Does AI Inevitably Lead to Power Concentration? 125 Years of Computing Efficiency History Offers a Powerful Counterargument

A Popular Argument and Its Hidden Flaw

In recent years, as training large models routinely consumes thousands of GPUs and tens of millions of dollars in compute costs, a viewpoint has become widespread in the tech world: AI structurally and inevitably leads to the concentration of power. The logic goes that since only giants with massive computing resources can train the most powerful models, the benefits and control of AI will inevitably concentrate in the hands of a few tech companies and nations.

This argument seems perfectly logical, yet it rests on a hidden premise — that AI's enormous demand for compute is permanent and unchanging. And this is precisely the assumption least able to withstand scrutiny. This line of reasoning ignores 125 years of super-exponential growth in computing cost-performance, mistaking a dynamically evolving technological phase for a static endgame.

What 125 Years of Super-Exponential Growth in Computing Cost-Performance Means

The improvement in computing cost-performance didn't begin with Moore's Law. If we trace computing back through mechanical calculators, relays, vacuum tubes, transistors, and integrated circuits, the computing power available for the same cost has increased by dozens of orders of magnitude over the past century-plus. This curve is not merely exponential — some researchers argue it is super-exponential — meaning the rate of growth itself is accelerating.

Starting with mechanical computing devices in the late 19th century, human computing capability has gone through five overlapping paradigms: mechanical computation (such as Babbage's Difference Engine), electromechanical relays (such as the Harvard Mark I), vacuum tubes (such as ENIAC), transistors, and integrated circuits. Each paradigm picked up the baton as the previous one's growth slowed, forming a continuous growth curve spanning 125 years. Futurist Ray Kurzweil systematically traced this curve in The Singularity Is Near, finding that the computing power purchasable per dollar roughly doubles every few years, and that this doubling rate itself is accelerating — meaning growth is not merely exponential (doubling at a fixed rate) but super-exponential, with the doubling period shortening over time.

This means that AGI-level capabilities requiring an entire data center to run today could very well be compressed into smaller, cheaper devices in the foreseeable future. A pattern that has repeated throughout history is this: any computing capability that initially depends on massive centralized infrastructure eventually becomes ubiquitous and decentralized. In the 1960s, IBM's System/360 mainframes required entire rooms to house and cost millions of dollars — only large enterprises, governments, and research institutions could afford them. The prevailing view at the time was that computing would always be a centralized service. Yet the invention of the microprocessor (Intel 4004 in 1971) spawned the personal computer revolution in less than fifteen years. By 2007, the iPhone put a processor exceeding the total computing power of the entire Apollo moon landing program into everyone's pocket. The computing power of the mainframe era now sits in everyone's pocket.

The assumption that "AGI will always require data centers to run" runs completely counter to the direction of technological evolution over the past 125 years.

Scaling Law Is Not a Law of Physics: The Limitations of Empirical Curves

Another pillar supporting the "compute concentration thesis" is the currently much-revered Scaling Law. People widely believe that model performance improves predictably with increases in parameter count, data volume, and compute, making "bigger is better" an industry consensus.

But a critical clarification is needed here: Scaling Law is not a law of physics.

The concept of scaling laws in AI was primarily systematized by OpenAI's Jared Kaplan and colleagues in their 2020 paper Scaling Laws for Neural Language Models. The study found that under the Transformer architecture, a model's cross-entropy loss exhibits an approximate power-law relationship with model parameter count, training data volume, and training compute — i.e., loss ≈ C/N^α, where N is the parameter count and C and α are constants. This finding directly gave rise to the industry belief that "brute force works miracles."

However, scaling laws are merely empirical relationships observed under specific architectures, objective functions, datasets, and other conditions. Change any one of these factors, and you get a different scaling curve. The power-law relationship itself depends on specific experimental conditions: the Transformer's self-attention mechanism, the autoregressive next-token prediction objective, and training corpora dominated by internet text. DeepMind's subsequent Chinchilla research revised the optimal ratio between parameter count and data volume, demonstrating that scaling curves are not immutable. Furthermore, scaling laws describe the trend of training loss reduction, not a linear improvement in downstream task performance — the mapping between the two is complex and nonlinear.

In other words, the scaling curves we observe today are characteristics of the current generation of machine learning methods, not intrinsic laws of intelligence itself. They describe the patterns of a specific combination: "using the Transformer architecture, with a next-token prediction objective, on internet text data." Once the architecture, training objective, or data undergoes fundamental change, the entire curve could be rewritten.

