The Lens · Frontier Thesis

The Off-Earth Economy

The next frontier isn't a place — it's an economy. Three things are starting to leave the planet at once: compute, mining, and manufacturing. Here's how they fit together, who's building it, and the honest timeline. Hover or tap any underlined term.

Dragonfly Lens · June 8, 2026 · A long-horizon thesis — mostly 2030s–40s. We flag the uncertainty throughout.

The short version

Why now: two things that never existed before now do — cheap launch (Starship) and autonomous robotics (AI). Together they make off-Earth economics work for the first time.
What's leaving: compute (orbital & lunar data centers), mining (lunar water/Helium-3, asteroid metals), and manufacturing (solar & structures built from space materials).
The key insight: these aren't three trends — they're one loop. You mine space materials to build in space to power compute in space.
The honest part: this is a 2030s–40s story, mostly private and pre-revenue. It's a trajectory to track, not a "buy now."

Why now — the two unlocks

People have dreamed about space industry for 50 years. It failed every time (remember the 2010s asteroid-mining startups that went bankrupt?) for one reason: the economics never worked. Two things just changed that.

Cheap launch. Starship aims to cut the cost of reaching orbit by 10–100×. When launch is cheap, everything downstream becomes possible.
Autonomous robotics. AI-driven autonomy removes the need for humans on site — the most expensive, dangerous part. Robots that can prospect, extract, and build on their own.

Cheap launch plus AI autonomy is the combination the 2010s never had. That's why this decade is genuinely different — not hype, just two enabling technologies finally arriving at once.

The three layers leaving Earth

Layer 1 · Compute

Data centers in orbit, then on the Moon

AI's two bottlenecks are power and cooling — and space splits them. Power is the easy win: in a dawn-to-dusk sun-synchronous orbit, solar panels get near-constant sun, no atmosphere in the way, no batteries. Cooling is the hard part — the opposite of most people's intuition (see the reality check). It's already started: Starcloud ran AI on an Nvidia GPU in orbit (Dec 2025), and SpaceX has filed FCC plans for orbital compute. The far-future vision: a lunar mass driver flinging compute into deep space without rockets — a decades-out idea, not a near-term plan.

Reality check — why cooling, not power, is the real obstacle. Space is cold (~−270°C), so the intuition is "the heat just gets pulled away." It's the opposite: a vacuum is an insulator, not a conductor. Heat escapes three ways — touching (conduction), air/water carrying it off (convection), and radiating it as infrared. In a vacuum the first two don't work at all, so the only exit is slow infrared radiation. It's the same physics that keeps a thermos hot — the vacuum gap traps the heat. A data center in orbit is essentially a computer sealed in a thermos. The numbers: Starcloud's own white paper pegs a radiator at ~633 W/m², over 1,000× slower than water-cooling chips on Earth — for a 1 MW center the radiators alone could weigh ~100 tons, roughly 10× the mass of the compute. And the economics hinge on launch cost: estimates say liftoff must fall below ~$200/kg by ~2035 to pencil out — which is exactly why a cheaper SpaceX Starship is the lever that turns this from sci-fi into a real (if long-dated) bet. How fast? See the four-timeline scenario table below — the honest range runs from "2027, Elon-speed" to "2035+, maybe never."

Meanwhile, on Earth, the same logic already runs — underwater. China switched on the world's first wind-powered undersea data center off Shanghai (May 2026): ~2,000 servers, 24 MW, 35m deep, using the ocean as a heat sink + offshore wind — a PUE under 1.15 vs the ~1.5 norm. Same idea as the space thesis — put compute where cooling and power are cheap — except water is a far better heat sink than vacuum, which is why the ocean works today and orbit is the longer-dated bet.

