The Ages of AI: A Compute Timeline of the Field

The history of artificial intelligence is usually told as a story of ideas. It is better understood as a story of compute. Each major age of AI began when a new way of using computation became possible, and each one accelerated as the compute behind it grew. Reading the timeline this way explains the long winters, the sudden breakthroughs, and why 2025 and 2026 felt like a step change. A second way to scale compute had arrived.

AI’s Real Timeline Is a Compute Timeline

It is tempting to mark AI history by its famous moments. A chess match won, a game of Go lost, a chatbot that suddenly wrote like a person. Those moments are real, but they are symptoms. Underneath each one sits a shift in how much computation a system could bring to bear, and how efficiently it could use it.

Researchers have measured this directly. In a 2022 study of 123 milestone systems, Sevilla and colleagues split machine learning into three compute eras, each with its own growth rate. The pattern is clear. When the way of using compute changed, the field changed with it. The sections below walk that timeline, then look at where the curve points next.

The Symbolic Age: Ideas Without Compute (1950s to 1980s)

The field opened with a question rather than a machine. In 1950, Alan Turing asked whether machines could think, and proposed a test for it. Six years later, the 1956 Dartmouth workshop gave the field its name. The early approach was symbolic, sometimes called good old-fashioned AI. Knowledge was hand-coded as rules and logic, and reasoning meant searching through those rules.

This worked for narrow problems and produced the expert systems of the 1970s and 1980s. It did not scale. Writing every rule by hand was slow, brittle, and unable to handle the messiness of the real world. Compute was scarce and grew slowly, so there was no way to brute-force the gap. Enthusiasm outran results, funding dried up, and the field passed through two AI winters. The ideas were ahead of the machines.

The founders are worth naming. John McCarthy coined the term artificial intelligence at Dartmouth and later created the Lisp programming language. Marvin Minsky, Claude Shannon, and others shaped the agenda alongside him. At Carnegie Mellon, Allen Newell and Herbert Simon built the Logic Theorist in 1956. It is often called the first program to prove mathematical theorems, and the General Problem Solver followed a year later. In 1958, Frank Rosenblatt built the Perceptron, the first trainable artificial neural network. That seed would not fully flower for another fifty years.

Where AI Was Born: The First Labs

AI was a global pursuit from the start. In the United States, John McCarthy and Marvin Minsky founded the MIT Artificial Intelligence Project in 1959. McCarthy then moved west to start the Stanford Artificial Intelligence Laboratory in 1963, while Carnegie Mellon grew around Newell and Simon. These three became the early centers of the field.

Europe was not far behind. At the University of Edinburgh, the wartime code-breaker Donald Michie built what many call Europe’s first AI research group in the 1960s. Edinburgh became the second university in the world, after Stanford, to teach the subject. In the Netherlands, the roots ran through Amsterdam. The Centrum Wiskunde en Informatica, founded there in 1946, became one of Europe’s oldest computing institutes and the cradle of Dutch computer science. The city remains a serious AI hub today.

Japan made the boldest national bet. In 1982, its Ministry of International Trade and Industry launched the Fifth Generation Computer Systems project. The ten-year program aimed to build machines designed around logic and parallel reasoning. It was ahead of its time and ultimately fell short of its goals, a reminder that ambition without the right compute and methods runs into a wall. That lesson would repeat.

The Statistical Turn: Learning From Data (1990s to 2000s)

The next age replaced hand-written rules with learning from data. Instead of telling a system the rules, researchers let it infer patterns from examples using statistics. Methods such as support vector machines and statistical language models pushed the field forward, and probability replaced rigid logic as the core tool.

Compute still grew at the steady pace of Moore’s law during this period, roughly doubling every 20 months. That set the speed limit. Progress was real but incremental, because the amount of computation a researcher could apply rose only as fast as general-purpose chips improved. The field was learning to learn, but it was still waiting for a faster engine.

The Deep Learning Era: GPUs Change Everything (2012)

The engine arrived in 2012. A neural network called AlexNet, built by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, won the ImageNet image-recognition challenge by a wide margin. It cut the top-5 error rate to 15.3 percent, against 26.2 percent for the runner-up. The result was striking, but the method was the real news. AlexNet was trained on two consumer NVIDIA GPUs.

That detail changed the field. GPUs could run the matrix math behind neural networks in parallel, far faster than ordinary processors. Deep learning was not a new idea in 2012, but for the first time there was enough affordable compute to make it work at scale. The effect on the growth curve was immediate. Sevilla and colleagues date a new era from around this point, with training compute now doubling roughly every six months rather than every 20.

