For three centuries, cognition was bound to biological substrate; the supply was effectively fixed at the human population. That is ending.

The unit is the token — the smallest priceable increment of machine reasoning. Cost per token has fallen by roughly two orders of magnitude in three years. Tokens per second per accelerator continues to compound. For the first time in human history, the unit economics of thought are visible, and they are deflationary.

The progress is not pure scaling. Algorithmic insight has done at least as much work as compute, and will continue to. But scaling and insight reinforce each other, and the trendline that emerges from their combination is what most institutions are still mispricing as a software cycle. It is not a software cycle.

This is the substrate we orient around. Everything else in our work exists because it materially compresses the timeline to it.

For most of modern history, value transfer required institutional intermediaries whose primary product was settlement. That layer is being replaced by protocols whose primary product is verification.

The deeper shift is not currency. Frontier intelligence is becoming an autonomous economic actor — agents that transact, contract, capitalize, and coordinate at machine speed. Institutional settlement does not run at machine speed and was never built to.

Programmable value is not a financial innovation. It is the prerequisite for autonomous intelligence to participate in the economy. The first substrate cannot scale without it.

The cost of placing mass in low Earth orbit has collapsed by roughly two orders of magnitude in twenty years. The trend has not finished.

What used to be a matter of state ambition is becoming a matter of industrial logistics. The first economies entirely native to orbit will exist within this decade.

Frontier compute is approaching terrestrial limits — power, cooling, and regulation. Orbit removes all three simultaneously. Compute, energy, and sensing primitives at orbital scale are not adjacent to the intelligence question; they are inputs to it.

These three are not coequal. Intelligence is the destination. Value and orbit are the two infrastructures that determine whether intelligence can scale on the timeline we believe it will.

Frontier training is becoming a thermodynamics problem before it is an algorithms problem. Without orbit lifting the ceiling, scaling laws collide with physics inside the decade.

Without programmable value, the economic surface area of frontier intelligence stays artificially small — capable models constrained by rails that cannot settle at machine speed.

Each substrate, taken alone, is being underbuilt. The mistake is to treat them as sectors. They are a single system constraint, and capital is misallocated accordingly.

Intelligence is the lever, not the destination. Treated as an end in itself, it is uninteresting. Treated as a means, it is the most powerful one ever assembled.

What this unlocks, in our reading, is the compression of scientific progress — drug discovery, materials, fusion, climate, fundamental physics — centuries of work into decades. Intelligence does the compressing; value lets it circulate; orbit lifts the ceiling. The bet worth orienting an institution around is the system, not the lever alone.

Building this well is not separable from building it at all. Capability without alignment is a different kind of failure than failure to build, but it is still a failure. We treat both with the same seriousness, and work alongside the teams that do the same.

The thesis is operational, not academic. To create or accelerate machine intelligence on the timeline implied above, the work has to happen on every layer at once — the labs, the rails beneath them, and the physical infrastructure that powers them.

We operate across all three substrates: building intelligence systems, the economic rails they transact on, and the orbital infrastructure they scale on. The majority of our work is operational — companies we run, products we ship, technology we develop. Research and engineering happen across our offices in San Francisco, Hong Kong, Tokyo, and Singapore. Where the work is more efficient outside our walls, we partner with and support the operators advancing it.