By Richard Stroupe

In 2019, I wrote the first Cape Fear Ventures check into SpaceX.

The company was worth around $30 billion. Starlink had just launched its very first 60 test satellites and had zero paying customers. SpaceX completed only 13 launches total that year. Reusability was still extremely early. Crew Dragon had not yet carried a single astronaut. And Starship was a stainless-steel prototype doing short hops in the Texas desert.

The S-1 dropped Wednesday, and the internet is already on fire.

Early indications point to intense demand at a $1.75 trillion valuation, with the IPO expected as early as June 12.

Then on Friday, SpaceX launched Starship V3, the first flight of its next-generation, significantly more powerful version. The upper stage performed strongly on its debut, successfully deploying its test payload of ~45 tons of Starlink mass simulators into orbit and executing a controlled splashdown in the Indian Ocean.

While the booster had some issues, the flight is widely viewed as a clear success for the upgraded V3 hardware.

Timing like that is hard to ignore.

After going over the S-1 filing, here’s what captured my attention most as a long-term believer:

#1 The $28.5 trillion TAM is not a valuation justification. It is a declaration of what SpaceX intends to become

The filing states a quantifiable TAM of $28.5 trillion. Of that, $26.5 trillion is AI-related. Within the AI figure, $22.7 trillion sits in enterprise applications.

Traditional space and connectivity, the businesses that built the original thesis, accounts for roughly $2 trillion combined.

On that split, SpaceX is not filing as a “space company that also does internet”. It is positioning itself as the indispensable physical infrastructure layer for the AI future.

The filing is direct: "The future of AI will be determined by the control of the physical stack.”

The real question is whether, at true planetary scale for AI, you eventually have to go through SpaceX infrastructure to play, and whether their own products are waiting for you when you get there

They are going vertical on every physical constraint that matters.

• Chips through Terafab
• Data centers through Colossus
• Power through orbital solar
• Connectivity through Starlink
• Launch capacity so no competitor deploys infrastructure at scale without paying SpaceX

The S-1 is explicit that these are management estimates, exclude China and Russia (roughly a third of global AI spend) and rest on assumptions about how layers that do not fully exist yet will eventually compound.

Regulatory constraints on orbital access, energy bottlenecks, or competitors building parallel infrastructure could shrink the realistic capture. That caveat belongs front and center.

But virtually every time SpaceX has made a claim like this, the market debated the number and missed what was happening.

Reusable rockets were called impossible until they landed themselves on a barge. Starlink was called economically unviable until it generated $11.4 billion in revenue at a 63% EBITDA margin, while deliberately cutting prices to grow faster.

#2 The orbital compute thesis is an energy story dressed as a space story

The S-1 describes plans to deploy orbital AI compute satellites as early as 2028, running AI workloads in space where solar power is effectively unlimited and cooling is free.

The stated target is 100 gigawatts of annual compute power, requiring thousands of launches per year and roughly one million metric tons to orbit.

This comes as AI data centers are running into a power wall.

The grid cannot keep pace. US electricity generation grew at less than 3% annually from 2023 to 2025 while AI compute demand exploded. Every major cloud provider is now queuing for grid connections measured in years.

The filing calls the resulting supply and demand gap "already imposing unsustainable strains on terrestrial power grids, supply chains, and the environment."

SpaceX's answer is to move the compute to where the power already is, where the sun never sets, there is no grid to connect to, and no fuel to buy.

On the economics, it states: "The marginal cost of energy for our AI compute satellites will be minimal because our satellites are powered by solar arrays in space."

The filing does not hide the risks.

Timelines will slip, capital requirements will grow, and the core premise (that orbital compute is cheaper than terrestrial at scale) has never been tested.

Execution depends entirely on Starship launching regularly, carrying enormous payloads, and turn around fast enough to move a million metric tons to orbit every year.

It’s expected that Starship will begin payload delivery to orbit in the second half of this year. Getting from there to thousands of launches a year is a different order of problem.

The orbital compute thesis needs Starship to become routine.

#3 The Anthropic deal and the Cursor option underline GPU scarcity

This month, SpaceX signed a deal with Anthropic for access to its data centers for $1.25 billion a month through 2029. Over the life of the contract, that is roughly $45 billion from a single customer. 90 day’s notice and either side can walk.

A month earlier, SpaceX signed a deal with Cursor, the AI coding tool used across a large portion of Fortune 500 companies. SpaceX is providing compute capacity and collaborating on model development, while holding an option to buy Cursor outright later this year for $60 billion. If it decides not to buy, it pays $10 billion for the joint work and walks away.

