A guide to where we are, and where we're going
Steam, then electricity, then oil — each one reshaped what humans could do by giving us more physical work for less effort. We are midway through a second wave that does the same for cognitive work. The curve this time is steeper, the doubling time shorter, and the consequences will arrive faster than anyone is prepared for. Here is the data, the history, and a forecast for the next decade.
01 — Two revolutions, side by side
The Industrial Revolution did one fundamental thing: it gave humans access to energy that didn't come from a horse, a slave, or a waterfall. Coal, then oil, then electricity. Once you can move atoms cheaply, you can build a factory, a railway, a city, an airplane. The Computational Revolution does the same thing for information. Once you can move bits cheaply — and now, once you can reason over them cheaply — you can build a search engine, a smartphone, a self-driving car, a synthetic colleague.
1769–1970 · ~200 years
Harnessed thermal and electrical energy to multiply physical work.
1956–???? · ~70 years and counting
Harnessing computational energy to multiply cognitive work.
Each tick is a decade. Top row: industrial breakthroughs. Bottom row: computational ones.
The point of the comparison
It took roughly 80 years for the Industrial Revolution to remake employment, productivity, and politics in Britain. The Computational Revolution has compressed similar shifts into 15 years. The doubling time of frontier training compute — six months — is faster than anything humans have ever sustained. That's not a metaphor. It's the actual measured pace, and it has held since 2010.
02 — Seventy-five years of compute
The chart below plots the training compute of every notable AI system from 1950 to today, in floating-point operations (FLOPs) on a log scale. Three regimes are visible. Pre-deep-learning (1950–2010): roughly Moore's-Law-paced, doubling every two years. Deep learning (2010–2018): a step change to 6× per year. Frontier era (2018–today): another step to 4–5× per year on systems that already cost more than a Boeing 747 to train.
Log scale, FLOPs. Each dot is a model.
Data approximated from Epoch AI's notable models database. The slope of the line steepens twice — once around 2010 (deep learning), once around 2018 (transformer era).
1950–1980
Hand-coded rules. Logic theorem provers, expert systems. Ran on machines with kilobytes of RAM. Two AI winters when the rules failed to scale to messy reality.
1980–2010
Support vector machines, random forests, hidden Markov models. Worked, but every domain needed bespoke feature engineering. Speech and search were the breakout commercial wins.
2012–2017
AlexNet (2012) won ImageNet by a country mile. GPUs from gaming get pressed into service. Google, Facebook, Microsoft pivot whole research orgs to neural nets within 18 months.
2017–2022
"Attention Is All You Need" (2017). BERT, GPT-2, GPT-3. Scaling laws discovered: bigger models with more data and more compute keep getting better, predictably.
2022–today
ChatGPT, GPT-4, Claude, Gemini. Then o1, o3, Claude 4 — a second axis of scaling: thinking time at inference. Capability gains accelerate even as pre-training plateaus.
03 — The four power laws
A power law is a relationship where one thing grows or shrinks as a fixed multiple of another. They show up as straight lines on a log-log plot. In computing, four of them rule. They have held for decades each, and there is no physical law saying they must stop tomorrow.
Law 01
+41% / year · since 1971
Gordon Moore observed in 1965 that integrated-circuit transistor counts were doubling every year, and revised it to two years in 1975. It has held — with some recent slowing — for fifty years. Intel's 4004 in 1971 had 2,300 transistors. NVIDIA's 2025 GB202 GPU has 92.2 billion. That is a 40-million-fold increase in a single human lifetime.
Critics have called the end of Moore's Law every five years since the 1990s. They've been consistently wrong. The pace has slowed somewhat at the bleeding edge (~3 years per doubling today), but it has not stopped, and 3D stacking, chiplets, and advanced packaging are giving it new lives.
Source: Our World in Data — transistors per microprocessor, Wikipedia.
Law 02
+216% / year · since ~2012
Named after NVIDIA's Jensen Huang, this is Moore's Law on steroids. It combines Moore's transistor scaling with architectural improvements specifically targeted at AI workloads — tensor cores, mixed precision, and memory bandwidth. From the K20 (2012) to the GB200 (2024), GPU inference performance for typical AI workloads has improved roughly 1,000× in twelve years.