June 2026 update: it stopped being a concept — meet AI1

On June 8, 2026 — four days before its IPO — SpaceX unveiled AI1, its first-generation orbital data-center satellite, and the supporting pieces that turn the white-paper debate above into checkable hardware claims:

The machine: ~70-meter wingspan (wider than a Boeing 747-8), 120 kW sustained / 150 kW peak of solar-powered compute, an interchangeable chip payload (so the satellite outlives any one GPU generation), laser links for data, and a 110 m² liquid radiator to shed the heat. Musk calls it "much simpler than a Starlink satellite."
The factory: an 11-million-sq-ft "Gigasat" plant in Bastrop, Texas, to mass-produce them — targeting 1 gigawatt of orbital compute per year by late 2027.
The ambition: an FCC filing for up to one million satellites — and the S-1 ties 300 million shares of Musk's own pay package to orbital data centers delivering 100 TW/yr of compute. The bet is written into his comp.

Run AI1 against the physics above — honestly. 120 kW through 110 m² of radiator implies ~1,090 W/m² of heat rejection — ~1.7× better than the ~633 W/m² Starcloud's white paper assumed (achievable in principle by running radiators hotter, which costs chip efficiency). That's a real engineering claim, not yet proven in orbit. And note what AI1 is for: training and batch compute, not your chatbot reply — orbit adds latency, so the low-latency inference market stays on the ground. The honest read: the physics objections haven't been answered yet — but for the first time, there's a dated, funded, factory-backed attempt to answer them.

Layer 2 · Mining

Resources, extracted in space

You can't ship everything from Earth — so you mine where you are. The near-term reality is ISRU: lunar water → rocket fuel, regolith → oxygen and metals. Interlune targets lunar Helium-3; AstroForge is already flying asteroid-prospecting missions. (Bringing bulk metals back to Earth is the hard, far-off part — see the timeline.)

Layer 3 · Manufacturing

Building in space, from space

The payoff of mining is making things on site: solar panels, the radiators that cool the data centers, habitat structures — all from lunar material instead of launching them. This closes the loop and is what lets the whole thing scale beyond anything possible from Earth.

The convergence (the part nobody connects)

Here's the insight that ties it together: these three layers are one machine.

Cheap launch (SpaceX) gets you there → AI robots mine the Moon → manufacturing turns that material into solar + radiators → which power compute in space → whose value justifies more launch. A self-reinforcing loop.

That's Elon's actual argument for going to the Moon before Mars: it's the nearest place you can close that loop and scale it. Each piece makes the others economic. No single company owns it — which is exactly why mapping the whole chain matters.

The honest timeline

Now – 5 yrs

Demos & proofs: orbital AI training (done), asteroid prospecting (flying), lunar water-ice extraction (NASA, 2025), Helium-3 offtake deals.

~10 yrs (mid-2030s)

First operational orbital data centers + lunar ISRU (water→fuel) + specialty exports. The in-space economy starts.

20–30 yrs

Integrated off-Earth economy: lunar manufacturing + compute at scale; mass drivers; the 1,000-TW vision — if the loop closes.

Read this honestly: almost every pure-play here is private, pre-revenue, and high-variance, and the boldest claims (1,000 TW, asteroid riches) are decades out and may never happen. This is a trajectory to track — each demo that works de-risks the thesis a little — not a buy-it-today trade. We flag it now so you see it coming, not so you chase it.

Four timelines for orbital compute — and how to know which one you're on

Forecasting a single date is pretense. The honest version is scenarios with checkable gates — and one historical fact that justifies taking the fastest lane seriously: when xAI built Colossus, Nvidia's own CEO said a 100,000-GPU supercomputer normally takes "three years to plan and a year to get working." xAI went from hardware delivery to training in 19 days (~122 days start-to-finish) — Jensen Huang called it "superhuman… there's only one person in the world who could do that." Musk's timelines are usually late by his own schedule and still years ahead of everyone else's.

