Analyst: AI's Trillion-Dollar Investment Bets on 'Unlimited Compute Demand' — Three Key Assumptions Carry Hidden Risks
nashnova research
Bloomberg strategist Simon White argues the trillions flowing into AI capex are really a bet that compute demand will grow forever — and the three assumptions holding that thesis together — Jevons paradox, scaling laws, and the task frontier — each carry fracture risk that could force a broad repricing.
What are the trillions actually betting on?
The surface bet is large-model commercialization. The real underlying assumption is far bolder: compute demand has virtually no ceiling.
OpenAI CEO Altman has called token demand "limitless"; Nvidia CEO Huang speaks of a "double exponential." Industry leaders are actively reinforcing this narrative.
This means → if the premise holds, any amount of capex gets absorbed. If it doesn't, trillions of dollars sit on quicksand.
Assumption one: will cheaper always mean more?
Bulls invoke the Jevons paradox — an economic pattern where lower unit costs drive higher total usage — to argue demand will expand without limit.
Token prices are indeed falling, reflecting rising model supply and intensifying competition.
In plain terms = price cuts do unlock new use cases, but only if enough "tasks worth doing with AI" actually exist. If supply grows faster than real task demand, the price war compresses margins without lifting demand.
Assumption two: will models keep getting stronger?
Scaling laws — the idea that more data, parameters, and compute yield steadily better models — are the core article of faith behind AI lab spending.
White notes this pattern has not been proven to hold indefinitely. Once capability gains slow, the business model shifts from "creating ever-stronger intelligence" to inference services.
This means → training accounts for roughly 20%–40% of token consumption. If scaling laws plateau early, training-side compute demand takes the first hit, and the lofty valuations of OpenAI and Anthropic lose their anchor.
Assumption three: how large is the "task universe"?
Even if models keep improving, whether compute demand can expand without limit depends on how many new tasks large language models can actually take on.
AI researcher François Chollet compares intelligence gains to "polishing a ball rounder" — the closer to optimal, the lower the marginal return.
This reflects a deeper constraint: many real-world problems are "computationally irreducible" — they cannot be efficiently solved by pattern-matching over training data. In plain terms = some problems are simply not suited to large models, and the ceiling on tasks is itself the ceiling on compute demand.
What is the market already sensing?
AI tech stocks this year are still buoyed by improving earnings and revenue expectations, yet forward P/E ratios are declining.
This means → investors are paying a shrinking multiple for long-term AI growth — near-term profits look fine, but confidence in the "future story" is being marked down.
White argues the market is alert to rising competition and margin pressure, but skepticism toward the core "limitless compute demand" assumption remains insufficient.
Where would a break in the chain hurt most?
The trillion-dollar bet relies on a single chain: models keep improving → addressable tasks keep growing → token demand keeps rising → compute demand keeps expanding.
If any of the three assumptions tops out early, the "demand never ends" narrative fails.
In plain terms = all three assumptions don't need to collapse at once — just one hitting a ceiling breaks the entire chain, and the AI infrastructure investment thesis built on it faces repricing.
市场有风险，内容仅供研究参考，不构成投资建议。