Space

This isn’t about Mars

I want to deal with that at the start, because it’s what most people assume a space thesis is.

Mars is genuinely interesting as an idea. If humanity ever becomes capable of establishing a self-sustaining civilisation on another planet, technological capability has reached an entirely different level from anything we’d recognise today. But that’s a statement about the far future, not a reason to allocate capital now, and any investment case depending on it is a story rather than a thesis.

What interests me is narrower and duller. It’s whether space becomes a place where economically useful infrastructure gets built. Not explored. Built.

That connects to the four pages before this one. T·01 describes intelligence becoming scalable, T·02 intelligence acting physically, T·03 compute becoming infrastructure, T·04 who gets access to it. All four assume an enormous and growing amount of physical infrastructure gets built somewhere. This page is about what happens when the somewhere becomes a problem.

Earth has room. Infrastructure doesn’t.

My original instinct was that we’d eventually run out of suitable places to build. That’s not quite right, and the accurate version is more interesting. There is plenty of land. What’s becoming scarce is the combination of things you need alongside it: grid connection, generation capacity, transmission, water where cooling requires it, permits, and the consent of people who live nearby. Any one of those can stop a project, and increasingly they do.

Texas is the clearest illustration, because it’s the state that wanted this development most. In January 2026, ERCOT had around 233 gigawatts of large loads waiting to connect to its grid. By August that figure had reached roughly 474 gigawatts, more than five times the state’s record peak demand, with about 90% of the requests coming from data centres. On 3 August 2026, Governor Abbott directed the state regulator and ERCOT to audit every data centre advancing through the interconnection process before any further projects could move forward, and said non-compliant projects must be denied connection. ERCOT suspended the Batch Zero classification notifications it had scheduled for that week and moved the verification step to the front of the study process instead.

That’s not a fringe jurisdiction discovering environmentalism. That’s the most development-friendly grid in America deciding demand had outrun its ability to assess it, and doing so to funded projects with financing milestones attached. Separately, Data Center Watch recorded at least 75 US data centre projects worth roughly $130 billion blocked or delayed in the first quarter of 2026.

I want to be honest about the counterargument. SemiAnalysis has argued that most cancelled projects were speculative rezoning applications never going to be built, and that opposition rarely kills a capitalised project already underway. That’s probably right about the headline cancellation numbers. It doesn’t dispose of Texas.

But I should be equally honest about the other direction. These constraints are real and slow to fix, yet none is permanent. Grid investment can accelerate. Permitting can be reformed. Nuclear and storage can expand. Chips get more efficient. Data centres can move to where the power is. Earth may well solve these problems faster than space solves its own, and that possibility is the heart of the bear case I’ll come back to.

What changed: getting things to orbit stopped being absurd

None of this would matter if putting infrastructure in orbit still cost what it used to.

The demonstrated change is remarkable on its own. Launching to low Earth orbit cost tens of thousands of dollars per kilogram in the Space Shuttle era. A Falcon 9, whose list price SpaceX raised to about $74 million in February 2026 for roughly 23 tonnes to low Earth orbit, works out closer to $3,000 per kilogram. That’s an order-of-magnitude improvement achieved by reusing the hardware rather than throwing it away. It’s also a customer price rather than SpaceX’s internal cost, which the company doesn’t disclose and which is almost certainly lower, particularly for its own Starlink launches.

Then there’s Starship, and this is where the distinction between demonstrated and promised matters most.

As of late July 2026, Starship had completed thirteen integrated flight tests. The programme has demonstrated stage separation, booster catches by the launch tower, payload deployment, an in-space engine relight and controlled re-entry, while also suffering multiple vehicle losses during development. SpaceX has said it expects the vehicle to begin delivering payloads to orbit in the second half of 2026, which is a prediction I’ll check rather than assume. What has not been demonstrated is any routine commercial cost per kilogram. Musk has talked about eventually reaching figures in the low tens of dollars per kilogram. I wouldn’t build a thesis on that.

