Compute per joule
Feedstock, lithography consumables, thermal materials, and the firms that make them — what decides how much computation a watt can buy.
Does this make a thought cheaper to have?
On a curve, and a chokepoint. Fail either test and it does not enter coverage.
We decline business and coverage every week. Not because either is unprofitable, but because a universe that drifts into general commodities and general tech loses the only edge a specialist has: knowing, for a countable number of nodes, every qualified producer, every grade that actually ships, and every lead time that is real rather than quoted.
Feedstock, lithography consumables, thermal materials, and the firms that make them — what decides how much computation a watt can buy.
Does this make a thought cheaper to have?
Transformer steel, conductors, superconducting tape, the cryogens that keep them cold, and the interconnect queue. A datacentre that cannot be energised is a shed.
Does this get power to where the compute is?
Permanent-magnet feed, precision drives, additive manufacturing — the step where a model stops advising and starts doing physical work.
Does this give intelligence hands?
Being on a curve is not enough. Most of a supply chain is substitutable, and therefore uninteresting. A node earns coverage when it gates a curve.
The unit of coverage is a chokepoint, not an asset class. This is how the universe reaches robotics, AI hardware and additive manufacturing without becoming “everything”: you arrive at them by tracing a chain you were already mapping.
A substance with a grade, a purity and a producer.
Public or private. Most of the interesting ones are private.
Where the process knowledge actually lives.
Tools with single-digit annual output and multi-year queues.
Know-how that does not transfer with a purchase order.
Research first, because it costs nothing but attention and it is the sourcing work the desk needs anyway.