Published on Thu Nov 06 2025 00:00:00 GMT+0000 (Coordinated Universal Time) by Orkid Labs

Computing is Energy

The Hidden Thermodynamic Constraint

A modern data center is, at its core, a giant bank of servers that consumes vast power and must dump an equal amount of heat. This isn’t a design flaw or an engineering challenge to be solved. It reflects a basic physical law: any information processing is literally an energy flow. Every bit flipped or erased costs energy—about 2.9×10⁻²¹ joules per bit at room temperature—and generates heat.

As one analyst put it: “Computation isn’t free: it costs energy and produces heat.”

Though the per-bit energy is minuscule, modern AI and blockchain systems process billions of bits per second. The result: even today’s supercomputers use about 10⁻¹³ joules per bit—roughly 200 million times the thermodynamic minimum. Large-scale computations burn huge power and heat.

The Core Insight: Treat Compute Like an Energy Plant

The thermodynamic framing delivers a clear takeaway: plan compute infrastructure like energy infrastructure. Doubling computing capacity roughly doubles electrical draw and heat output.

The International Energy Agency projects that data-center electricity use will nearly double from ~415 TWh in 2024 (~1.5% of global power) to ~945 TWh by 2030—about Japan’s entire annual consumption. This is largely due to AI workloads.

In practical terms: running large AI models or blockchain farms means building new power plants and chillers. A single rack of GPUs can dissipate 50–100 kW of heat—far beyond what ordinary air cooling can handle. Firms now install exotic liquid-cooling systems or immersion tanks.

Put bluntly: if you flip on more computer “engines,” you must also build bigger generators and radiators. An AI cluster drawing megawatts will hit grid, cooling, and carbon limits just like any heavy industry.

The Risks of Ignoring Physical Limits

Grid bottlenecks. Building new data centers without adequate power or cooling hookups leads to delays or outages. An estimated 20% of planned centers could be held up by lack of grid capacity. Regions that concentrate data centers can strain local grids, risking brownouts.

Higher emissions and costs. More compute without efficiency means more power from any source available. Even if renewables supply much new capacity, forecasts warn we’ll end up extending reliance on gas plants to meet AI demand. Bitcoin mining alone once consumed on the order of 500 TWh per year—more than many countries—burning hundreds of kWh per single transaction.

Reliability hazards. Heat crashes hardware. A high-powered GPU chip heated beyond its tolerance can fail in seconds. For critical systems (healthcare AI, finance, defense), even a brief cooling failure could cause mission-critical outages. Cooling systems become safety systems: a cyber-attack or fault that disables chillers could instantly fry expensive hardware.

Scaling ceiling. Physical limits loom as we scale up. Efficiency improvements are slowing: where once computation per joule doubled every ~1.5 years (a 100× gain per decade), today it’s roughly every 2.6 years (only ~16× per decade). Even exotic ideas like optical or reversible computing would only raise the floor a bit. In the near term, every extra unit of computing still means extra heat. An AI system that tries to run twice as fast or with twice as many chips will, without extraordinary cooling, literally risk “frying its circuits.”

The Metaphor: Engines and Exhaust

Think of a supercomputer as a car engine. You can make the engine more efficient, but you can’t violate the laws of physics: every fuel burn produces exhaust heat. Landauer’s principle is like saying “no engine is 100% efficient—every bit of work creates a bit of waste.”

A quip from the field: training today’s AI can feel like “boiling the oceans” of energy. In other words: massive computing makes massive heat.

Another metaphor: every bit erased is like burning one match. One match hardly warms a room, but millions of matches certainly do. At global scale, our data “match list” is astronomical.

Here’s a simple example: OpenAI reported that generating about 100 pages of text with GPT-3 consumed roughly 0.4 kWh of electricity—about 2.5 Wh per page. That’s the energy of keeping a 40 W light bulb on for 10 hours. Now imagine millions of users chatting with AI. It adds up.

The Bottom Line: Power Up, Cool Down—or Pay the Price

Computation is fundamentally tied to energy. The thermodynamic balance equation—power in equals heat out plus stored energy—reminds us that you cannot endlessly scale processing without scaling electricity and cooling.

A strategy blind to this will face surprise costs, equipment failures, or regulatory pushback on emissions. By contrast, an infrastructure plan that treats compute as an energy question will guide smart trade-offs: efficiency versus performance, investments in grid and cooling upgrades, and new technologies.

In plain terms: for large AI, crypto, or digital systems, always budget in the power plant and the radiator before you flip the switch. Ignoring that is like building skyscrapers without elevators or fire exits.

It won’t end well.

Written by Orkid Labs