4. NVIDIA Corp (NASDAQ:NVDA)
Number of Hedge Fund Investors: 180
The AI boom is just getting started and NVIDIA Corp (NASDAQ:NVDA) is a clear leader in the AI chips industry which makes the stock one of the suitable options to buy and hold. As AMD and some other companies begin to make AI chips, there are concerns NVIDIA Corp (NASDAQ:NVDA) might not be able to continue its momentum. But many analysts believe the demand for AI chips is so huge that a few new entrants in the market would not be able to dent NVIDIA Corp (NASDAQ:NVDA).
A latest analysis shows that despite a 236% increase in stock price over the past one year, NVIDIA Corp (NASDAQ:NVDA) is the cheapest of the big three semiconductor companies (Intel, AMD, NVIDIA).
In December, Stacy Rasgon, Bernstein senior chips analyst, said while talking to CNBC that NVIDIA Corp (NASDAQ:NVDA) remains in the driver’s seat in the chip sector.
Blue Tower Asset Management made the following comment about NVIDIA Corporation (NASDAQ:NVDA) in its Q3 2023 investor letter:
“In addition to the use of larger datasets, the training speed of AI models has increased dramatically. NVIDIA Corporation (NASDAQ:NVDA)’s stock almost tripled in the first 3 quarters of this year with a 197% gain, and a large reason for this is the huge role they have played in recent AI improvements. Nvidia’s single GPU AI training speed performance has increased by a dramatic 1000x in 10 years with only 2.5x coming from Moore’s Law3 driven increases in chip density. Besides better chip manufacturing, there were three other improvement factors at play: simplifications in number representation for the weights of the neural networks, more complex mathematical instructions for reducing the computational overhead involved in mathematical calculations, and increased neuron sparsity (in neural networks, some neurons are useless and can be pruned from the network without reducing performance significantly). In addition to these single GPU improvements, Nvidia also made improvements in data center scale architecture that allows groups of GPUs to work more efficiently together.
It is noteworthy that the vast majority of the improvement came from hardware architectural and software data improvements, rather than transition density. These improvements were likely the low-hanging fruit of training speed improvements as researchers will eventually converge on an ideal architecture. The simplification of going from 32-bit to 8-bit floating point numbers for measuring weights is a one-time gain that can’t be repeated again. I expect the rate of improvement to slow down over the next ten years and eventually approach the levels of Moore’s Law improvements in chip efficiency. The historical trend for computer hardware is for it to eventually be commoditized, and I believe this will eventually occur for Nvidia’s GPUs as well.”
