How big is the open-model threat to AI hyperscalers?

S&P Global’s ‘known unknown’ is that open-weight models could close the performance gap with expensive frontier models — via local small-language models or large cloud-hosted open models. Stanford research (Jon Saad-Falcon, Avanika Narayan; John Hennessy co-author) finds SLMs often match or beat frontier models on simple everyday queries and are catching up on reasoning; routing simple queries locally could cut compute/energy costs 60–80%. Panmure Liberum’s Joachim Klement says that, if right, data centres and cutting-edge semiconductors could be replaced in four of five use cases — but the article stresses scaling and electricity assumptions remain unresolved, and Alphabet (Hennessy chairs) still plans >$1tn capex through end-2029. Cloud-hosted open weights still use data-centre capacity, so their sharper threat is to AI labs (OpenAI, Anthropic) more than hyperscaler infra. Hugging Face hosts >3mn open models; Moonshot’s Kimi K3 (2.8tn parameters) is described as outperforming closed models built >3 months earlier. OpenRouter: closed models fell from ~3/5 of queries at year-start to ~1/4 recently. DoorDash, Siemens and Airbnb shifted toward open models for savings; AT&T’s open-model share rose from 20% (May) to 40% and could reach 60%, with savings up to 80%. IP-sensitive users: Latham & Watkins buying Nvidia hardware to host/fine-tune Nemotron 3; asset managers (ApexE3) resist single-provider lock-in. Financial counterweight: hyperscalers have profitable cores and valuable capacity, while cash-burning labs depend on investors, debt, leases and forward power; BIS has warned worse returns could trigger an investment bust. Closing line: the whole world won’t run on open weights soon.