Translated to dollars: the cost of one floating-point operation has fallen by an order of magnitude every four years on AI hardware. This is what makes massive training runs that would have cost $10 billion in 2015 dollars affordable today.
Source: NVIDIA GB200 spec sheets, Huang's Law explainer.
Law 03
Power-law decline · Kaplan 2020, Hoffmann 2022
In 2020, OpenAI researchers (Kaplan et al.) discovered that language model performance follows a clean power law: as you increase parameters, data, and compute, the test loss falls in a predictable way. DeepMind refined this in 2022 (Hoffmann/Chinchilla) by figuring out the optimal split between model size and training data.
The practical consequence is enormous. Capability became a budget question. If you want a model with X loss, you can compute the FLOPs, the dollars, and the wall-clock time required, before training. Every major lab runs models the way Boeing runs airframes — engineered to spec, not discovered by accident.
Sources: Kaplan et al. 2020, Hoffmann et al. 2022 (Chinchilla).
Law 04
~10× cheaper / year · since 2022
The most important number nobody talks about. To get GPT-4-class output on a benchmark, the API cost in March 2023 was about $36 per million tokens. By December 2024 the same quality cost about $0.13 per million tokens via gpt-4o-mini and Llama 3.1 70B. That's a 280× drop in 21 months.
This is the deflation that matters for the model on the other site you and I built. AI capability is getting exponentially cheaper for the consumer, even as frontier training costs explode. The two are not contradictions; they are the supply-curve and demand-curve sides of the same productivity wave.
Source: a16z, "LLMflation"; Epoch AI cost data.
04 — Where we actually are right now
Frontier models trained at GPT-4 scale or above
30+
across 8 labs · Epoch AI
Largest training cluster, 2026
100K+ GPUs
xAI Colossus, Meta, Microsoft
Frontier training power demand
~150 MW
a small city · Epoch
Active monthly AI users (top 5 products)
~1.2B
ChatGPT, Gemini, Copilot, Claude, Perplexity
Annualised AI revenue, top 5 labs
~$50B
2026 run-rate, mostly subscriptions + API
Capex committed to AI infra, 2024–2026
~$700B
hyperscalers + sovereigns
Frontier model performance against expert humans on standardized tests, mid-2026:
The pattern is consistent. Anything that resembles standardized testing — bounded, well-defined, verifiable — is at or above expert human level. Anything requiring multi-hour agency, novel research, or genuinely open-ended judgment is still developing. That gap is where the next two years of progress will land.
Two specific capabilities just crossed the threshold
Software engineering. Frontier agents can now close real GitHub tickets autonomously at ~75% success on the SWE-bench Verified benchmark. In production at companies like Anthropic, Google, and Cognition, AI agents now write substantial fractions of new code.
Scientific reasoning. On graduate-level physics, chemistry, and biology problems (GPQA Diamond), top models outperform domain PhDs. This is not pattern-matching; the problems are designed to be Google-proof. The implication is that AI is now a genuine research collaborator, not a transcription tool.
05 — What compounding looks like through 2035
The point of compounding is that small differences in rate produce enormous differences in outcome. Frontier compute has grown 4–5× per year for fourteen years; let's run that forward a decade and see what falls out. Three scenarios — plateau, continuation, and acceleration. The middle one is the trend extrapolation.
All three lines start at GPT-4-scale (~10²⁵ FLOPs) in 2023. Plateau holds at 1.5× / year (Moore's Law alone). Continuation holds at 4× / year (the historical trend). Acceleration assumes 5× / year for five years before tapering — what would happen with a successful research-automation flywheel.
2027
The agentic year
Reliable multi-hour autonomous agents — they file the expense report, write the codebase, book the trip, and only escalate edge cases. Software engineering becomes 5–10× more productive at frontier shops. White-collar entry-level roles begin restructuring.