Scenario	First commercial orbital compute	What has to be true	You'll know you're on this track when…
Conservative (the physics skeptic)	~2035+, maybe never at scale	Starship cost/kg stalls; radiators underperform in orbit; chip radiation losses eat the economics; terrestrial power gets cheap faster (nuclear/SMRs).	AI1 slips past 2027; Gigasat output stays in the dozens; no paying tenant signs.
Balanced (our base case)	~2030–2033 for meaningful capacity	AI1-class hardware works but iterates slowly; launch falls toward ~$200/kg by early 2030s; orbital wins the training/batch niche only.	First AI1s fly 2027–28, perform near spec; a hyperscaler signs a pilot offtake by ~2029.
Optimistic	~2028–2030	Radiators hit the claimed ~1,090 W/m²; Gigasat ramps like Starlink's factory did (the one assembly line SpaceX has already proven at scale); Starship flies weekly.	1 GW/yr production rate confirmed by late 2027; compute-per-launch-dollar beats new terrestrial DC build in a published deal.
"Elon-speed" (the Colossus precedent)	First revenue-bearing orbital compute 2027, GW-scale by ~2029	Everything above, plus the 19-days-to-training execution culture transfers from Memphis to orbit — and the IPO's $75B war chest funds it without dilution pain.	AI1 launches this year; Anthropic/Google-style tenants extend their ~$26B/yr compute deals to include orbital capacity.

One asymmetry worth noticing: even the conservative scenario is bullish for the suppliers — the radiators, rad-hard chips, laser links, and launch cadence get paid for whether or not orbital compute ultimately beats terrestrial. The scenario risk lives almost entirely in the orbital operator's economics, not the chain that feeds it.

The ripple map — and why every problem on it is a business

If orbital compute scales on any of the middle timelines, the effects run far beyond SpaceX — and the pattern frontier investors should internalize: every "fatal problem" on this list is an unsolved market waiting for its company. The objections are the opportunity list:

The problem ("why it'll fail")	The business it creates	Who's positioned / how to watch
Heat rejection — the thermos problem above	Advanced radiators, deployable thermal systems, two-phase cooling loops — the "transformers of space"	Mostly private/defense today; watch who Gigasat sources from. The picks-and-shovels of the whole thesis.
Radiation kills chips — no rad-hard H100 exists	Radiation-tolerant AI compute — shielding, error-correction, custom silicon ("non-terrestrial chips" are already in the Terafab plan)	NVDA flies on tolerance + redundancy today; Terafab's space-chip line is the tell to watch.
Orbital debris / collision risk — a million more satellites	Space situational awareness, traffic management, deorbit services — mandatory infrastructure, likely regulated into existence	RKLB, RDW (public); LeoLabs, Slingshot (private). Starlink already runs ~300k collision-avoidance maneuvers/yr — 1M sats makes SSA non-optional.
Things break, no technicians in orbit	Orbital servicing & robotic repair — the space tow-truck/mechanic industry	Early private players; AI1's interchangeable payload design hints SpaceX plans swap-don't-fix.
Getting data down — lasers need ground stations	Optical ground stations, laser-link networks, edge gateways	Watch who builds SpaceX's downlink web; terrestrial fiber at the gateways still wins (the old fiber map matters again).
Astronomy + spectrum backlash	Regulatory navigation, dark-sat coatings, spectrum coordination — compliance as a moat	The FCC million-sat filing is the test case; expect this fight to be loud and slow.
Insuring a million satellites	A new actuarial class — space insurance at constellation scale	Specialty insurers; premiums per kW in orbit will be a public signal of how risky the pros think it really is.

Downstream — what it does to the rest of our map: if orbital compute arrives on the balanced-or-faster track, it becomes a relief valve for the terrestrial grid — but not before the late 2020s. The grid/transformer bottleneck stays binding through exactly the window our thesis covers (2026–2030); orbit is the next chapter, not a contradiction. Energy prices, data-center siting wars, and nuclear PPAs all keep their near-term logic — and if orbit works, the eventual loser isn't the grid names, it's marginal terrestrial data-center real estate in the 2030s. Up the chain, the winners repeat: launch cadence (SpaceX), chips that fly (NVDA/Terafab), and the laser/thermal/SSA suppliers above.