OpenAI measured the same acceleration from a different angle. In a 2018 analysis, it found that the compute used in the largest training runs had doubled about every 3.4 months since 2012. That was a more than 300,000-fold increase over six years. For comparison, Moore’s law would have delivered roughly a sevenfold rise in the same span. Compute, not just chips, had become the engine of progress.

The breakthrough rested on data and on decades of patient work, not compute alone. The ImageNet dataset, created by Fei-Fei Li and collaborators at Stanford, gave the field a benchmark large enough to prove deep learning at scale. Behind AlexNet stood years of neural-network research by Geoffrey Hinton, Yann LeCun, and Yoshua Bengio. In 2018 the three shared the Turing Award as the godfathers of deep learning. The idea had simply waited for the compute and the data to catch up.

The Age of Scale: Transformers and Scaling Laws (2017 to 2022)

The deep learning era found its defining architecture in 2017, when Vaswani and colleagues introduced the transformer. Its attention mechanism processed sequences in parallel and scaled gracefully with both data and compute. Almost every large model since has been built on it.

The transformer arrived just as a new question came into focus: what happens if you simply make models bigger? In 2020, Kaplan and colleagues published scaling laws showing that model performance improves predictably as a power law of compute, parameters, and data. Two years later, Hoffmann and colleagues refined the recipe with the Chinchilla result, showing that compute is used best when model size and training data grow together. Scaling stopped being a hope and became something close to an equation.

This was the large-scale era that Sevilla’s team identified. Leading labs began training models with 10 to 100 times the compute of the deep learning trend that came before. The result was the modern wave of large language models. Their capabilities followed the curve the scaling laws predicted, which is why the period felt like a series of step changes rather than gradual gains. We traced the economics of running these models in AI Inference Costs at Scale.

The Reasoning Era: A Second Axis of Compute (2024 to 2026)

For about a decade, the recipe for a better model was to train a larger one. In 2024 and 2025, a different idea took hold. Instead of only scaling the compute used to train a model, you could scale the compute a model uses to think at the moment it answers. This is test-time, or inference-time, scaling.

OpenAI’s o1, released in 2024, was the first widely used model to do this, extending its chain of thought to reason for longer before responding. In January 2025, DeepSeek released R1 and showed that pure reinforcement learning could produce comparable reasoning. Its R1-Zero variant reached a pass@1 score of 86.7 percent on the AIME 2024 mathematics benchmark. Work from DeepMind confirmed the deeper point, that the scaling laws first found for training also apply to inference.

This is why 2025 and 2026 felt like a step change. A second axis of scaling had opened. Progress no longer depended only on building bigger models, which is slow and expensive. It could also come from spending more compute per query at inference, on hardware already deployed. The two axes compound, and total demand for compute rose accordingly. The shift also moved the center of gravity toward inference, which we examined in why energy is overtaking hardware as the defining input.

Where the Curve Points Next

The frontier is now two-dimensional. Labs scale training runs and inference compute at the same time, so total demand keeps climbing even as individual models grow more efficient. That has a consequence. The binding constraint is shifting away from raw chips and toward the things that feed them.

Memory is the first wall. Research on the memory wall shows that compute has outgrown memory bandwidth for years, so modern accelerators are often starved for data rather than arithmetic. We covered how this reshapes model design in why mixture-of-experts makes the GPU shortage worse. Energy is the larger wall. As demand grows, the question is less whether models can scale. It is more whether the power and grid connections exist to run them, a theme we set out in the data center energy constraints piece.

The through-line across every age holds. Compute has been the defining input of AI since the symbolic systems that lacked it, and it remains so today. What changes from era to era is the form the constraint takes. It was rules in the symbolic age, chips in the deep learning era, and now memory, packaging, and power. The next age of AI will belong to those who can secure and use compute efficiently, wherever it is needed.

2027 to 2030: The Next Age

What do experts expect through the rest of the decade? On raw scaling, the answer is cautious optimism. Epoch AI examined whether the recent pace, roughly a fourfold increase in training compute each year, can hold to 2030. Its conclusion is that training runs of around 2e29 operations will likely be feasible by then. That is the leap from GPT-2 to GPT-4, applied once more. The limit is not imagination. Epoch ranks four constraints, and power is the first to bind, followed by chip manufacturing, then data and latency.

Data is the second pressure. Human text suitable for training is finite. Epoch projects that the usable stock could be largely consumed between about 2027 and 2028 at compute-optimal scale, and sooner if models are overtrained. That points to a shift already under way. Models will lean on synthetic data, on multimodal sources such as video and audio, and on far more efficient use of the data that exists. The age of simply scraping more of the web is ending.

