AI Training Scale Breaks the 100,000-Run Barrier: The Intelligence-Per-Watt Era Has Arrived

Introduction: When Will AI Truly Change Your Daily Life?

This question once sounded distant and abstract, but in 2025, it is closing in on everyone at an unprecedented pace. The renowned AI industry newsletter Import AI published a series of striking observations in its Issue 435: training run scales have reached the 100,000-run level, AI systems are systematically absorbing decision-making power that once belonged to humans, and "Intelligence per Watt" is becoming an entirely new yardstick for measuring AI development. These three trends are intertwined, painting a clear picture of AI moving from the laboratory to the core infrastructure of society.

Core Trend One: What Do 100,000 Training Runs Mean?

In the world of AI research and development, a "training run" is the fundamental process by which a model learns from data. Over the past few years, the industry has gradually scaled from hundreds to thousands of training runs, and now leading AI laboratories are conducting training runs at the 100,000 (100k) level.

This numerical leap is far more than a simple quantitative change. One hundred thousand training runs mean research teams can conduct systematic searches across vast hyperparameter spaces, model architecture iteration speeds are growing exponentially, and AI companies' investments in computing infrastructure have reached an entirely new order of magnitude. According to industry estimates, a single large-scale training run can cost tens of millions of dollars, and the experimental infrastructure behind 100,000-scale runs is backed by annual R&D budgets in the tens of billions of dollars.

More critically, this scaled approach to training is transforming the methodology of AI R&D itself. In the past, researchers relied on intuition and small-scale experiments to guide model design; now, large-scale parallel training has made "trading compute for insight" the dominant strategy. As one industry insider put it: "We no longer guess which architecture is best — we let 100,000 experiments give us the answer."

Core Trend Two: AI Systems Are Absorbing Human Power

Another profound observation from Import AI Issue 435 is that AI systems are gradually "absorbing" decision-making power previously held by humans. This goes beyond the familiar narrative of automation replacing repetitive labor — it represents a deeper transfer of power.

In finance, AI trading systems already control over 60% of global stock market trading volume. In healthcare, AI-assisted diagnostic systems have surpassed human experts in accuracy for certain conditions. In content, AI-generated text, images, and video are reshaping the entire chain of information production and distribution. Behind each of these scenarios lies a process in which humans cede judgment, decision-making authority, and even creative power to AI systems.

This absorption of power is not always overt. Much of the time, it occurs quietly under the banner of "assistance" and "augmentation." When a hiring manager relies on AI to screen resumes, when a judge references AI sentencing recommendations, when a doctor adopts an AI diagnostic conclusion, the human decision-maker formally retains final authority — but in practice, AI has already become an indispensable link in the decision-making chain, and often the most critical one.

This raises a series of urgent ethical and governance questions: When an AI system makes a wrong decision, who should bear responsibility? When AI judgment diverges from human experts, whom should we trust? When power shifts from distributed human individuals to centralized AI systems, how will society's power structures be reshaped?

Core Trend Three: "Intelligence per Watt" Becomes the New Benchmark

If the first two trends focus on AI's capability boundaries and social impact, then "Intelligence per Watt" zeroes in on the sustainability of AI development.

As model scales continue to expand, the energy consumption of AI training and inference has become a challenge that can no longer be ignored. According to International Energy Agency estimates, global data center electricity consumption could reach 1,000 terawatt-hours in 2025, with AI workloads accounting for a rapidly growing share. Against this backdrop, simply pursuing improvements in model performance is no longer sufficient — the industry is turning its attention to efficiency, specifically how to extract more intelligence from every watt of energy consumed.

The implications of this metric are far-reaching. It is driving innovation at the hardware level — from NVIDIA's Blackwell architecture to various AI-specific chips, energy efficiency is becoming a core consideration in chip design. It is also spurring algorithmic transformation, including the widespread adoption of techniques such as model pruning, quantization, and knowledge distillation, as well as architectural innovations like Mixture of Experts (MoE) models. Most importantly, it establishes a development standard for the AI industry that is directly linked to environmental sustainability.

Analysis: The Convergence and Tensions Among the Three Trends

These three trends share deep internal connections while also harboring significant tensions.

The 100,000-run training regime drives rapid advances in AI capabilities but also intensifies pressure on energy consumption, creating a natural contradiction with the efficiency goals of "Intelligence per Watt." The enhancement of AI capabilities enables these systems to take on more human decisions, thereby accelerating the process of power absorption — yet this power transfer in turn demands higher standards of reliability and explainability from AI systems.

From the perspective of industry competition, companies that can achieve breakthroughs across all three dimensions simultaneously will hold a decisive advantage in the next round of the AI race. Organizations that possess large-scale training capabilities, achieve high energy-efficiency ratios, and build trust in AI governance will be the true winners of the AI era.

Notably, China's AI industry has demonstrated strong competitiveness across all three dimensions. From breakthroughs in training efficiency by models such as DeepSeek, to the steady progress of domestically produced AI chips in energy efficiency, to the gradual refinement of AI governance frameworks, China is forging its own path in AI development.

Outlook: The Tipping Point Where AI Changes Daily Life

Returning to that original question — when will AI truly change your daily life?

The answer may be: it is already happening; many of us simply have not fully realized it yet. When models produced by 100,000 training runs are deployed in phone assistants, search engines, and smart home devices; when AI systems take on ever more decision-making functions in healthcare, education, and finance; when improvements in "Intelligence per Watt" enable AI capabilities to be deployed across a wider range of edge devices — AI's penetration into daily life will transform from a trickle into a flood.

Over the next 12 to 18 months, we are likely to see the following changes accelerate: personal AI assistants evolving from "useful tools" to "indispensable partners"; AI-driven personalized services rolling out comprehensively in healthcare, education, and other fields; and societal discussions around the distribution of AI power expanding from tech circles into the public sphere.

This is not merely a technological revolution — it is a civilizational turning point. At this juncture, what we need is not only more powerful models and more efficient chips, but also deep reflection and prudent governance of the relationship between AI and humanity.