Goldman Sachs Warns: AI Ecosystem Narrative Shifts from Two-Horse Race to Multipolar Competition
N.R. Finch
Goldman's One-Delta desk head Rich Privorotsky says the AI narrative is moving from a two-leader frontier race to a highly competitive multi-polar landscape, with enterprise inference migrating to smaller models and hyperscaler capex logic under pressure.
What exactly shifted in the AI narrative?
Privorotsky argues last week marked a significant turning point in how markets perceive AI: the story moved from "two leaders racing on frontier models" to "intense multi-polar competition."
Three threads define the shift: CEOs warning that AI compute spending is high and near-term returns low; rising demand for cheaper, local, non-frontier models — especially from Chinese competitors; and hyperscalers' free cash flow turning negative as debt quietly builds.
This means → the market is no longer just asking "whose model is best" — it is asking "whose spending pays off."
Can frontier models still defend their pricing moat?
Meta and xAI have shipped API products that match frontier quality at sharply lower prices, intensifying token-pricing competition.
Bulls argue cheaper tokens will unlock exponential demand, benefiting hardware suppliers. Privorotsky pushes back: that demand elasticity is not certain.
The Ramp AI Index shows enterprise paid AI adoption has stabilised above 50%. In plain terms = most companies that will use AI already do — a price cut alone may not trigger a fresh demand wave.
Where is real enterprise deployment heading?
Production workloads are migrating fast toward smaller distilled models — models that compress a large model's capabilities into a compact form — optimised for document processing, compliance, and customer service. They cost less, run faster, and can operate locally on edge devices or CPUs.
Hugging Face CEO Clément Delangue has noted the same trend: cost pressure is pushing enterprises from renting proprietary frontier APIs to running their own open-source models.
This reflects the emerging long-term enterprise AI architecture: frontier models serve as "high-end reasoning engines," while day-to-day production inference shifts to dedicated local clusters.
Does hyperscaler capex logic still hold up?
Enterprise inference decentralising → API/token pricing faces structural deflation, making it harder for hyperscalers to justify massive capital spending.
Privorotsky argues the key question is not whether hyperscalers slow spending, but when shareholders demand higher capital efficiency.
This means → once investors push for capital discipline, valuation multiples on hardware beneficiaries will compress — even if earnings are still growing.
What defines the ultimate winner?
Privorotsky's core conclusion: the winner will not be the company spending the most on AI, but the one generating the highest effective output per dollar of AI investment.
He sees the "effective output per watt / per dollar" efficiency metric trending downward. In plain terms = more money is pouring into AI, but each dollar is buying less real-world result.
He therefore argues the market should look beyond AI for the next beneficiary — and he personally favours the healthcare sector.
Content is for reference only, not financial advice.