The next age also has a name forming around it: the age of agents. Building on reasoning models, AI is moving from answering questions to taking actions, using tools and completing multi-step tasks with limited supervision. Gartner expects roughly a third of enterprise software to include agentic AI by 2028, up from under one percent in 2024. It also expects around 15 percent of routine work decisions to be made autonomously. The same analysts add a sober note. More than 40 percent of agentic AI projects may be cancelled by the end of 2027, as cost and reliability catch up with the hype. Every age has had its winter, and agents will not be exempt.

Put the forecasts together, and a picture forms for 2027 to 2030. Scaling continues, but on a tighter leash held by energy and chips. Models rely more on synthetic and multimodal data. The frontier shifts from training ever-larger models toward running ever-more-capable agents, which operate constantly and multiply inference demand. In every version of this future, the binding input is the same. It is compute, and increasingly the power to run it.

What This Means for Building AI Today

For anyone building in AI now, the timeline offers a practical lesson. Each age rewarded the teams that could apply compute the moment it became useful. The reasoning era raises the stakes, because inference now scales as well as training. Sustained, well-used capacity matters more than ever, and access to it is becoming the real differentiator.

That is the gap Axe Compute is built to close. Deploying on infrastructure that already exists, rather than waiting years for new builds, lets teams reach compute when a workload needs it. It also puts that compute in the regions where it needs to run. The planning side of that decision, across training and inference, is covered in Enterprise GPU Strategy in 2026. The ages of AI have always turned on who could harness compute. That has not changed, and it is not about to.

Frequently Asked Questions

What are the main eras or ages of AI?

Read through the lens of compute, AI moves through five rough ages. First, the symbolic age (1950s to 1980s) ran on hand-coded rules. Next, the statistical age (1990s to 2000s) learned from data at the pace of Moore’s law. The deep learning era then began in 2012 with AlexNet and GPUs. After that came the age of scale (2017 onward), with transformers and scaling laws. Most recently, the reasoning era (2024 to 2026) added test-time compute, where models think longer at inference. Each age is defined by how fast training compute grew.

Why did AI accelerate so sharply in 2025 and 2026?

Because a second scaling axis opened. For a decade, progress came mainly from training larger models on more data, with the largest training runs doubling their compute roughly every few months. In 2024 and 2025, reasoning models such as OpenAI’s o1 and DeepSeek-R1 showed that spending more compute at inference time, letting a model reason for longer, also improves results. That added inference scaling on top of training scaling, multiplying total demand for compute.

How fast has the compute used to train AI grown?

Very fast, and faster than hardware alone. Before 2010, training compute grew roughly in line with Moore’s law, about doubling every 20 months. In the deep learning era it doubled roughly every six months. OpenAI found that between 2012 and 2018 the compute in the largest training runs doubled about every 3.4 months, a more than 300,000-fold increase over that period.

What comes after the reasoning era of AI?

Most forecasts point to an age of agents through 2027 to 2030, with autonomous, tool-using AI built on today’s reasoning models. The frontier stays two-dimensional, scaling both training and inference, so total compute demand keeps rising. Epoch AI expects scaling can likely continue to around 2030, with power the first constraint to bind, followed by chips and data. The next age will be shaped less by whether models can scale and more by whether the energy and infrastructure exist to run them. Compute remains the defining input.

Sources

• Sevilla et al., “Compute Trends Across Three Eras of Machine Learning,” arXiv:2202.05924 (2022): pre-2010 ~20-month doubling, deep learning era ~6-month doubling, and a distinct large-scale era from ~2015-2016
• OpenAI, “AI and Compute” (2018): largest training runs doubled ~every 3.4 months from 2012 to 2018, a >300,000x increase
• Vaswani et al., “Attention Is All You Need,” arXiv:1706.03762 (2017): the transformer architecture
• Kaplan et al., “Scaling Laws for Neural Language Models,” arXiv:2001.08361 (2020)
• Hoffmann et al., “Training Compute-Optimal Large Language Models” (Chinchilla), arXiv:2203.15556 (2022)
• DeepSeek-AI, “DeepSeek-R1,” arXiv:2501.12948 (2025): reinforcement-learning reasoning; R1-Zero reaches 86.7% pass@1 on AIME 2024
• AlexNet (2012): ImageNet top-5 error of 15.3% vs 26.2%, trained on two NVIDIA GPUs (Krizhevsky, Sutskever, Hinton)
• Gholami et al., “AI and Memory Wall,” IEEE Micro / arXiv:2403.14123 (2024)
• Epoch AI, “Can AI Scaling Continue Through 2030?” (2024): ~4x/year pace, 2e29 FLOP runs likely feasible by 2030, with power the first binding constraint, then chips, data, and latency; data wall projected around 2027-2028
• Gartner: agentic AI forecasts, including ~33% of enterprise software featuring agentic AI by 2028 and 40%+ of agentic projects cancelled by end of 2027