The Folly

In the landscape gardens of the eighteenth century there is a particular kind of building called a folly. It is a tower, a ruined abbey, a temple on a far hill — built to be seen and never entered. The point of a folly is the view of it. It dignifies the prospect; it draws the eye across the valley; it photographs, in the language of the time, beautifully. What it does not do is keep the rain off anyone.

There is a folly being drawn in the sky at present, and it is magnificent. Data centres in orbit: solar power without a grid queue, cold without a cooling tower, compute lifted clean out of the terrestrial constraints I wrote about in The Cost of Light. The renderings are extraordinary. The market is beginning to price some of them as though they were buildings you could move into next year. They are follies. That is not an insult — a folly is real masonry, real engineering, a genuine object on the hill. It is simply not a structure anyone can yet occupy, and you should not pay occupancy rents for the view.

The category error, in orbit this time

The economic ruler here is not in dispute, only in provenance. The most-cited reconstruction has orbital compute running roughly four times the terrestrial cost today, narrowing toward something like parity around 2040 under aggressive launch assumptions.[^1] I will not pretend the dated point estimates — parity in this year for edge inference, that year for delay-tolerant cloud — are measurements. They are outputs of a deterministic scaffold built on public-estimate inputs, and a scaffold fed its own assumptions can only ever confirm its own shape. The two-decimal precision you see attached to these models is the tell of a thing that wants to feel measured. Hold the qualitative shape; discard the decimals.

A scaffold fed its own assumptions can only ever confirm its own shape. The two-decimal precision is the tell of a thing that wants to feel measured.

Letter 015 — The Folly

The shape that survives is honest enough: launch cost per kilogram and radiator specific power together govern most of the timing — they are co-equal levers, and anyone who tells you radiative heat rejection is the single binding constraint is overfitting the noise. The physics floor is hard and real. The financeability is absent. Those are two different sentences about the same object, and the market keeps reading the first as if it answered the second.

The foundation nobody is photographing

Here is the part the renderings obscure. The orbital parity date is not a fixed star; it recedes every time the ground gets better. Onsite generation at the rack — small modular reactors, behind-the-meter gas, the whole apparatus of not waiting for an interconnect. Direct-to-chip and immersion cooling that quietly relaxes the thermal limit. Storage drifting toward fifty dollars a kilowatt-hour. Each of these is a load-bearing improvement to the terrestrial structure, and each one pushes orbital parity outward — plausibly past 2050.

This is the inversion worth holding. The better the foundation gets, the longer the folly stays a folly — which means the same engineers and balance sheets straining to make orbit competitive are, by improving the ground, the thing that defeats it. The investable inflection is not in the sky. It is the unglamorous terrestrial build-out that the orbital narrative is, in effect, subsidising with its attention. The crowd is looking up. The load is being transferred at ground level.

A survey, not a forecast

I do not predict when the folly becomes a building. I do what a surveyor does: I know where the joints are, and I visit them on a schedule — especially when nothing appears to be happening. There are exactly three observable gates that separate "feasible physics" from "financeable security." As of this writing, all three are unmet. Keep this table taped inside the cover of the notebook; it is the only part of this letter you need to act on for the next several years.

Gate	What clears it	Trigger year	Status
Launch economics	Starship sustained under $200/kg — audited, not promoted	2030	Unmet
Thermal rejection	A flight-qualified, scaled radiator with a verified W/kg	2028	Unmet
Ground interface	A flight-relevant downlink above 100 Gbps per node	—	Unmet

The third gate is the one the spectacle hides, and the one I would watch hardest. Satellites can move on the order of a terabit per second between themselves; what reaches a paying customer through weather, turbulence, and a finite number of ground stations is one to ten gigabits per node — two to three orders of magnitude less, and gated as much by ITU and FCC spectrum coordination as by any optical terminal. A node that trains beautifully in orbit and cannot downlink the result is a stranded asset. It is the locked door of the folly: you can admire what happens inside; you cannot get it out.