Tyler Cowen exposes the empty math of AI revenue projections

By Nikhil Raghavan · Reporting from San Francisco ·

As economists on the a16z podcast note, crushing price competition and coordination bottlenecks mean frontier AI models will deliver normal economic growth rather than double-digit takeoff.

The Math Behind the Revenue Mirage

When economist Tyler Cowen asked Bay Area tech executives which economic sectors would grow quickly enough to justify the eye-watering revenue projections for Anthropic and OpenAI, nobody could name one. On a recent episode of the a16z podcast, Cowen and economist Alex Imas pulled the plug on Silicon Valley’s favorite fantasy: double-digit GDP growth powered by autonomous artificial intelligence.

The mechanism behind their skepticism is straightforward if you look at the unit economics instead of the pitch decks. Citing corporate spending data from Ramp, Imas noted that while token consumption is growing exponentially as agentic workflows replace basic chatbots, frontier lab revenues remain flat. Severe price competition between Anthropic, OpenAI, Google DeepMind, and open-weight models is driving the cost per token down faster than enterprise usage can scale. Labs are burning exponentially more compute to execute complex loops, but they are taking in the same revenue.

For frontier AI to drive a macroeconomic singularity, Imas explained, the economy's idea production function must exhibit increasing returns—a model parameter greater than one. Empirical estimates put that parameter decidedly below one. Furthermore, as automation drastically reduces the cost of digital tasks, non-homothetic consumer preferences kick in. The cheap, automated output accounts for a shrinking share of total GDP, while economic value concentrates in real-world bottlenecks like land, energy, and human oversight.

Parallel Processing Is Not Coordination

The core technical failure in the runaway growth thesis lies in the distinction between raw compute and coordination. On the podcast, Imas highlighted findings from his paper on "organizing intelligence," pointing out that current agentic swarms are merely running parallel processing loops. They are not coordinating. In standardized testing, two AI agents splitting a task often perform worse than a single agent operating alone. Human teams achieve exponential gains through division of labor and adaptive communication.

In infrastructure engineering, we learn early that throwing uncoordinated nodes at a shared state problem yields lock contention, latency spikes, and degraded throughput. The same reality applies to multi-agent software. An agentic system that routinely books your flight in the seat right next to the lavatory because it cannot navigate commercial choice architecture is a system bound by hard implementation friction.

Cowen added that fast technological progress in narrow domains only forces systems to hit real-world bottlenecks faster. Roughly 40 percent of the United States economy—including healthcare, education, and federal administration—is locked behind bureaucratic inertia and regulatory drag. As Cowen dryly observed, the IRS still relies on fax machines, and enterprise edtech software remains behind consumer standards. You cannot code your way around a municipal permitting board with a faster LLM.

Why Stagnation Rules Supreme

Advocates of rapid takeoff argue that simulation will soon replace physical experimentation, allowing AI to iterate scientific breakthroughs in a closed virtual loop. But as Cowen pointed out, simulation is a poor substitute for empirical reality. When physical validation is required, real-world latency wins every time.

The Bay Area tech consensus mistakes model capability for economic absorption. We have spent fifteen years watching software fail to transform legacy physical systems at the speed promised by venture capital. Expecting agentic AI to break through 200 years of economic history and deliver 10 percent annual growth ignores how production functions actually scale. When token prices collapse toward marginal cost and uncoordinated agents choke on real-world edge cases, the page for systemic failure will land where it always does: on the human institutions that have to clean up the deployment.

Sources

  1. What History Tells Us About the AI Revolution | Tyler Cowen & Alex Imas