Every time a new AI model goes down (GPT update, DeepSeek, Gemini), huge size, complexity, and increasingly, these large models feel more hungry. Suppose these models are defining the resource requirements of the AI revolution.
This assumption is wrong.
Yes, large models are eager. However, the greatest pressure on AI infrastructure is not from a few Mega models, it comes from the silent diffusion of AI models across industries, each model is fine-tuned for specific applications, and each application consumes computation at an unprecedented scale.
Despite potential winner competition in LLM, the AI landscape is not concentrated and it is scattered. Every business isn’t just using AI, they need to train, customize and deploy private models that meet their needs. It is the latter situation that will create an infrastructure demand curve that cloud providers, businesses and governments are not ready yet.
We’ve seen this pattern before. The cloud has not solidified its workloads; it creates a huge hybrid ecosystem. First, it’s server sprawl. Then the VM spreads. Now? AI spreads. Each wave of calculations leads to diffusion, not simplification. AI is no different.
AI Sprawl: Why AI’s future is one million models, not one
Finance, Logistics, Cybersecurity, Customer Service, R&D – EAST has its own functions to optimize its own AI model. Organizations do not train an AI model to rule their entire operation. They are training thousands of people. This means more training cycles, more computing, more storage requirements, and more infrastructure sprawl.
This is not theoretical. Even in industries where technology has traditionally been cautious, AI investment is accelerating. The 2024 McKinsey report found that organizations now use AI on average three business functions, manufacturing, supply chain and product development leading (McKinsey).
Healthcare is a great example. Navina, a startup that integrates AI into electronic health records, has just raised $55 million in Series C funding from Goldman Sachs (Business Insider). Energy is no different – Industry leaders have launched an open AI alliance to optimize AI for grid and factory operations (Axios).
Calculate pressure no one is talking about
AI is already breaking the traditional infrastructure model. The assumption that the cloud can be infinitely scaled to support AI growth is wrong. AI does not scale like traditional workloads. The demand curve is not gradual, its index is an index, and high standards cannot keep up.
- Power constraints: Currently focusing on power availability, not just network backbone, but on AI-specific data centers.
- Network bottleneck: Hybrid IT environments become difficult to manage without automation, and AI workloads will only intensify.
- Economic pressure: AI workloads can consume millions of dollars in a month, resulting in financial unpredictability.
Data centers already account for 1% of global electricity consumption. In Ireland, they now consume 20% of the national power grid and are expected to rise sharply by 2030 (IEA).
Plus the imminent pressure on the GPU. Bain & Company recently warned that AI growth is driven by explosive demand from data center bargaining chips, laying the foundation for semiconductor shortages.
At the same time, the sustainability of artificial intelligence is becoming increasingly important. Analysis for 2024 Sustainable cities and societies Unless targeted efficiency offsets, the widespread adoption of AI in healthcare can significantly increase the industry’s energy consumption and carbon emissions.
AI spreads more than the market, this is a matter of state power
If you think AI spread is a corporate issue, think about it again. The most important driver of AI decentralization is not the private sector, its governments and military defense agencies deploy AI at scales without high standards or enterprise matching.
The U.S. government alone has deployed AI in more than 700 applications from 27 institutions, covering intelligence analysis, logistics, etc. (FedTech Magazine).
Canada will invest up to $700 million to expand domestic AI computing capabilities, posing national challenges to strengthening sovereign data center infrastructure (innovation, science and economic development Canada).
Moreover, there is a growing call for the “Apollo Project” of AI infrastructure, which is a heightened improvement in AI from commercial advantages to national orders (MIT Technology Review).
Will military AI be effective, coordinate or optimize costs, driven by national security mandates, geopolitical urgency, and the need for closed sovereign AI systems. Even if companies control AI spread, who will tell the government to slow down?
Because when national security is online, no one stops to ask if the grid can handle it.