2030
The scientist year
AI systems generate, run, and interpret novel experiments with minimal human direction. Drug discovery cycle times collapse from years to months. Materials science, fusion research, and climate modeling each see 10× acceleration in iteration speed.
2035
The unknown year
If trends hold, frontier models train on ~10²⁹ FLOPs — the rough scale at which some researchers expect human-equivalent general intelligence. This is where forecasting breaks down. Either we hit a wall, or we don't, and either is civilizationally important.
The honest caveat
Predictions of compounding curves consistently overshoot in the short run and undershoot in the long run. Three things could stop this trajectory: energy (frontier training already uses small-city power; 1 GW clusters are coming), chips (TSMC, ASML, and a handful of fabs are single points of failure), and policy (export controls, compute caps, or a serious safety incident). My base case is that two of these three throttle the curve to ~3× per year by 2030, but none stop it.
06 — What it means & how to position
The most consequential general-purpose technologies — fire, agriculture, the printing press, electricity, the internet — have one thing in common. The people who recognized them early and built around them captured most of the value. The people who were skeptical or waited for clarity captured very little. None of those technologies waited for permission.
AI is on that list. It has been, for at least a decade. The curves above are not speculation; they are measured. The base case isn't maybe AI will be a big deal in fifteen years. The base case — the trend extrapolation, the thing that has happened consistently every year since 2010 — is that frontier intelligence per dollar will be ~1,000× cheaper five years from now than it is today. Plan for that.
For a risk professional, that means three things. First, every risk model that assumes today's labor cost or today's judgment-as-a-bottleneck is wrong on a 5-year horizon. Second, the firms that integrate AI deeply — not as a chatbot, but as a re-architecting of their core processes — will compound advantages that compound advantages. Third, the second-order risks (concentration, model failure modes, systemic prompt injection, regulatory whiplash) are not yet priced into most risk frameworks, and the gap between the firms that price them and the firms that don't will be enormous.
None of this is forecasting. It is bookkeeping on a curve that has held for fourteen years. The only real question is whether you choose to be the person who saw it — and adjusted — or the person who didn't.
Three specific things to do this quarter
Glossary
A floating-point operation — one multiply or add on a fractional number. The unit of computational work for AI training. GPT-4 used roughly 2 × 10²⁵ FLOPs, which is twenty-trillion-trillion arithmetic operations. For context, all human brains alive today, thinking for one second, perform very roughly 10²⁵ operations.
Training is the one-time cost to build a model — months on tens of thousands of GPUs. Inference is the per-query cost to run it — milliseconds on a few GPUs. Training compute has been growing ~4×/year. Inference cost per fixed-quality output has been falling ~10×/year. Both numbers matter; they capture supply (training spend) and demand (inference price).
Empirical observations that model loss falls as a smooth power-law function of parameters, data, and compute. First clearly published by Kaplan et al. (OpenAI, 2020), refined by Hoffmann et al. (DeepMind, 2022). They turned AI from "research" into "engineering" — capability became a budget question.
Three places to spend compute. Pre-training: read the internet, learn statistics. Post-training (RLHF, RLAIF, fine-tuning): turn a base model into a useful assistant. Inference-time: spend more compute per query for harder problems — the o1 / o3 / "thinking" approach. The third is a new axis of scaling, discovered in 2024, and it's still early.
AGI (artificial general intelligence): an AI that matches or exceeds human capability on most cognitive tasks. ASI (artificial superintelligence): an AI substantially more capable than the best humans at essentially all cognitive tasks. Definitions vary wildly; what matters is the underlying capability curve, not the label.
The neural network architecture introduced in the 2017 paper "Attention Is All You Need" (Vaswani et al.). Underlies essentially every frontier AI model today — GPT, Claude, Gemini, Llama. Its key insight (self-attention) lets a model relate every token to every other token, in parallel, scalably. The single biggest research result of the decade.
A lot. Robotics and embodied AI (improving fast, not yet on the same exponential as language). Specific safety and alignment risks (real, deserve their own guide). Geopolitics and the US-China race (probably the most important governance question of the decade). Open-source vs closed-source dynamics. Each of these compounds with the picture above; none change its direction.