Goldman says AI agent economics may turn profitable sooner, favors AVGO, NVDA, AMD

Goldman Sachs said AI infrastructure investments may begin generating profits earlier than investors expect, as token prices stabilize while underlying computing costs continue to fall. In a report published May 13, 2026, the bank projected AI token gross margins could turn positive in the first half of 2026. Goldman said compute costs tied to platforms from Nvidia (NVDA), AMD (AMD), Alphabet (GOOGL) and others are falling 60% to 70% annually. It forecast global AI queries could rise to 23 billion per day by 2030 from about 5 billion in 2025, with AI agents handling as much as 30% of demand. Combined consumer and enterprise agent usage could lift monthly token consumption to about 120 peta tokens by 2030, more than 24 times current levels. The bank named Broadcom (AVGO), Nvidia and AMD as preferred semiconductor plays; Alphabet, Amazon (AMZN) and Meta (META) in cloud and internet; and Microsoft (MSFT), Cloudflare (NET) and Accenture (ACN) in software and IT services.