The "Achilles' heel" of the NVIDIA dynasty: Three major risks of GPUs

June 18th, global GPU leader NVIDIA becomes the world's largest company by market value

On June 18, 2024, at the close of the U.S. stock market, NVIDIA's market value reached $3.33 trillion, surpassing Microsoft, Apple, Google, and Amazon to become the world's largest company by market value.

AI's demand for GPU computing power is the biggest driving force for GPUs

Taking NVIDIA as an example, the driving force of computing power for GPUs is evident. In Q1 2024, the data center business, with a revenue share as high as 86.6%, saw a revenue growth of 426% to reach $22.5 billion. In 2016, NVIDIA's data center business revenue share was only 6.77%.

Turning Point 1: AI competition slows down, large model startups close down

On May 6th, Huanfang initiated a price war in deep exploration, followed by Alibaba, Baidu, Tencent, ByteDance, and other large models. Currently, the mainstream application of large models still remains in question-and-answer mode. The background of price reduction is that the profit model cannot be well implemented and can only be charged through tokens. If large model startups do not change their money-burning model without profitability or lack of profit expectations, there will be a risk of cash flow interruption in the future, and a large number of startups exiting will reduce the demand for GPUs.

Turning Point 2: Samsung invests in GPUs, new competitors join to lower industry profit margins

Industry leader NVIDIA's gross profit margin in 24Q1 is as high as 78.35%, with a net profit margin of 57%. High profit margins easily attract more competitors to join. Samsung publicly announced its investment in GPUs, indicating direct competition with GPU industry giant NVIDIA. Unlike Google and other internet giants that use the fabless model to make GPUs, Samsung has experience and technological accumulation in making GPUs: 1. Samsung's semiconductor business follows the IDM model, closely integrating design and process. 2. Samsung has successful experience with processors like Exynos, providing reference for chip design in terms of computing power for GPUs. 3. Samsung has high-bandwidth memory chips (HBM) that are compatible with GPUs, which can enhance the chip's computing power at the system level.

Turning Point 3: Algorithmic shift from matrix computation to non-matrix computation

Reducing GPU consumption through algorithm improvements is a significant technological risk. The first two risks impact the short to medium-term performance of GPU companies within the existing technological framework. However, the algorithmic shift is a major change, similar to the shift from gasoline cars to electric cars, directly eliminating the demand for gasoline. Matrix multiplication (MatMul) is the most computationally intensive operation in Transformer large language models (LLM), consuming high computational costs and massive memory requirements. On June 18, 2024, the paper "Scalable MatMul-free Language Modeling" (Fifth Edition) on arXiv introduced a new Transformer architecture that significantly reduces GPU and memory dependencies through non-matrix multiplication (MatMul-free), and uses an FPGA solution instead of GPUs for training and inference, processing models with billions of parameters at close to human brain efficiency with 13 watts of power Investment Recommendation: Currently, it seems that the three major turning points have shown some initial signs. If these signs continue to develop in the future, it may have a substantial impact on the GPU industry. In particular, the third turning point will fundamentally change the investment logic of GPUs. It is recommended to pay attention to other companies in the computing power industry chain: domestic CPU Longxin Zhongke, and advanced manufacturing SMIC.

Author: He Lizhong (SAC License No.: S0110522110002), Source: Shouchuang Securities, Original Title: "Turning Points of GPUs"