Barclays: LLM computing demand far exceeds expectations, AI chip spending wave has not peaked yet
Barclays analysts say that spending on artificial intelligence chips needs to increase to meet the expectations of model developers. Analysts believe that future computing demand will far exceed expectations, with nearly 20 million chips expected to be needed to support cutting-edge models by 2027. Market differentiation will continue, with solutions from NVIDIA and AMD suitable for training and inference, while custom chips will be used for professional workloads. The inference market is also expected to remain strong
According to the Zhitong Finance APP, Barclays stated that spending on artificial intelligence (AI) chips must increase to meet the minimum expectations of model developers.
Barclays analyst Tom O'Malley stated that since NVIDIA (NVDA.US) released its financial report at the end of August, the selling of AI chip stocks has triggered a reaction throughout the AI ecosystem, with increasing doubts about whether the AI chip spending wave is "approaching its peak." Analysts believe that this view does not properly consider future computing needs.
The analyst added that at least 9 independent organizations are developing cutting-edge, large-scale language models (LLMs) with massive parameters. Due to various reasons (investment returns, funding, training data limitations, roadmap issues, etc.), most of these companies cannot guarantee that they will continue to advance the next round of model development. However, just a few model iterations require an incredible amount of computation.
This demand far exceeds the industry's current expectations, and analysts believe that the following three factors are key to covering the stocks.
First, the gradual realization of expected capacity is needed. Analysts estimate that by 2027, nearly 20 million chips will be needed for training just three models with about 50T parameters each. One major reason for the high demand is that the growth rate of new model computing requirements is expected to far exceed current expectations.
Second, there is a way for both businesses and consumers to win. Analysts believe that a dual approach should be taken in terms of AI accelerators, with NVIDIA and AMD's commercial solutions more suitable for training and inferring cutting-edge models, while custom chips will be used for more specialized workloads within data centers.
The analyst added that they see this situation developing overall as expected, with a few exceptions such as Apple (AAPL.US) using TPU. They continue to expect that this market differentiation will play a role in the future.
Third, the inference market will be very strong. O'Malley stated that NVIDIA recently claimed that about 40% of its chips are used for inference, and with other accelerator providers refocusing on the inference market, this provides support for the development of this emerging market.
Inference will become the main channel for monetizing cutting-edge models currently under development. Analysts state that investing more resources to reduce inference costs will help improve the return on investment in model development