Originally published on Medium.
Imagine you’re trying to build the smartest robot brain ever. Every time you make its brain cells (computer chips) twice as smart, they also get twice as hungry for food (electricity) and produce twice as much heat.
Introduction: The AGI Dream: Promise vs. Reality
The idea of Artificial General Intelligence (AGI) — machines that can think, learn, and adapt like humans across many different tasks — is incredibly exciting. It promises to change our world in ways we can barely imagine. However, reaching AGI isn’t just a software challenge; it comes with some very real, physical hurdles, especially concerning the massive amounts of computing power and energy it would need. As the saying goes, “Compute & Energy Chips” are hotter than rockets; AGI could gulp terawatt-hours without breakthroughs.” This isn’t just a minor technical issue; it’s a fundamental roadblock that we need to understand and address.
Even today’s advanced AI models consume huge amounts of electricity and generate a lot of heat, pushing our current technology to its limits. AGI, which would be constantly learning and adapting, is expected to need far, far more. This raises big questions about whether we can actually sustain such a future. This article will explore these compute and energy challenges, explain what they mean for AGI, and look at some of the clever solutions scientists are working on.
What is Artificial General Intelligence?
AGI is a theoretical concept about creating AI that has human-like intelligence. This means it could teach itself new things and perform tasks it wasn’t specifically programmed for. Unlike the AI we use today, which is very good at specific jobs (like recognizing faces or translating languages), an AGI system would be able to solve complex problems in many different areas, much like a human. It’s about replicating human thinking abilities, including being able to control itself, understand itself, and learn new skills in new situations.
How AGI Differs from Current AI (Like ChatGPT)
Today’s most impressive AI, including the large language models (LLMs) like ChatGPT, are forms of “narrow AI.” They are excellent at their specific tasks, whether it’s generating text, understanding speech, or identifying objects. But they can’t suddenly switch to a completely new, untrained area on their own.
For example, while ChatGPT can write amazing articles, it still relies on humans to set up the problem, design its core structure, and select the data it learns from. It’s great at “well-structured problems” (where the rules are clear), but it needs human help to simplify “ill-structured problems.” This dependence on human input means that current AI models, despite their power, aren’t truly general. To reach AGI, we need a major shift in how AI systems learn and adapt independently. Simply making current models bigger won’t be enough; it will just make the energy problem worse if the basic approach isn’t efficient for true general intelligence.
The Current AI Energy Footprint: A Glimpse into the Future
How Much Power Do Today’s AI Models Use?
The current generation of large AI models already uses a lot of energy. For instance, training a model like GPT-3 consumed between 324 to 1,287 MWh (megawatt-hours) of electricity. To put that in perspective, that’s roughly the same amount of electricity used by 120 US homes in a whole year. Newer models like GPT-4 and Claude 3 Opus are even larger, likely needing 2 to 4 times more energy to train.
But it’s not just about training. The daily use of these models, called “inference,” is quickly becoming the biggest energy user. A single query to ChatGPT, especially for advanced models like GPT-4o, uses about 0.43 watt-hours (Wh) of energy. This is ten times more than a typical Google search. With millions of queries happening every day, the total energy used for inference is now greater than that used for training. This is a big change. While training is a huge, but one-time, energy cost, inference is a continuous and growing demand that increases as more people use AI. For AGI, which is imagined as being “always-on” and constantly adapting, the total energy cost of its ongoing operation will be far greater than its initial training. This means we need to focus on making AI efficient for continuous use, not just for the initial setup.
The Self-Accelerating Loop: AI Demanding More AI
A worrying trend is that AI is now being used to design more efficient chips. As AGI develops, it could learn to re-train itself, control advanced computers to optimize its own models, and even design better hardware. This creates a cycle: “More AI → More Compute → More Energy → More Emissions.” This means the energy demand isn’t just growing linearly; it could grow exponentially and become self-perpetuating. If AGI can improve its own hardware and algorithms, its energy needs could increase much faster than human innovation can predict or control. This makes it urgent to develop sustainable solutions before AGI reaches a point where it can rapidly improve itself, as the problem could quickly become too big to manage.
AGI’s Insatiable Appetite: Projections and the Scaling Challenge
How Much Compute Would Human-Level AGI Need?
Estimates for the computing power needed for human-level AGI vary a lot. Some researchers think a powerful gaming PC (like an Nvidia GeForce RTX 4090, with about 8.3e13 FLOP/s) might be enough for human-level AGI. This is based on comparing AGI to the human brain’s computational power.
However, the median forecast from 108 AI experts is much higher, at 3e17 FLOP/s (floating point operations per second). This is orders of magnitude more powerful than current consumer hardware.
The human brain is incredibly efficient at learning. To “train” a human to be an engineer takes about 20 times the compute of one year of their work. Compare this to current AI language models, where this factor can be “several hundred million” times higher! This huge difference in efficiency, where the human brain learns and operates with far less training overhead, points to a big “algorithmic efficiency gap” between biological and artificial intelligence. If AGI is truly general and always learning, it can’t afford the current AI paradigm’s training inefficiency. This means we need fundamental breakthroughs in algorithms, perhaps inspired by how the brain works, just as much as we need raw hardware power.
Projected Energy Demands for AGI-Scale Models
If current trends continue and the “scaling hypothesis” proves true, AGI-scale models are expected to be “power devourers.” They will likely need constant computation, real-time adaptation, and much larger models trained with more data and longer cycles. Forecasts suggest AGI-scale models could use 10 to 100 times more energy than current Generative AI.
Data center projections really show the scale of this challenge: In 2023, AI-related data centers used about 4.5 GW (gigawatts) globally. This is expected to jump to 14–18.7 GW by 2028, potentially making up 20% of all data center energy use. All global data centers together used about 415 TWh (terawatt-hours) in 2024 (about 1.5% of global electricity). This is expected to more than double by 2030, possibly reaching 1,000–1,300 TWh. The International Energy Agency (IEA) predicts that data centers’ total electricity use could exceed 1,000 TWh by 2026, which is roughly equal to Japan’s entire annual electricity consumption. In the United States, AI-driven data centers could account for almost half of the growth in electricity demand by 2030. This comparison to a whole country’s energy use shows that this isn’t a small increase, but a huge national-level energy requirement that will put immense strain on power grids worldwide. This means the energy challenge for AGI isn’t just a tech problem; it’s a global infrastructure and geopolitical issue.
Conclusion: Navigating the Energy Crossroads to AGI
While an individual’s daily use of ChatGPT uses very little electricity, the combined effect of widespread AI adoption is huge. The total daily queries for just one advanced model like GPT-4o already use as much electricity annually as tens of thousands of U.S. homes, and this includes all the significant energy used by data center infrastructure. Looking ahead, the growth of AI-driven data centers is projected to double global data center electricity consumption by 2030. This increasing demand poses major challenges for our energy grids.