Who's building it

Layer	Name	What they do
Compute	Starcloud	Orbital data centers; trained AI in space ('25); ~$1.1B	private
Compute	Lonestar / SpaceX	Lunar data backup; SpaceX "non-terrestrial" chips (Terafab)	private
Mining	AstroForge	Asteroid prospecting & refining (flying now)	private
Mining	Interlune	Lunar ISRU + Helium-3 specialty exports	private
Logistics	TransAstra	Space logistics + resource harvesting	private
The enabler	SpaceX (SPCX)	Starship = the cheap launch that makes it all possible	IPO ~Jun 12
Public proxies	LUNR, RKLB, NVDA	Lunar infrastructure, launch, the chips that fly	public

The pattern you'll notice: the pure-plays are private and early; the buyable names are the enablers and the picks-and-shovels. That's how every frontier starts — you can't buy the dream yet, so you buy the shovels and watch the pipeline.

The moonshots — the far-future unlocks

Everything above is the credible, fundable near-term. These are the "we don't have the tech yet, but physics allows it" tier — the ideas that, if any one lands, don't just grow the off-Earth economy, they transform it. Clearly speculative; included because each is a target someone may actually go solve.

Self-replicating robot factories moonshot — the real scaling unlock. Robots that mine lunar material and build copies of themselves would grow capacity exponentially with no new launches — the convergence loop above, on autopilot. Needs autonomy + in-space manufacturing we don't have yet.

The lunar mass driver moonshot — an electromagnetic catapult flinging Moon-made compute and cargo into deep space with no rocket at all (the Moon's low gravity + no atmosphere make it possible). Decades out, but it's how the 1,000-TW vision actually reaches scale.

Space-based solar, beamed to Earth moonshot — collect constant orbital sun, beam it down by microwave. It flips the thesis: instead of moving compute to space for power, move the power to Earth — potentially defusing the whole grid bottleneck.

Asteroid redirect moonshot — instead of chasing a metal-rich asteroid across the solar system, nudge a small one into orbit near the Moon and mine it there (NASA studied exactly this). Asteroid-metal economics only work if you don't have to fly all the way out to them.

The honest tag: none of these are investable today — they're the ceiling on the thesis, the reason the upside is "whole new industries" rather than "a bigger launch market." Track the first credible demo of any one — that's the signal the off-Earth economy just jumped a tier.

See it before the crowd

We map the frontier early.

Dragonfly Lens connects the threads — compute, space, energy, AI — into theses you can act on, in plain English, fully sourced. Years before they're obvious.

Join the Lens →

Sources: space-based data centers + the cooling/radiator reality — World Economic Forum, Scientific American, NVIDIA / Starcloud, McKinsey; the ~$200/kg-by-2035 launch-cost threshold — Google Suncatcher via Scientific American; China's undersea data center — Tom's Hardware; AI1 satellite specs (120/150 kW, ~70m span, 110 m² radiator, interchangeable payload, Jun 8 2026 unveiling) — Tom's Hardware, Yahoo Finance; Gigasat factory (11M sq ft, 1 GW/yr by late 2027) — Tom's Hardware; FCC filing for up to 1 million satellites — Data Center Dynamics; Colossus "19 days / superhuman" (Jensen Huang) — Entrepreneur, Business Today; orbital data centers in Musk's pay package (100 TW/yr condition) — S-1 exhibit (SEC).

Educational research, not personalized investment advice. Dragonfly Lens is not a registered investment advisor. This is a long-horizon, forward-looking thesis as of June 8, 2026 — timelines are uncertain, many claims are speculative, and most companies named are private and not publicly tradeable. Verify every name against primary sources. Past performance does not guarantee future results.