Some thoughts on AI - Good read

Came across this from Jefferies....There will be problems with AI as a theme one day.....

Key Takeaways from Meeting with Ex-MSFT AI Expert

AI expert Dr Fong sees LLMs commoditizing, with value shifting to AI apps/infra. Anthropic and OAI need to develop vertical apps to grow, but success is not guaranteed. Open-weight models handle ~80% of execution, but closed models capture ~90% of spend (with far higher pricing & accuracy). AI adoption is gated more by compliance/risks than ROI, but ROI is still hard to quantify. AI would disrupt UI/IT outsourcing/seat-based SaaS, but we think ERP is resilient.

Models are commoditized, pushing AI labs into the app layer with no guarantee of success. Dr Fong highlighted that: 1) OAI is moving into vertical apps, such as healthcare, financial services, insurance, and manufacturing, by hiring vertical experts; 2) Anthropic is moving horizontally, locking in customers via apps such as Claude Code (CC); 3) ~85% of his clients use CC for AI coding, instead of the cheaper Codex; 4) using his clients as examples, open-weight models now handle 75-85% of tasks, up from near-zero, but enterprises still use frontier closed models (GPT, Claude) for complex task planning and final verification; these capture ~90% of dollar spend; 5) despite model commoditization, Chinese models’ benchmark gains are overrated because, in real-world use, they often cost clients more time and money due to more hallucinations; 6) more trillion-parameter models may not be needed, as current models can adequately handle most daily tasks; and 7) super large models are driven partly by competition and partly by future use cases, such as world models, real-time AR, and multi-agent orchestration, which are further away in commercialization. JEF view: We agree models are commodities, but do not think model vendors could dominate the app layer as: 1) verticals are diverse; 2) industry know-how and data are often implicit and hard to acquire by model players; and 3) potential cultural clash between AI researchers and vertical app experts. Dr Fong's observations on enterprise spending on open/closed models are mainly on Western enterprises. In China, the majority of enterprise AI spend would still be on Chinese open-weight models via MaaS or API.

AI adoption is gated by compliance/risks rather than ROI, but ROI is still hard to quantify. Our expert thinks enterprise ROI is almost an obsolete question, as it is too hard to quantify yet. Currently, the AI budget is independent of the IT budget, and some key drivers are competition and headcount cuts. Among AI pilots and proofs of concept, only 10-15% of those are deployed, given governance, compliance and agent security issues. He sees financial services, insurance, healthcare, and legal leading in AI adoption as these industries have structured policies, compliance, and rules, which AI can absorb and then apply fast. JEF view: Investors and corporate CEOs/CFOs (the end-users of AI) need to know ROI eventually. Competition is driving down the token price and likely has much more to go, which would drive enterprise adoption via better ROI, but may not be good for AI players and the hardware supply chain.

UI/Seat-based SaaS and back-office/IT outsourcing are set to face disruption soon. Dr Fong believes that within 12 months, agents could disrupt back office/IT outsourcing, and headless applications to replace those that develop UX/UI. In 12–24 months, the agent marketplace could replace SaaS, and AI will offer outcome-based pricing to disrupt seat-based SaaS subscription and current API model pricing. Moreover, he expects enterprise IP to shift from models to a persistent memory layer that makes the underlying LLM easily swappable, so that enterprises can keep that IP value.

JEF view: ERP SaaS is hard to displace, as they are systems of record and deeply integrated with enterprises' core processes. We also think multiple revenue models will coexist.