The S-1 states that SpaceX expects to sign more deals like these.

The underlying reason it can is that SpaceX built its data centers so fast it had more capacity than it could use internally, so it rents out what it does not need. The first facility came online in 122 days, the second in 91.

The rest of the industry takes roughly two years to build something comparable.

When you build that fast, you end up with spare rooms, and SpaceX is charging by the month for them.

The risks are meaningful. $1.25 billion per month from a single counterparty with a 90-day exit clause is real concentration risk. As SpaceX's own AI products grow, the spare capacity available to rent shrinks. And regulators are paying close attention to how the Musk companies do business with each other.

But take a step back from the deal mechanics.

Anthropic builds Claude, one of the leading frontier AI models. Cursor is among the most widely adopted AI coding tools in enterprise software. Both went to SpaceX for compute.

That tells you where the supply constraint sits.

#4 Starlink is cutting prices on purpose before Amazon shows up

Starlink generated $11.4 billion in revenue in 2025, up 50% on the prior year. EBITDA was $7.2 billion. It had 10.3 million subscribers across 164 countries as of March, roughly double where it was a year earlier.

Average revenue per user has fallen from $99 per month in 2023 to $66 in Q1 2026. Most coverage treats that trajectory as a warning. But it’s the opposite in my view.

The filing describes Starlink today as "The sole low-latency network available globally."

SpaceX is using that monopoly the same way Amazon used AWS in its early years: cutting prices to lock in customers and build scale before serious competition arrives.

The telling detail is that the EBITDA margin is holding even as prices fall. Scale efficiencies are outrunning the price cuts.

On the 18% Average Revenue Per User decline itself, Sam Korus of ARK put it well this week: even as prices fall, Starlink’s income from operations has surged 843% over 3 years, and the cost of putting bandwidth into orbit has dropped roughly 15,000-fold since 2004.

SpaceX is deliberately trading near-term price for scale while the underlying cost curve bends sharply in its favor.

Starlink is the recurring cash engine funding everything else in the stack and has been since the first paying customer signed on.

Amazon Leo (formerly Project Kuiper) is now in enterprise beta with Verizon, AT&T, Vodafone, JetBlue, and NASA already signed on, targeting full commercial launch this year.

Amazon has deep pockets, patient capital, a large enterprise base, and a long track record of subsidizing a product to gain position.

The first 10 million Starlink subscribers were largely people with no other option. The next 10 million will have one. That is a harder problem, and Amazon knows how to compete on price.

Starlink’s lead is extremely strong but Amazon Leo is not a footnote.

#5 The CEO's biggest payday requires a 1-million inhabitant city on Mars

The board approved a package of one billion performance-based restricted Class B shares for Musk, split across 15 tranches tied to rising market cap milestones, from $500 billion up to $7.5 trillion.

But every single tranche carries the same second condition:

"A permanent human colony on Mars with at least one million inhabitants"

Market cap alone unlocks nothing. The entire package sits frozen until Mars is achieved.

That’s the most demanding performance condition ever written.

A CEO whose largest personal financial outcome requires building a self-sustaining city on another planet spends differently than one whose bonus resets every December.

Every dollar going into Starship reusability, life support research, and long-duration mission infrastructure is rational given that incentive structure.

The governance concern is real. Voting control, concurrent leadership across multiple companies, no independent check on capital allocation decisions. That tension between long-horizon conviction and concentrated control is part of the deal.

For anyone buying at IPO, the honest question is whether your definition of patient capital and his are measuring time in the same units.

The IPO is a floodlight. Everything underneath SpaceX will become visible.

SpaceX going public next month reprices everything around it.

The parts manufacturers, sensor builders, and propulsion companies that this infrastructure depends on have been building in relative obscurity. That ends when the world's most watched IPO makes orbital infrastructure the defining investment category of the next decade.

At Cape Fear Ventures, we have been building positions there for the past few years.

Scout Space gives satellites real-time awareness of their orbital environment, directly addressing the collision-avoidance problem SpaceX faces with hundreds of thousands of Starlink maneuvers each year.

Raven Space Systems manufactures aerospace-grade composite parts in days rather than months, cutting lead times that have historically grounded programs entirely.

Juno Propulsion is developing rotating detonation engines that could make frequent orbital repositioning economically viable rather than cost-prohibitive.

Samara Aerospace builds solar panels that double as satellite steering systems — delivering power and agility on a single platform with no tradeoff.

Lux Aeterna is working on applying SpaceX-style reusability logic to the satellite itself.

If the stack described in this S-1 filing gets built at anything close to the stated scale, the demand for what these companies make explodes with it.