Why 200-300% growth isn't always a good thing

[anthonyms]

@GauchoRico

• If data center buildout is only 5% complete as Jensen has indicated, then after three more doublings we’re only at 40% complete. If NVDA maintains 80%+ market share then a 40% buildout is only 50% of NVDA’s potential. And data center demand won’t stand still but will keep growing so the TAM will exceed 100% of today’s number.
• There is strong evidence that demand exceeds supply. H200 will be available in a couple of months, and B100 will be available in about 7 months. So why are H100s still selling out when chips that are a lot better will be available in weeks? I think it’s because customers need compute and they need it yesterday (see #4 below for why I think this is the case); so they say “give me all the H100s you can give me and I’ll also take the H200s and the B100s”. I think this will continue through 2027 with the B200s, the R100s, and the R200s all of which are already in either full production (H200 and B100) or full development (B200, R100, and R200) as per Jensen’s keynote at Computex last week.

Let’s not forget that the AI demand driven revenue growth is a function of both price and volume. In a supply constrained world a lot of historic and future growth will come from pricing as much as unit sales growth. The H100 was over 2x the price of the preceding A100. You can bet that the H200 will be more expensive again and maybe that is why demand is there for H100 is still strong because chips are and will be selling on value prices not cost + or TCO basis.

Ant

[anthonyms]

@intjudo

• graphics rendering
• blockchain
• simulations of complex systems (weather, power grid, social networks, molecule interactions etc.)
• validating quantum algorithms
• cryptography

Hi Intjudo:- it is an interesting point you are making that the GPU HPC market is indeed made up of many more segments and use cases than just AI, however apart from continued terminal growth rates in these segments, surely these were the mature numbers pre-existing in the Nvidia revenues at the Q1 2024 stage @ $4.2bn in the quarter, the meteoric triple digit growth in revenues that has us enthralled right now is all AI related. TBD is still unknown.

I’m not sure how much AI compute demand could necessarily spark a leg up in terminal growth rates of the other mature segments.

Ant

[intjudo]

Hi @anthonyms , I think I agree? I was responding though to @stewaj1 :

@stewaj1

I think it’s inevitable that at some point their chips get better but at smaller margins and customers will upgrade less often and growth will slow

Part of what I’m trying to convey is that $NVDA so far has had two significant steps-up in revenue that were 100% un-anticipated by the market; the first was a blockchain application (bitcoin mining) and the second was a AI application (ChatGPT).

So anyone selling $NVDA because sales of graphics-processors waned, or because sales to bitcoin miners waned, lost a LOT of money, and they lost it by dong what we’re “supposed” to do: by following the numbers. Should we follow their example and sell $NVDA when the demand for AI systems levels off?

HPC has so many potential applications that imo additional super-apps, currently unknown to the market, are probably on the way; I think they will arrive with increasing frequency. So: “yes” eventually demand for GPU for AI will level off, same as it leveled off for graphics processing and bitcoin mining. But by that time, we’ll probably be in the throes of the next HPC super-app, whatever that turns out to be.

Another way to frame this is: HPC is demonstrating emergent properties. To me this mitigates the risk that demand for AI systems might eventually level off.