More useful is a number from a party with no incentive to flatter SpaceX. Google’s Suncatcher research states that its analysis of historical and projected launch pricing suggests prices may fall below $200 per kilogram by the mid-2030s, and that at that price point the cost of launching and operating a space-based data centre could become roughly comparable to the reported energy costs of an equivalent terrestrial facility per kilowatt per year.

That’s the most concrete threshold I’ve found. It isn’t full data-centre cost parity, but it gives me something measurable to watch, and it comes from a company that would rather build on the ground if the ground worked.

What actually exists in orbit today

This is where I have to be most disciplined, because the categories blur easily and the enthusiasm is running ahead of all of them.

Demonstrated in orbit: Starcloud launched a satellite in November 2025 carrying an NVIDIA H100, and used it to train a small language model in orbit and run a version of Gemini. Real hardware doing real work above the atmosphere, for the first time.

Tested on the ground, flying later: Google’s Project Suncatcher, announced in November 2025, ran its Trillium TPU under a proton beam and concluded the chip could survive a five-year low Earth orbit mission without permanent failure. Its modelled design is a formation of 81 satellites at around 650 kilometres. Two prototypes fly with Planet in early 2027.

Announced or deployed at small scale: Axiom deployed free-flying orbital data centre nodes in January 2026. NVIDIA announced a space computing product line in March.

Filed with a regulator: SpaceX applied to the FCC in January 2026 to deploy up to a million solar-powered data centre satellites. Starcloud filed for 88,000. A five-month-old startup filed for 100,000 in June.

Those last ones tell you about ambition and about how cheap it is to file. A regulatory application is not a business, and a constellation size in a filing is not a deployment plan.

So the honest position is that orbital compute has moved from slideware toward real hardware, which is a genuine milestone, and remains far from commercial scale. That distinction is the one to hold onto as the announcements keep coming.

The physics that could kill this

Two problems deserve more attention than they get, and one is usually stated backwards.

Cooling. People assume space is cold and therefore cooling is easy. It’s the opposite. A vacuum has nothing to carry heat away, so you can’t blow air across a heatsink or run a cooling tower. Waste heat has to be radiated, and radiating enough of it from high-density compute requires substantial radiator area, which means substantial mass. Mass is the thing you’re paying to launch, so cooling in orbit isn’t an advantage that offsets launch cost. It’s a constraint that adds to it. Sundar Pichai said plainly when Suncatcher was announced that thermal management and reliability were both challenging, and he was describing a research paper rather than a product.

Obsolescence. In T·04 I described how the useful life of AI hardware is genuinely uncertain, and how the largest operators on Earth disagree about it. Now put that hardware somewhere you can’t reach. Google’s radiation testing is encouraging on physical survival, but physical survival isn’t the issue. A chip can work perfectly and still be economically worthless if something four times faster per watt is available on the ground. Replacing it means another launch.

That gives orbital compute an unusual exposure: it competes against a terrestrial cost curve that is itself falling fast, while carrying costs terrestrial facilities never pay. I’ll come back to it, because it’s the centre of the bear case.

Space solar, and a distinction that matters

Space solar gets discussed as one thing. It’s two, and conflating them has confused a lot of people.

The first is generating solar power in orbit and using it there. The physics genuinely favour it. Google’s own wording is that in the right orbit a solar panel can be up to eight times more productive than on Earth and produce power nearly continuously, reducing the need for batteries. That’s a statement about energy yield from a panel in a dawn-dusk sun-synchronous orbit, not a claim that orbital power is eight times cheaper. The panel still has to be built, launched and maintained. But if you’re building compute in orbit anyway, the power source is already there with far better uptime than anywhere on the ground.

The second is generating power in orbit and beaming it to Earth. NASA’s Office of Technology, Policy and Strategy assessed this and concluded such systems would produce power substantially more expensively than terrestrial alternatives under its baseline assumptions. Competing against solar panels, batteries and nuclear on cost, having first paid to launch everything, is a difficult starting position.

I don’t need that argument, and I’d rather my thesis were never attached to it. Worth noting that at least one company founded to beam power to Earth, Aetherflux, now lists an orbital data centre powered by its own solar collection among its stated commercial targets. Whether that’s a pivot or an additional line I can’t tell from outside, but it points where the physics does.

