The power wall

For two decades, progress in computing felt like a law of nature: transistors doubled, chips got faster, workloads scaled. That era is over. Dennard scaling broke around 2005. Moore's Law continues in transistor count but not in the predictable performance dividend software once rode for free. The industry pivoted to parallelism — GPUs, clusters, tensor cores — and that pivot accidentally created the hardware landscape on which Transformers and modern language models became economical.

We are now hitting a different wall. Not the transistor. The grid.

The new binding constraint

Training and serving large models already consume electricity at scales that are legible in household-years and small-city power budgets. Inference at global deployment scale compounds daily. This is not an environmental aesthetic preference. It is a planning constraint for any institution — university, hospital network, public agency, or research lab — that assumes compute-intensive AI can grow without a matching energy strategy.

The constraint bites harder as workloads shift from single-shot prompts to multi-step reasoning: agents that loop, retrieve, draft, verify, and loop again. Each step extends context. Attention cost grows quadratically with context length in standard Transformer architectures. Chain-of-thought is not a free lunch; it is a metered resource.

We have been modelling this formally as contextual depth: the number of iterative reasoning steps a system takes, each adding to the running context. When output from step n becomes input to step n+1, costs do not add linearly — they can compound sharply. Institutions budgeting "one API call per user question" are already mis-budgeting systems that reason in ten implicit steps.

What this means for research and public institutions

First, the frontier is not only algorithmic. A lab or civil-society partner evaluating "whether to adopt AI" must ask where the megawatt-hours come from — and what happens when reasoning depth doubles.

Second, verification has an energy cost. Our work on delegation and lock-in treats verification as a social and institutional problem. It is also a compute problem. Systems that delegate cheaply but verify expensively may converge on shallow checking not only because humans are tired, but because organizations cannot afford thorough review at scale. Energy and epistemics couple.

Third, hardware roadmaps and energy roadmaps are merging. Major technology firms are signing long-horizon power agreements — including nuclear and other baseload sources — not as public relations, but because cluster expansion is power-limited. Research funders who treat compute as someone else's capex are outsourcing a strategic dependency.

What we are not claiming

We are not arguing that AI progress stops. We are not selling a particular chip architecture or energy vendor. We are not using "singularity" language — the future is not guaranteed to any single exponential.

We are claiming that serious civilian research programs must internalize energy and depth-aware cost models the way they internalize IRB review, data governance, and reproducibility. A partnership that cannot say how many reasoning steps it expects, at what context sizes, on what hardware, for how many users, is not yet ready for production — regardless of benchmark scores.

Three fronts that actually matter

Algorithms. Architectures that reduce quadratic attention costs, sparse or recurrent alternatives, distillation, and task-specific models that do not default to largest-available general models. Not every problem needs a frontier LLM. Many need a right-sized model with a verified evaluation harness.

Silicon integration. Advanced packaging, memory bandwidth, and domain-specific accelerators still yield gains — but they are engineering gains, not the automatic doublings of the 1990s.

Energy and siting. Where power is clean, reliable, and accountable; how heat and water use affect communities; what happens to research continuity when electricity prices spike. These are civilian infrastructure questions.

Why this note exists

We wrote this in July because partners and funders were asking whether Neocortic's research agenda — learning, language, verification, disaster-risk modelling — could simply "scale up on cloud APIs." The honest answer is: only with a power and depth budget, and only where institutional verification keeps pace.

Public-interest AI that touches education, health, labour markets, or disaster response cannot treat compute as infinite. The communities those systems serve pay for failure in trust, access, and sometimes lives. Energy is part of that ledger.

Open questions

• Can we publish depth-aware cost models alongside model cards — expected reasoning steps, context growth, and energy bands — so funders can compare systems fairly?
• Where does energy-constrained verification create new moral crumple zones — humans blamed for not checking outputs an organization could not afford to check?
• What research programs remain worth doing at smaller scale rather than at frontier scale — and how do we evaluate them without frontier envy?

We welcome collaborators who treat energy, measurement, and institutional capacity as first-class design constraints — not as afterthoughts once the demo works.

If this fits your mandate, start a conversation.