Copenhagen AI
CPH.AI
Approach
Capabilities
Insights
Research Institute
Copenhagen AI
COPENHAGEN AI
ENGINEERING EXCELLENCECREATIVE RENAISSANCEHYPER OPTIMIZATION

We function as the strategic bridge between sovereign infrastructure and autonomous intelligence. Bridging the gap between frontier breakthroughs and systematic industrial execution.

The AI Suite

  • Runestone
  • Bedrock
  • Ledger
  • Vector
  • Aegis
  • Prism

Institute

  • Academic Partnerships
  • Open Source
  • Research Blog

Careers

  • Open Roles
  • The Residency
  • Interviewing
  • Culture
Global Offices
© 2026 Kæraa Group. All Rights Reserved.
Terms of Service|Privacy Policy|Responsible Disclosure|Accessibility Statement
Back to Intelligence Stream
PhysicsSustainabilityHardware

Thermodynamics of Intelligence

Exploring the energy cost per token as the fundamental limit of agentic scaling.

Research Group OCT 14, 2025 11 MIN READ

The Thesis: The Landauer Wall

Intelligence is an energy-intensive process. The human brain operates at roughly 20 Watts. A cluster of H100s training a frontier model consumes megawatts. As we scale agentic systems, we approach a Thermodynamic Wall.

Limit
Log(Energy) / N(Params)

The Efficiency Equation

We model the Energy Cost per Token Et as a function of parameter count N and hardware efficiency η:

Et ∝
Nparams
ηflops/joule

The BitNet Revolution

The shift from FP16 (Floating Point 16-bit) to INT1 (1-bit integer) weights has been the defining hardware shift of 2025. By eliminating multiplication in matrix operations and replacing it with addition, we reduce energy consumption by ~70%.

INT1 Quantization Active

The Jevons Paradox of Inference

Efficiency does not lead to reduced consumption; it leads to increased usage. This is the Jevons Paradox. As inference costs drop towards zero (via BitNet and specialized ASICs), we will not see a reduction in energy use. Instead, we will see an explosion in "Agentic Density".

We will move from one-shot answers to agents that "think" for hours—generating millions of internal tokens to verify a single output. The grid demand will shift from Training Clusters (bursty) to Inference Swarms (continuous baseload).

Subscribe to The Transmission

Get these engineering field notes delivered directly to your inbox. Signal only.