The prediction’s timeframe is “within roughly the next few years (by around 2029).” As of November 30, 2025, we are only about three months past the podcast release (August 22, 2025), so it is too early to judge whether this 2029‑ish claim will ultimately be correct.

What we can say so far:

• There is a visible industry trend toward more modular and specialized approaches, e.g.:
  • Mixture‑of‑experts (MoE) architectures and router‑based systems in both research and production settings.
  • Specialized / small language models (often called SLMs or task‑specific models) such as Microsoft’s Phi series and many on‑device models targeting phones and edge devices, marketed as lower‑cost and more efficient per token than frontier LLMs.
• There is also strong economic and regulatory pressure to reduce energy use and cost per token in large‑scale AI deployments (e.g., data‑center power constraints, GPU scarcity, and cloud cost concerns discussed widely in AI industry reporting in 2024–2025).

However, the prediction contains three separate long‑range components, each aimed at around 2029:

1. A significant transition of AI toward architectures built from networks of smaller, specialized models.
2. Achieving roughly 10x–100x reductions in energy and dollar cost per token relative to 2025 levels.
3. A resulting substantial increase in total token production versus 2025.

Current public information (as of late 2025) is not sufficient to:

• Confirm that SLM‑network architectures will dominate over large monolithic frontier models by ~2029.
• Verify an achieved 10x–100x aggregate reduction in cost/energy per token versus 2025 baselines, at production scale across the industry.
• Measure industry‑wide token production in 2029 relative to 2025.

Because the target year (≈2029) has not yet arrived, and the necessary market‑wide metrics are not yet observable, the prediction cannot be definitively evaluated.

Therefore the correct status as of November 30, 2025 is: inconclusive (too early to tell).