Mistaking an empirical curve for a universal law is the deep cognitive error underlying the claim that "AI inevitably leads to concentration."

The Lesson of the Brain's 20 Watts: An Existence Proof for Efficient Intelligence

If you need an existence proof for "efficient intelligence," the human brain is the most compelling evidence.

The human brain consumes only about 20 watts of power — roughly equivalent to an energy-saving light bulb — yet achieves general intelligence encompassing language, reasoning, creativity, and planning that the most advanced AI systems today still cannot fully match. By comparison, the energy required to train and run large models exceeds this by several orders of magnitude.

The human brain has approximately 86 billion neurons, transmitting information through roughly 100 trillion synaptic connections. Unlike the von Neumann architecture of digital computers, the brain employs a massively parallel, event-driven, analog-digital hybrid computing paradigm. Neurons fire at extremely low frequencies (typically only a few dozen times per second), far below the GHz-level clock frequencies of modern CPUs, but the brain achieves extreme energy efficiency through extraordinarily high parallelism and sparse activation (only about 1%-5% of neurons are active at any given moment). Moreover, the brain's synaptic plasticity enables "storage" and "computation" to occur on the same physical substrate, avoiding the enormous energy bottleneck of data movement in von Neumann architectures. The field of Neuromorphic Computing is attempting to draw on these principles — chips like Intel's Loihi and IBM's TrueNorth aim to perform neural-like computation at energy efficiency levels approaching the brain's.

This leads to a disruptive inference: the current Scaling Law may not be a "feature" but a "bug."

If the brain serves as a reference point, truly capable AGI is likely to be very efficient. The current scaling laws reveal how inefficient the machine learning algorithms we've discovered so far are, not how much compute intelligence itself must consume.

In other words, the reason we need such enormous compute today is precisely because existing algorithms are extremely inefficient. Once true general intelligence is achieved, its computational requirements will likely be far lower than current expectations. Inefficiency is a temporary technological phase, not the destination.

Reassessing the "AI Power Concentration" Thesis

Taken together, the fragility of the claim that "AI structurally and inevitably leads to power concentration" stems from treating the current technological status quo as a permanent structural fact. Two independent downward forces are acting simultaneously:

• At the hardware level: Computing cost-performance continues to improve at a super-exponential rate — yesterday's supercomputer is tomorrow's edge device.
• At the algorithm level: The inefficiency of existing algorithms implies enormous room for optimization — more efficient architectures will dramatically lower the compute threshold needed to achieve equivalent capabilities.

On the algorithm front, efficiency improvements in recent years have been far more significant than most people realize. Knowledge Distillation can "compress" large model capabilities into smaller models; Mixture of Experts (MoE) dramatically reduces the computation needed per inference through sparse activation; algorithm-level optimizations like FlashAttention reduce the memory overhead of attention mechanisms from quadratic to near-linear; quantization techniques (such as GPTQ and AWQ) compress model weights from 32-bit floating point to 4-bit integers, reducing compute and storage requirements by several fold with almost no loss in performance. Since 2024, teams like DeepSeek have further demonstrated that with more sophisticated training strategies, one can achieve performance levels close to leading models with far smaller compute budgets. All of these advances indicate that a large amount of "inefficient redundancy" in current compute requirements can be systematically eliminated.

The combined force of these two trends points in precisely the opposite direction from "centralization" — together, they drive AI capabilities toward smaller, cheaper, and more distributed forms.

Of course, this doesn't mean that in the short term, there are no competitive advantages and governance challenges arising from compute concentration. At the current stage, institutions with access to large-scale compute do possess first-mover advantages. But to conclude from this that AI "structurally" must concentrate power is to freeze a dynamic technological snapshot into an immutable historical verdict.

Conclusion: From Efficiency Breakthroughs to Institutional Design

The most common mistake in technology forecasting is extrapolating far into the future using today's constraints. Compute hunger is a phase-specific characteristic of this generation of AI, not an essential property of intelligence. Using 125 years of computing evolution as a mirror and the brain's 20 watts as a benchmark, we have ample reason to believe that truly mature AGI will trend toward efficiency, rather than endlessly depending on data centers.

For those concerned about AI's societal impact, rather than worrying about an "inevitable concentration" built on a false premise, it would be better to focus attention on how to accelerate breakthroughs in algorithmic efficiency, and how to build the institutions and ecosystems that ensure efficient AI truly benefits everyone.