Cliff Sosin shows why physical reality will limit AI returns
By Klaus Berger · Reporting from Frankfurt ·
On Invest Like the Best, Cliff Sosin correctly shows that physical limits and collapsing token prices will shift value from AI models to physical assets.
The Mathematical Limits of Pure Mind
On the podcast Invest Like the Best, investor Cliff Sosin put forward a sobering thesis for venture capitalists underwriting multi-billion-dollar compute clusters. Sosin argued that superintelligent artificial intelligence is not an unbounded scalar capable of solving every human problem. Instead, it is merely a tool for interpolating between a thin scattering of known facts. In complex non-linear dynamic systems—whether predicting weather, mixing cream into coffee, or forecasting human interactions—compounding initial errors makes pure mathematical calculation useless. To know the future, one must run real-world experiments. As Sosin dryly observed, even a 300 IQ model cannot speed up a biology experiment that requires breeding mice. The mice still take a year to grow up and die.
The Fiber-Optic Cautionary Tale
From an institutional and balance-sheet perspective, Sosin’s framing cuts directly through the magical thinking currently gripping Silicon Valley. Capital markets have redeployed roughly 3% of American gross domestic product into AI infrastructure build-outs. This bids up the cost of electricians, plumbers, and power grid capacity for the rest of the economy. On the podcast, Sosin acknowledged the bull argument: Jevons paradox, which suggests that cheaper intelligence will unleash exponentially higher demand. Yet he contrasted this with a chilling historical precedent from the '90s telecom boom. Technologists then built vast fiber-optic networks under the assumption of endless data demand. While data traffic indeed exploded, the cost of moving information collapsed even faster. Today, Cogent Communications routes massive volumes of internet traffic, yet trades as a modest business worth only a few billion dollars. When marginal costs approach zero, market size can shrink rather than expand, leaving late infrastructure investors with stranded assets.
Value Shifts to Physical Bottlenecks
If intelligence becomes cheap, ubiquitous, and commoditised, economic value does not disappear; it migrates to its complements. Sosin argued on the podcast that ultimate winners will hold physical scale, proprietary data, local proximity, and human attention. He cited Meta, which spends its revenues on data centers and inference, as a prime candidate to convert falling costs into expanding operating margins. Conversely, pure software tools that merely represent "crystallised intelligence" face severe commoditisation. Furthermore, autonomous AI agents will soon act for boundedly rational consumers. Automatically switching bank accounts or garbage collectors to save 50 cents a month will evaporate friction-dependent rents. Markets will sharpen dramatically, favoring low-cost physical operators like Carvana over digital middleman software that charges high rents.
Balance Sheets Outlast Market Manias
Advocates for the AI frontier contend that elite models will create new industries, driving self-reinforcing demand loops for supply-side investors. But history shows that capital cycles driven by cheap money and technological euphoria invariably overreach. Sosin's cold accounting reminds us that rules outlast politicians and physical reality outlasts software models. The long-term winners will not be hype-driven frontier labs selling collapsing commodity tokens. Instead, unglamorous, asset-backed entities holding physical infrastructure, local density, and genuine human attention will win. In the end, balance sheets cannot be filled with prompt tokens—they require cash flow.