Compute
Semiconductors
The Intelligence Stack is a story about software until it is a story about wafers, wattage, and yield.
Mission
Why this layer exists.
To treat compute as an architectural layer: devices, interconnect, and systems designed for intelligence rather than borrowed from it.
Problem
What is unfinished.
The last decade rented its intelligence from a handful of fabs and a handful of accelerators. That was rational. It is not a strategy. Energy, memory bandwidth, and supply concentration are now the binding constraints on thought. A company that builds intelligence without a point of view on silicon is building on someone else’s timetable.
Vision
Thought has a physics. We intend to design it.
Celestra approaches semiconductors as the floor of the stack. Not a vanity fab. A research and systems program: architectures matched to models, to agents, to physical AI. The aim is sovereignty of design — the ability to specify the machine that intelligence deserves.
Architecture
The system.
- 01
Workload truth
Agents, training, and control have different shapes. Compute must stop pretending they are one benchmark.
- 02
Systems
From device to rack to facility. Intelligence is an energy story as much as a transistor story.
- 03
Co-design
Model and machine specified together. The stack is allowed to change both.
- 04
Supply as strategy
Partners, process, and geography treated as design inputs, not afterthoughts.
Future
What this becomes.
The civilizations that lead in intelligence will be the ones that can still make the substrate. Celestra intends to remain fluent in that fact.
Related Research
Adjacent layers