Sunil Venkataram
Back to garden
•
🌱 budding
Two Exponentials - AI Capability vs Economic Diffusion
type/notetopic/aieconomicsinvesting

Two Exponentials — AI Capability vs Economic Diffusion

The two curves

Capability exponential. Pre-training scaling laws (2017-present) plus RL scaling laws (emerging 2025-2026) drive log-linear improvement in what models can do. Amodei frames this progression as smart high schooler → smart college student → PhD-level → beyond PhD, with coding already past PhD equivalent. This curve follows compute investment and shows no signs of plateau.

Diffusion exponential. How fast capability gets absorbed into the economy. This curve is downstream of the first and shaped by: enterprise procurement cycles, security and compliance reviews, change management, budget approval chains, and the gap between individual-user adoption and organizational adoption.

The two are connected but offset. Individual developers adopt Claude Code within weeks of release. Series A startups follow within months. Large enterprises follow quarters to years later, even when actively trying to adopt.

Why the gap matters

For investors: the capability curve is visible now (benchmarks, demos, research papers). The revenue curve lags. Investing based on capability alone overestimates near-term returns. Investing based on current revenue alone underestimates the trajectory. The right frame: how fast is the diffusion exponential accelerating?

For incumbents: the diffusion lag is the window for adaptation. Category A incumbents (per The Agent Is the Customer - A Convergence Thesis on Where AI Value Accrues) use this window to ship AI-native products. Category B incumbents waste it.

For workers: displacement is real but lagged. The capability to automate a job arrives before the organizational will to act on it. This creates a planning window, but the window is shorter each cycle as adoption accelerates.

The “country of geniuses” timeline

Dario Amodei gives 90% probability that a system equivalent to “a country of geniuses in a data center” arrives within 10 years. His personal 50/50 estimate: 1-3 years (2026-2027). The transition from that capability to trillions of dollars in annual economic impact: before 2030, likely by 2028.

Anthropic’s own revenue trajectory as evidence of the diffusion curve: $100M (2023) → $1B (2024) → $9-10B (2025). 10x per year, supply-constrained.

Connection to existing thesis

This resolves a tension in AI and Investing Thesis: TAM of Intelligence is Infinite describes the capability ceiling (unbounded demand for cognition), while the two-exponentials frame explains why we aren’t there yet. The lag is diffusion, not capability. Absorb Automate Unbundle - Three Phases of Technology Deployment maps the diffusion curve’s shape: absorb phase is fast, automate phase is medium, unbundle phase is slow but where the real value accrues.

The capital allocation question: companies spending ahead of the diffusion curve (building capacity for demand that hasn’t arrived) look unprofitable. Companies riding the diffusion curve look like hypergrowth. The Circular Financing as a Bubble Signal in AI risk is that some of the apparent demand is recycled capital, not genuine diffusion pull.

Counter-argument: the second curve is probably not an exponential

The frame’s persuasive power comes from calling both curves exponential. Diffusion of general-purpose technologies has historically been logistic — slow, then fast, then flattening — and the flattening arrives from a direction the frame does not model. Electrification took roughly four decades to show up in productivity statistics, and the constraint was never awareness of electricity. It was the complementary reorganisation of factories, which is precisely the mechanism Productivity J-Curve - Why Transformative Technologies Suppress Measured Output Before Harvest describes. Reorganisation is a human and institutional process, and those do not run on exponentials.

Naming the second curve exponential therefore assumes the conclusion. It converts an open question — how fast can organisations actually absorb this — into a parameter.

The framing is also hard to falsify in the short run. Any absence of measured impact is absorbed as diffusion lag rather than counted as evidence that capability is overstated on economically relevant work. A thesis that explains every observation equally well is not doing predictive work. Excessive Automation - When AI Displaces Without Real Productivity Gains offers the competing account — that some of the missing gains are genuinely missing rather than merely delayed — and the two are not currently distinguished by any stated test.

Finally, the 2028 date originates with a frontier lab CEO. That does not make it wrong, and Amodei has been directionally right before. It does mean the estimate carries an incentive, and it should be weighted as a forecast from an interested party rather than as a neutral base rate.

The empirical tell, stated in advance. Pick the diffusion metric now — revenue per employee in AI-exposed sectors, or share of task-hours automated — and check whether its second derivative is positive or turning negative. Exponential diffusion requires acceleration; logistic diffusion shows an inflection followed by decay in the growth rate. Committing to the metric before the data arrives is what keeps this falsifiable, and the absence of such a commitment is the frame’s main weakness today.

Backlinks (7)
Connected Notes