.

$NVDA

The DeepSeek sell-off:

Analyst Reactions:
🔸 JPMorgan (Sandeep Deshpande): Suggests the AI investment cycle might be overhyped; DeepSeek's efficiency could lead to a more efficient future.

🔸 Jefferies (Edison Lee): Proposes two strategies post-DeepSeek: continue…

— *Walter Bloomberg (@DeItaone) January 27, 2025

Interestingly, analysts at Citi and Bernstein have adopted a similar bullish view on NVIDIA in light of DeepSeek's advances. However, Raymond James analysts believe the development bodes ill for "large GPU clusters."

Follow Wccftech on Google to get more of our news coverage in your feeds.