The AI Revolution Is Underhyped: Why Even Insiders Underestimate What Comes Next

01 The Scaling Hypothesis

The single most important discovery in artificial intelligence over the past decade is not a new algorithm or a novel architecture. It is the empirical observation that neural networks improve predictably as you increase three things: the number of parameters, the amount of training data, and the compute budget. This relationship, known as the scaling hypothesis, was first documented by researchers at OpenAI in 2020 and has held with remarkable consistency across model generations and laboratories.

The implication is profound. If capability scales with resources, then AI progress is not waiting for a breakthrough. It is waiting for money, electricity, and silicon. Every time someone doubles the compute used to train a model, the model gets better by a measurable amount. This is why the largest AI companies are spending tens of billions of dollars on data centers, and why Nvidia's market capitalization has reached levels that would have seemed absurd five years ago. The scaling hypothesis turned AI development from a research question into an engineering problem, and engineering problems get solved by throwing resources at them.

02 From Chatbots to Agents

The first wave of large language models, from 2020 through 2023, demonstrated that AI could generate fluent text, answer questions, and write code. The second wave, now unfolding, is about agency. An AI agent does not just respond to a prompt; it plans a sequence of actions, executes them, observes the results, and adjusts. This shift from passive generation to autonomous execution is the difference between a tool that helps you write an email and a tool that books your flight, orders your groceries, and files your taxes without step-by-step instructions.

Agentic systems build on the same foundation models but add scaffolding: memory, tool use, planning loops, and the ability to call external APIs. The technical challenge is reliability. A language model that occasionally hallucinates a fact is tolerable in a chatbot. An agent that occasionally executes the wrong action in a real system is not. Solving this reliability problem is the central engineering challenge of the current moment, and the approaches under development, from reinforcement learning to constitutional AI to formal verification, represent some of the most active research in computer science.

03 The Compute Bottleneck

If AI capability scales with compute, then the limiting factor is not intelligence but infrastructure. Training a frontier model now requires tens of thousands of GPUs running for months, consuming enough electricity to power a small city. The cost of a single training run has crossed $100 million and is expected to reach $1 billion within a few years. This creates a dynamic where only a handful of organizations can afford to train the most advanced models, concentrating power in ways that have no precedent in software history.

The bottleneck extends beyond GPUs. Data center construction, power procurement, cooling systems, and the supply of training data are all constraining the pace of progress. Some of the largest AI companies are now negotiating directly with utility providers for dedicated power plants, and there is serious discussion about whether nuclear energy will be needed to meet the electricity demands of future AI training runs. The compute bottleneck is not a temporary problem; it is a structural feature of the scaling era.

Model parameters have grown by roughly 10x every two years, following the scaling hypothesis

04 Economic Disruption and Productivity

The economic impact of AI is already visible in specific sectors. Software development has been transformed by coding assistants that can generate, review, and debug code. Customer service is being automated at scale, with AI agents handling complex multi-turn conversations that previously required human operators. Content creation, from marketing copy to video production, is being reshaped by generative tools that reduce the cost of creative work by orders of magnitude.

But these are the early applications. The deeper transformation comes when AI systems can perform cognitive tasks that currently require specialized human expertise: legal research, medical diagnosis, financial analysis, scientific literature review. The productivity gains from automating even a fraction of these tasks could be enormous, but the distributional effects are uncertain. Workers whose cognitive skills were previously protected from automation may face displacement, while those who can effectively integrate AI tools into their workflows may see significant productivity gains.

05 The Alignment Problem

As AI systems become more capable, ensuring that they behave in ways aligned with human values becomes harder, not easier. A system that can plan and execute complex actions can also plan and execute harmful ones, whether intentionally or as a side effect of misaligned objectives. The alignment problem, as it is known in the field, asks how we can specify objectives that capture what we actually want rather than what we literally say.

The challenge is that human values are complex, context-dependent, and sometimes contradictory. An AI system optimizing for engagement might amplify polarizing content. A system optimizing for accuracy might refuse to make useful but uncertain claims. Current approaches to alignment, from reinforcement learning from human feedback to constitutional methods that encode principles of behavior, are promising but incomplete. The field acknowledges that alignment techniques that work for current models may not scale to future, more capable systems.

Training compute has grown by roughly 4-5x per year, far outpacing Moore's Law

06 The Geopolitical AI Race

The United States and China are engaged in a competition for AI dominance that will shape the coming decades. The United States holds the lead in frontier model development, with OpenAI, Google, Anthropic, and Meta producing the most capable systems. China has invested heavily in domestic AI capability, with companies like DeepSeek, Alibaba, and ByteDance producing increasingly competitive models, often at lower cost. The competition is not just about capability; it is about influence, economic power, and national security.

Export controls on advanced GPUs and lithography equipment, restrictions on data flows, and investments in domestic chip manufacturing are all instruments in this competition. The risk is that the internet and the AI ecosystem fragment along geopolitical lines, with incompatible standards, separate model ecosystems, and reduced collaboration. The outcome will determine not only which countries lead in AI but also what values and constraints are embedded in the systems that increasingly mediate human interaction and decision-making.

07 Why the Trajectory Is Underhyped

Eric Schmidt's argument that the AI revolution is underhyped rests on a simple observation: the rate of improvement is accelerating, not decelerating. Each generation of models is better than the last by a larger margin. The capabilities demonstrated by frontier models in 2025 were considered years away by most experts in 2022. The trend shows no sign of plateauing, and the infrastructure investment now underway, in data centers, custom silicon, and power generation, suggests that the next five years will see capabilities that current observers find difficult to imagine.

The reason even insiders underestimate AI is that human intuition is calibrated for linear change. We expect tomorrow to be slightly better than today, and we extrapolate trends as straight lines. But exponential growth means that the change between year five and year ten is vastly larger than the change between year one and year five. If the scaling hypothesis continues to hold, and there is no evidence yet that it will not, then the AI systems of the early 2030s will make the systems of 2025 look primitive. That is the meaning of underhyped: not that we overestimate the present, but that we systematically underestimate the future.