In a recent report on semiconductors, Barclays analysts highlighted three themes they believe are having an influence on the industry, this time from the perspectives of training and inference demand across the consumer and enterprise environment.

According to the investment bank, the market for training and inference accelerators from simply frontier models is expected to reach $14 billion by 2024. They also predict that demand will increase to over $1 trillion by 2027 and to $180 billion the following year.

Barclays anticipates that the majority of Training and Inference growth will come from frontier models and the adoption of Consumer and Enterprise applications in 2025 and beyond, even if other businesses such as leasing GPUs on the public cloud, tier-2 cloud, and Sovereign AI are also anticipated to develop moderately.

The experts identified the increase in custom computing as the first of three factors affecting semis. They claim that by 2028, bespoke ASICs, such as non-AMD and non-NVDA solutions, will account for 40% of consumer inference compute demand, up from 10% currently.

They predict that by 2028, bespoke ASICS will account for 25% of business Inference compute demand, up from the current 5%.

Nvidia will continue to be the "favoured accelerator vendor," but according to Barclays, AMD will hold around 10% of the consumer and 5% of the enterprise inference markets. It is also expected to reach about $4 billion in accelerator sales for AI inference the following year, $10 billion in 2026, and more than $90 billion by 2028.

The analysts' discussion on the buildout phase of cloud-hosted accelerator products (Tier 1 and Tier 2) is their second point. Over the past few years, both Tier-1 and Tier-2 cloud providers have made an effort to expand the number of AI accelerators they have constructed for rental clients. We chose to limit our research to frontier models, even if there is some disagreement about whether or not this expansion of rental accelerators has lost momentum," they stated.

Considering that more than $T in AI accelerator spending will be required to fulfil compute demand projections by 2027, they also note that the supply chain must grow.

At the moment, Nvidia, AMD, Broadcom, and Marvell shares have overweight recommendations from Barclays.

(Sources: investing.com, reuters.com)