Franklin Templeton Portfolio Manager Matthew Moberg and Research Analyst Emily Elott published a white paper in September 20261 that is neither a hype document nor a bear case. It is a structured attempt to explain the mechanics behind artificial intelligence's sustained, accelerating improvement, and to place that improvement in its correct historical context. What emerges is a compelling argument: the world is not experiencing a technology cycle. It is living through an industrial revolution.
The S-Curve Problem
The paper's first move is to establish the normal pattern of technological progress. Most innovations advance along an S-curve: a brief period of exponential improvement, then a plateau. The lightbulb doubled in brightness twice after Edison's 1879 invention, then leveled off. Commercial aircraft cruise speeds surged from 1933 to the late 1950s, then stalled. The iPhone introduced most of its foundational features within its first decade and has offered diminishing improvements since. As Moberg and Elott observe, "most innovations haven't improved further than the third square of the chessboard, much less the second half."
The chessboard analogy is the paper's organizing structure, and it earns its prominence. A peasant inventor asks a Maharaj for rice, doubled across each of 64 squares. By the 32nd square, the Maharaj owes roughly 540 camel loads, a manageable sum. By the 64th, he owes 18.4 quintillion grains. The second half of the chessboard contains 99.9999999% of all the rice. "When improvement extends into the second half of the chessboard," Moberg and Elott write, "that is when we enter the territory of industrial revolutions."
The Semiconductor Exception
The semiconductor is the rare technology that never stopped doubling. For 60 years, transistor counts on a chip roughly doubled every two years, what Gordon Moore identified in 1965. Today's leading chips carry over 200 billion transistors. If lightbulb lumens had scaled at the same rate, viewing one would permanently destroy a person's vision. If the automobile had scaled at the same rate, it would travel at approximately 10 billion miles per hour.
Moore's Law is now slowing, constrained by both physics and economics. Transistor gates are just 10 to 15 silicon atoms thick. Heat dissipation has become a ceiling. Each additional doubling requires exponentially more engineering capital. The paper is direct: Moore's Law carried the world through the first half of the chessboard. It built mainframes, personal computers, and smartphones. Its era is closing.
AI Scaling Takes the Baton
In 2022, NVIDIA CEO Jensen Huang declared Moore's Law dead and introduced the H100 GPU. That same year, ChatGPT launched. Moberg and Elott view those two events as the opening of the Fourth Industrial Revolution. The H100 delivered roughly three doublings in raw TFLOPS over its predecessor, the A100, through architectural innovations that far outpaced transistor density gains alone. Blackwell, introduced in 2024, added two more. Rubin is projected to contribute several more still. The paper's GPU generation table makes the trajectory concrete: peak AI compute has moved from 21 TFLOPS in 2016 to a projected 100,000-plus TFLOPS in 2027.
The source of these gains has shifted fundamentally. As Moberg and Elott explain, "AI compute advancements no longer derive from just an individual chip. Rather, improvement comes from the entire system operating as one machine." Memory, networking, power, packaging, and software are now co-optimized as unified infrastructure. The authors conclude that the next S-curve is not a single curve but many reinforcing S-curves compounding simultaneously.
The paper also surfaces the "Bitter Lesson," a 2019 finding by AI researcher Rich Sutton: brute-force computation consistently outperforms elegant, handcrafted programming. Algorithms matter, but those that succeed rely on general methods that improve with more compute and data. As Moberg and Elott note, "Moore's Law may be waning, but compute power is accelerating." That single sentence is the paper's central investment thesis.
5 Key Takeaways for Advisors and Investors
- AI is not a single S-curve event. It represents a cascade of reinforcing S-curves across chips, memory, networking, and power, sustaining exponential compute growth well beyond what any single innovation could deliver on its own.
- The GPU roadmap is a credible earnings signal. The TFLOPS trajectory from H100 through Rubin Ultra maps directly to infrastructure spending by hyperscalers, making AI capex both durable and traceable.
- Moore's Law served the Third Industrial Revolution. Its slowdown does not signal an end to compute progress; it signals a structural handoff to system-level AI scaling operating at the second half of the chessboard.
- The "Bitter Lesson" is an allocation principle. Compute scale, not algorithmic elegance, drives AI capability gains. This keeps hardware infrastructure central to the investment case for the AI era.
- AI capex topping $1 trillion annually is a structural observation, not a cyclical one. With 68% of S&P 500 companies citing AI on earnings calls as of Q2 2026, the adoption curve is already in its steepening phase.
Footnote:
1 Moberg, Matthew J., and Emily Elott. The Limit Does Not Exist: Rethinking the S-Curve in the Age of Compounding Innovation. Parts I and II. Franklin Templeton, September 2026, https://franklintempletonprod.widen.net/s/pnrhq6mh9j/the-limit-does-not-exist-s-curves-parts-1-2.pdf.