What SpaceX actually has

I think SpaceX has an extraordinary lead in the infrastructure needed to build anything in space. I also want to be precise, because I’ve caught myself overstating this.

SpaceX reports having launched more than 80% of global mass to orbit each year since 2023, with a greater than 99% mission success rate on Falcon rockets, and operating approximately 9,600 Starlink satellites serving customers across 164 countries as of the end of March 2026. Those are company-provided figures and I’d treat them as such, but the scale isn’t seriously disputed.

What matters more than any single number is accumulated repetition. SpaceX has flown and recovered orbital-class boosters hundreds of times, with individual boosters now flying twenty times or more. Reliability in this business is bought with repetitions, and no competitor is close on that count.

What SpaceX does not have is 90% of space, and I’ve said something close to that before. It’s wrong. Launch is one layer of a much larger industry, and dominance in launch is not dominance in everything built on top of it.

Why Starlink matters, and why it isn’t the thesis

Starlink is the evidence that the interesting thing about cheap launch may not be launch.

SpaceX didn’t just lower the cost of putting other people’s satellites in orbit. It used its own launch capability to build a business on top, and the reported economics show where the value is currently landing. In 2025 the Connectivity segment, which is essentially Starlink, generated $11.4 billion of revenue, up about 50%, with $4.4 billion of income from operations. The Space segment generated $4.1 billion and reported an operating loss of around $657 million.

That comparison needs a caveat, and it matters. The Space segment isn’t only the launch business as customers experience it. It covers Falcon 9, Falcon Heavy, Dragon and Starship together, and it absorbed roughly $3.0 billion of Starship development spending in 2025. So the loss is substantially a development cost for a vehicle that doesn’t fly commercially yet, not evidence that selling launches is inherently unprofitable.

The honest version of the point is still striking. The business built using SpaceX’s launch capability is currently producing far stronger reported economics than the segment containing that launch capability. That’s the pattern I care about, and it cuts both ways: it supports the idea that cheap access enables valuable businesses, and it warns against valuing transportation itself too highly.

The pattern also appears to be continuing. SpaceX acquired xAI in February 2026 and listed publicly in June, and the stated rationale for the merger was space-based AI: the argument that terrestrial power and cooling cannot meet what AI needs and that orbit is the way to scale. I’d treat that as an interested party arguing its own book rather than as evidence, and the argument is one I’m making on this page too, so I have to be careful not to mistake agreement for confirmation. What it does tell me is that the company with the most launch capability is choosing to build in exactly this direction rather than remain a transport provider.

But Starlink alone doesn’t justify the opportunity I’m describing. It’s a satellite internet company, and a good one. If Starlink and launch are all this becomes, the thesis on this page is far too grand for what it’s describing. Something bigger has to emerge in orbit for my broader space view to be right.

Competition, and what would actually close the gap

I’m not attached to SpaceX and my conviction would weaken if someone reached comparable launch costs, reliability and scale.

Blue Origin’s New Glenn flew its first commercial missions in early 2026 and won a NASA lunar cargo contract worth up to $468 million. What makes that credible rather than promising is the combination: a flying vehicle, a demonstrated booster landing, and a customer willing to put a lunar payload on it. Then it was grounded after a pad explosion in May, a reminder of how brutal this business is. Rocket Lab’s backlog has more than doubled and its reusable Neutron is aimed squarely at the payload class Falcon 9 dominates, though Neutron hasn’t flown and its economics are projected rather than demonstrated.

The competitor I take most seriously is China. It flew a record 83 orbital missions in 2025 and is developing reusable methane-fuelled vehicles. What makes it structural is the combination behind it: state capital, strategic motivation, and state-backed constellation programmes guaranteeing internal launch demand for years. That’s the same combination that let SpaceX build its repetitions, and it’s the only place I see all three assembled.

Closing the gap needs hundreds of successful recoveries, a manufacturing base to match, and internal demand large enough to justify flying constantly. To me that looks more like years of accumulated execution than something closed with one successful vehicle. That’s a judgement rather than a forecast, and it’s why I think the lead is real without thinking it’s permanent.

