🔮 Exponential View #553: The story of 100 trillion tokens; China’s chip win; superhuman persuasion, Waymo ethics, Polymarket & hydrology++

An insider’s view of AI and exponential technologies

I haven’t read anyone as thorough as you on the AI bubble and related topics. – Miguel O., a paying member

---

Hi all,

In this edition:

• Superhuman persuasion. New data shows that AI is more effective than TV ads at changing minds, even when it lies.

• The chip sanctions failed. How “stacking” old tech is neutralizing the US blockade (and validating my 2020 thesis).

• Solving the unsolved. Terence Tao calls AI “routine” and autonomous agents are finally cracking the backlog of neglected science.

• The end of Hollywood. Financialization, not AI, may be the root cause.

In my latest video, I break down how I think about the overlapping technology S-curves that are driving the market upheaval:

There are more exponentials in this AI wave doing their work than just ChatGPT and large language models. These new technologies make new things possible: we start to do things we didn’t do before, either because they weren’t possible or because they were too expensive.

Listen on Apple Podcasts or Spotify

---

RIP chatbot, hello glass slipper

OpenRouter, an aggregator routing traffic across 300+ AI models for five million developers, released an analysis of 100 trillion tokens of usage. Their data offers a unique, cross-platform look at the market. It’s worth a deep read, but for now I’d highlight two directions:

First, the “glass slipper” effect – retention is driven by “first-to-solve,” not “first-to-market.” When a model is the first to unlock a specific, high-friction workload (like a complex reasoning chain), the cohort that adopts it shows effectively zero churn, even when cheaper or faster competitors emerge. This confirms my long-held view: customers don’t buy benchmarks; they buy solutions. Once a model fits the problem, like a glass slipper, switching costs become irrelevant.

Second, the shift to agentic inference is undeniable. In less than 12 months, reasoning-optimised models have surged from negligible to over 50% of all token volume. Consequently, average prompt lengths have quadrupled to over 6,000 tokens, while completion lengths have tripled. The insight here is that users aren’t becoming more verbose; the systems are. We are seeing the mechanical footprint of agentic loops iterating in the background.

---

China’s chip tipping point

China’s drive for semiconductor independence is accelerating faster than predicted. The recent Shanghai IPO of Moore Threads, a leading AI chipmaker, surged 425% on debut, signaling voracious domestic capital support for “China’s Nvidia” alternatives. This aligns with a bold forecast from Bernstein, that China is on track to produce more AI chips than it consumes by 2026, effectively neutralizing the intended chokehold of US export controls.