Back to Insight
ResearchJune 25, 2026

The Carcinisation of Compute

Why Data Centers Are Starting to Look Like Crypto Farms

As AI changes the physical and economic constraints of infrastructure, compute is converging toward the same industrial attractor state — driven by power, heat, density, speed, and cost per unit of compute.

Different Origins. Same Destination.

In evolutionary biology, carcinisation describes a strange but powerful phenomenon: unrelated species repeatedly evolve into crab-like forms. They do not start from the same origin. They do not follow the same path. But under similar environmental pressures, they converge toward the same physical structure.

Different SpeciesSame Physical ConstraintsConvergent Crab-Like Form

The reason is simple: the crab works. When nature faces the same constraints again and again, it often arrives at the same answer.

The Internet-Era Data Center

For decades, data centers were built under the assumptions of the internet era. They were designed around latency, urban proximity, Tier III / Tier IV redundancy, HVAC air cooling, raised floors, heavy UPS systems, and human-friendly maintenance.

Urban Proximity

Latency to end users

Tier III / IV

Maximum redundancy

Air Cooling

HVAC and raised floors

That model made sense for enterprise software, financial systems, web applications, and consumer internet workloads. But AI changes the environment.

When Constraints Change, Morphology Changes

AI training clusters do not primarily optimize for proximity to users. They care less about being 50 milliseconds closer to a city. They care much more about whether thousands or tens of thousands of GPUs can stay powered, cooled, and economically productive at scale.

Old Constraints

  • Latency to users
  • Uptime and permanence
  • Hardware redundancy
  • Architectural elegance

New Constraints

  • Power and compute density
  • Thermal efficiency
  • Cost per MW and deployment speed
  • Silicon utilization

When the constraints change, the morphology changes.

What Looked Primitive May Have Simply Been Early

Crypto mining facilities evolved under one brutal variable: maximum compute per kilowatt-hour. To survive, they stripped away everything non-essential.

Cheap Power

Close to stranded energy

Simple Structures

Industrial wrappers

Fast Deployment

Low CapEx, high output

Crypto farms were not solving the same problem — they were solving a more industrial one: how to convert electricity into computation as efficiently as possible.

The Carcinisation of Compute

AI training is now pushing hyperscale data centers toward the same industrial logic. Data centers are not literally becoming Bitcoin mines. But they are increasingly responding to the same physical and economic pressures.

The same pressures produce the same morphology

Power

Heat

Density

Speed

Cost / Compute

As those pressures intensify, the old one-size-fits-all framework for data center design begins to break.

The End of One-Size-Fits-All

The traditional model assumed all serious data centers should maximize uptime, redundancy, permanence, and urban proximity. AI forces a more segmented view. Training, inference, storage, and enterprise workloads have different engineering requirements.

Training

Density and power first

Inference

Latency-segmented

Storage

Capacity-optimized

Enterprise

Uptime and proximity

A single universal data center model no longer makes sense. The industry is moving toward ruthless practicality.

Hyperion: Engineering Radicalization

Meta's Hyperion project is one of the clearest signals of this shift. It represents engineering radicalization at the infrastructure level — compromises that were previously deemed extreme are now becoming rational.

Rapid deployment structures reduce the importance of permanent concrete-heavy buildings.
Direct liquid cooling replaces traditional air-cooling assumptions.
Software checkpointing reduces dependence on massive hardware redundancy.
Large-scale power infrastructure becomes part of the data center itself.

The facility is no longer designed around perfection. It is designed around throughput.

At Gigawatt Scale, Architecture Becomes Capital Allocation

At small scale, these choices look like engineering details. At gigawatt scale, they become financial strategy.

Traditional Hyperscale

$8M–$10M

per MW

Stripped-Down AI Compute

$4.5M–$5.5M

per MW

At 5GW Scale

Every dollar saved from overbuilt redundancy and excessive building complexity can be redirected toward the assets that actually produce value: silicon, power, cooling, substations, transmission, and network fabric.

Convergence Toward the Same Attractor

The old data center was built to serve humans and applications. The AI factory is built to feed machines. That difference changes everything.

Nature builds crabs because crabs work. Economics builds crypto farms because industrial compute has its own attractor state.

The carcinisation of compute has already started.

Learn more about Leviathan

Explore how we capture power-first infrastructure for AI compute.

    Built with v0