Transistor counts, launch costs, and AI token prices all followed exponential curves — but the stories they tell are very different.
Understanding that gap is the only way to make sense of technological innovation in 2026.
The Exponential Curve That Started It All
In 1965, Gordon Moore noticed that the number of components on integrated circuits was roughly doubling every year, later revised to every two years.
From the 1970s onward, transistor density followed that exponential trend for more than five decades, with counts doubling about every two years on a log-linear plot.
Our World in Data’s analysis shows the consequences clearly: computational capacity, measured in FLOPS, grew exponentially from 1975 to 2009, roughly doubling every 1.5 years.
Over the same period, computing efficiency (energy per computation) halved every 1.5 years, and the price per terabyte of storage dropped along a similar curve.
I’ve seen senior engineers roll their eyes at “exponential tech” slides, but the Moore’s law data is not hype — it’s a century-scale graph that explains why your phone today beats 1990s supercomputers without anyone calling it a miracle.
Exponential Tech, Linear Institutions
Moore’s law never was a law of physics; it was an empirical observation of how manufacturing, design cleverness, and demand interacted over time.
The same article that popularized Moore’s law also led to “Moore’s second law”, or Rock’s law: the capital cost of semiconductor fabs rises exponentially as we push each new generation of chips.
Experience curve effects show up elsewhere too: every doubling of cumulative production in many industries tends to cut unit cost by a fixed percentage, from solar panels (Swanson’s law) to LEDs (Haitz’s law).
In all cases, the exponential improvement in performance or cost comes from painstaking, incremental work — better processes, tighter tolerances, quieter revolutions in factories.
I’ve been in boardrooms where someone throws “exponential” on a slide to justify a moonshot, ignoring the other side of the curve: R&D, tooling, and capital costs that also compound, sometimes faster than the benefits.
SpaceX and the Economics of the Exponential Drop
Rocket launches are a good example of exponential change that broke an entire industry’s spreadsheet.
Before SpaceX, every orbital rocket’s first stage — the most expensive part — was discarded after one use, like throwing away a Boeing 747 after every flight.
SpaceX has now landed and reflown first-stage boosters more than 300 times, with one booster flying 22 missions.
As a result, the cost per kilogram to orbit fell from roughly $54,000 in 2000 to around $2,700 today, a 95% reduction.
That is not cosmetic optimization; it’s a restructuring of space economics.
Satellite operators repriced constellations, Starlink became viable, and in 2026 SpaceX filed to build a “space cloud” of up to one million satellites — about 100 times larger than Starlink — to support AI computing in orbit.
From a consulting perspective, this is the kind of exponential change that breaks every historical cost model. If your spreadsheet still assumes $50,000/kg, your business plan belongs in a museum.
Token Economics: Exponential Deflation, Exploding Bills
AI tokens are following a different exponential path.
Benchmarks like the MyTokenTracker AI Cost Index show frontier models at a blended $4.64 per million tokens, with inputs around $2.31 and outputs around $11.63, while budget models cluster around $0.70 per million tokens.
Analysts point out that per-token prices have fallen roughly 99.7% compared to GPT‑3-era rates, and one deflation curve estimates AI inference costs dropped 300x since GPT‑4 launched at $30 per million tokens in 2023.
Despite that, enterprise AI bills have tripled: agentic workflows multiply token usage by 50–500x per task, and more than 70% of production AI cost now sits outside the model invoice, in orchestration, retrieval, retries, and observability.
Microsoft’s FinOps guidance captures the new reality succinctly: in AI applications, tokens are now cost, and token economics deserves architectural attention equal to model choice.
Several enterprise frameworks urge leaders to treat AI as an economic system governed by unpredictable token-based costs, not just as “a model running somewhere in the cloud.”
I’ve watched teams celebrate a switch to a cheaper model, only to discover that their new agent architecture silently multiplied usage so much that their monthly bill went up anyway. The exponential moved from unit price to volume.
Exponential Innovation, Asymmetric Value
What ties semiconductors, rockets, and tokens together is simple: exponential curves change not just capabilities but who captures value.
Moore’s law expanded total computing capacity, but Rock’s law pushed fabrication costs so high that only a handful of players can afford leading-edge fabs.
SpaceX’s 95% cost reduction to orbit made new constellations possible but also concentrated launch economics around a single provider with unmatched reuse experience.
AI token deflation opened the door to budget models near $0.70 per million tokens, but the real money still flows to whoever controls agent orchestration, routing, and the workflows that drive 50–500x token multipliers.
Token economics research now frames tokens as production factors, exchange media, and units of account, unified across micro (single agent), meso (multi-agent systems), and macro (ecosystems).
Future directions in that work include differentiable token budgets and dynamic markets, where budget allocation itself becomes a learned, optimized process.
From where I sit — working with teams trying to tame AI infra costs — the pattern is clear: exponential technological progress shifts the bottleneck from “can we do this?” to “who owns the economic levers that decide whether this is worth doing?”
The Next Exponentials to Watch
Moore’s law may be slowing at the transistor level, but exponentials are alive and well in stacked 3D chips, specialized accelerators, reusable rockets, and AI agent economies.
Launch cost curves and token price indices are both flattening today, yet usage is still on an exponential trajectory, driven by more ambitious agentic workflows and orbital infrastructure plans.
The interesting question for the next decade isn’t “will technology keep improving exponentially?” — we already know it will, in multiple domains.
The real question is which exponentials we choose to architect around: cheaper tokens with smarter budgets, cheaper launches with orbital compute, or something we haven’t graphed yet.
And whether the organizations funding those curves actually understand the economic math behind them, or are still treating exponentials as a buzzword on a slide.