OK, can NVDA increase its share of the market as it has for the last 5 years going forward? Ofc not. We can extrapolate that curve to the point where NVDA will reach 100% of the market. 200% of the market. Impossible!just frank,
I'm a US-based index investor, mostly corresponding to the SP500. So whatever % that is, I haven't looked it up today.
I am saying that Nvidia in the 2020s is like Intel in the 90s. Is tech overblown? Maybe. Maybe not. But Nvidia is not Pets.com.
But, that was not the relevant question. The relevant question is whether NVDA stock price can go up at this trajectories forever. It didn't happen for Intel. So, why is this possible for NVDA?
KlangFool
NVDA is clearly 'having a moment'. They are filling a very large latent need (for cheaper and faster compute). The tech companies have had 10 years of good ideas about how to improve their products, many of which were not feasible bc of the stalling of computation improvements.
Now NVDA has rung the dinner bell... and all those companies are innovating like mad their whole backlog of good ideas, using compute that is cheaper than they ever imagined.
And NVDA has (for the moment) a proprietary and highly versatile, hierarchical architecture for multiprocessors.
A decade of latent demand from all the deep pocketed tech companies, combined with an effective monopoly on a breakout product line? Those chips and servers are being bid up to the stratosphere, and NVDA is selling everything they can make at up to 10x their (estimated) cost of manufacturing. E.g. the H100 fits on a single large chip (with a chiplet architecture) that pulls about 300W. The card running on the chip costs $30k MSRP in 2024. Most people thing it takes TSMC (in Taiwan) maybe $3k to build that chip.
AI is just part of this story. This is about the future of tech and computation. And automation and robotics. And labor.
So, will NVDA get to keep it 10x markup forever? Will sales slow down (or their rate of growth cool slightly) as latent demand catches up? Well, before Moore's Law stalled a decade ago, a lot of folks thought... oh, well, computers are fast enough now to do anything we want... we won't need faster computers. And those people were wrong. Faster computers opened up new services and lower costs for existing services... up until Moore's Law stalled.
So I'd bet that in 2035 we will want 10x the compute as in 2025. Can NVDA hold the crown forever? How long did Intel stay on top? That is a management question.
For the nerds, the H100 chip has 14,000 (simple) separate processors onboard, each running at >1 GHz, sharing 80 GB of on-board high speed RAM over a bus that is 5000 bits wide and runs at 16 trillion bits per second bandwidth, with very low latency, not counting their cache memories. A total of 80 billion transistors built with a 5 nm process. If you care about 'floating point operations', it can nominally do 50 million million of them per second (FP32), or 50 TFlops.
https://www.techpowerup.com/gpu-specs/h ... 0-gb.c3899
The 7 year old cards I use are about 1/3rd the specs of an H100, and are for sale used on EBay for $500, about $30/TFlop.
Statistics: Posted by just frank — Sat Jun 22, 2024 9:19 am