The argument I take most seriously

Everything above assumes a causal order: Earth gets constrained, launch gets cheaper, orbit becomes economically accessible, infrastructure moves, new industries follow.

The alternative order is entirely plausible and would leave this thesis stranded. Terrestrial infrastructure improves faster than orbital economics. Grid connections get built, nuclear and storage arrive, permitting gets reformed, chips get more efficient, and Earth simply stays cheaper. Launch grows into a solid business serving satellites, and the larger frontier never materialises.

This is the version I can’t dismiss. For general-purpose orbital compute to compete with Earth, falling launch and orbital operating costs have to overcome continued improvement in terrestrial compute economics, while additionally paying for radiators, radiation tolerance, replacement launches and the inability to send an engineer. Both curves are moving. Only one has to move faster. Google’s sub-$200 per kilogram figure is the most useful marker I have for where the crossover might begin, and it marks the point where orbital power costs enter the terrestrial range rather than the point where a whole orbital data centre becomes cheaper.

The crossover also doesn’t have to happen everywhere at once. Specialised workloads could cross far earlier: processing Earth-observation data where it’s collected rather than downlinking it, serving other satellites, or anything where the alternative is moving enormous volumes of data to the ground. Those cases don’t require beating a terrestrial data centre on general-purpose cost. They require the round trip to Earth being worse.

One thing I should correct in my own thinking. I’ve been tempted to argue that orbit’s advantage is that it doesn’t need anyone’s permission. That’s wrong. Orbital infrastructure needs launch licences, spectrum, regulatory approval, debris compliance and international coordination, and those constraints will tighten as constellations grow. The accurate version is narrower: orbital infrastructure may avoid some of the specific local land, grid connection and community constraints terrestrial data centres face, while acquiring a different set of its own.

What would make me wrong

This is the most speculative page on this site, and I’d rather say that plainly than bury it.

The infrastructure I care most about does not exist at commercial scale and may never. If orbital data centres and other economically meaningful space infrastructure don’t become viable, my broader space thesis weakens materially. Not partially. Materially. Launch and satellite internet would remain real businesses, but the frontier argument would be wrong.

Specifically, I’d reduce conviction if terrestrial power and permitting constraints ease substantially, if Starship never approaches its target economics, if thermal management or radiation prove more limiting than current work suggests, if hardware obsolescence makes orbital deployment uneconomic regardless of launch cost, if debris or regulation restrict deployment, or if a competitor matches SpaceX on cost, reliability and scale.

And there’s the risk that applies to every page here, in its sharpest form. Space could become genuinely important and still be a poor investment, because the capital required is enormous, the build times are long, and the returns arrive late or not at all. This is the thesis where the gap between being right about the technology and being paid for it is widest.

How I hold this

I’ve owned smaller space companies before, including Rocket Lab, AST SpaceMobile and Intuitive Machines. They were higher-risk positions, sized accordingly, and I later sold them as I moved capital toward what I considered stronger leaders.

That’s the process rather than a track record. When I can’t identify a winner, I’d rather hold several small positions than one large one. As execution proves itself and evidence accumulates, conviction can rise and a position can grow. When a much clearer leader becomes investable, I’m willing to rotate toward it. That’s what happened here, and it could happen again in the other direction.

Worth stating plainly: the thesis on this page is not that SpaceX wins. It’s that space becomes somewhere meaningful infrastructure gets built. SpaceX is currently how I’d express that view, because of its launch scale, accumulated reusability experience, Starlink, and the internal demand that keeps it flying. If those advantages weaken, or someone matches them, my allocation changes without my view of space changing at all.

What I’m watching is narrow enough to state: whether launch costs keep falling toward the range where orbital economics start working, whether the demonstrations of 2026 and 2027 turn into operating businesses, and whether terrestrial constraints ease.

If space becomes a place where meaningful infrastructure gets built, the companies controlling access to it will matter enormously. If it doesn’t, I’ll have owned a good satellite business and a powerful launch platform, which is still a considerably smaller idea than the one on this page.