Recorded with the SpaceX IPO roughly 24 hours away, this episode was Mach 33 Research's chance to put intuition behind the numbers it published earlier in the week. The article presenting the model's outputs had gone out with little commentary, so the hour was spent walking through what the figures actually mean, fielding a steady stream of audience questions along the way.
The team opened with its standard framing: the figures come from a fundamentals-driven financial model built over the past several months, not a prediction of price action. Mach 33 Research sells rigorous, physics-grounded analysis to institutional clients with three-to-five-year-plus horizons, and it does not intend to publish price targets. Guessing what an irrational market does next quarter is somebody else's job. (For those keeping score anyway, Oppenheimer put out a $190 target the day of recording, with street ranges running roughly 160s to 190. The more interesting part of that note, in the team's view, was that sell-side analysts are starting to articulate the technical opportunity, not the dollar figure itself.)
Two million tons, and a nod from Elon
The week's launch-mass animation drew a comment from Elon Musk himself. The visualization is operationally driven, building up launch-by-launch, and implies SpaceX will have launched about 2 million tons to orbit by 2040, with roughly a million tons active after deorbits. Musk's response: SpaceX is targeting a million tons per year by 2030. That puts the model at least 5x below the company's own ambitions, which is by design. Mach 33 Research layers conservatism deliberately, on the logic that space delays happen and the numbers look attractive without heroic assumptions. One striking coincidence from the base case: about 370 kilotons of active mass generates $1 trillion in annual revenue.
How the model works
The valuation rests on a 5,000-trial Monte Carlo across roughly 350 assumption ranges. Each simulation draws randomly within sensible bounds for every uncertain input, so areas of greater uncertainty widen the output distributions, and the curves flatten the further out in time they go as compounding stacks the upside ranges. The team reads medians rather than peaks, since long right tails pull the averages around. Most published SpaceX models, by contrast, hardcode a single case or a few ranges; the Monte Carlo approach shows how hundreds of variables interact.
The choice of valuation method changes the story considerably. Gordon growth in perpetuity, a conservative method suited to companies that have stopped growing, values near-term certainty: under it, Starlink contributes $1.8 trillion against compute's $1 trillion. Switch to EV/EBITDA, which the team considers the most suitable consistent lens for SpaceX over the next 15 years, and the order flips: Starlink contributes about $2 trillion while compute contributes $3.3 trillion, with an upside tail reaching $10 trillion. By 2040, with every star aligned, compute alone could be worth $37.9 trillion within the model's assumption ranges.
What does star alignment mean? The tornado chart puts four variables on top: ODC adoption steepness, demand for orbital compute, the multiple the market sustains, and how cheap chips get per operation, with GPU-hour pricing close behind. If demand proves effectively unbounded, multiples hold, cost per flop keeps falling fast, and GPU-hour pricing stays strong, that is where the ultra-high outcomes live.
The episode also showed off a 3D scatter tool plotting all 5,000 simulations against the IPO price floor, with selectable variables on each axis and a time slider through 2040. It surfaced things like the correlation between chip price efficiency and GPU-hour price CAGR, and one amusing rogue simulation that started high and sank despite maximum ODC adoption. The team could not immediately explain what went wrong in that particular future, and resolved to track the dot down.
Where the demand comes from
Asked where trillions in compute demand could possibly originate, the team pointed past knowledge work, which it considers the small part despite its own heavy daily usage, toward physical AI. The token math is the tell: a paragraph of text is hundreds of tokens, while live video streaming for a humanoid robot runs into the millions. A figure attributed to Musk suggests something like 30% of a humanoid robot's compute will need to come from off-board, and space offers the global coverage to serve it.
The framing offered for investors: place yourself on the demand spectrum. If AI demand fizzles within a year or two, orbital data centers do not pencil. If demand keeps compounding, the math works. The breakpoint where compute must go to orbit sits somewhere between 100 gigawatts and a terawatt of total demand, by the team's own admission a guesstimate.
Cash allocation, debt, and the $75 billion surprise
The model allocates capital each year based on prior-year IRR, a blunt tool that works. Starlink absorbs about $1 trillion of cumulative allocation by 2040 and throws off at least $300 billion a year from the mid-2030s onward. The constellation matures in the early 2030s, enters a harvesting period, then catches a second wave as the addressable market expands mid-decade. Crucially, once Starlink matures, all that launch capacity pivots to lofting compute, which is when AI revenue explodes. Mars and the Moon absorb $300 to $400 billion over 15 years and, in the model's unsentimental view, function as a cash sink with no IRR before 2040.
Debt enters through two facilities, for TerraFab and for ODC satellites, the two bottlenecks that throttled growth in earlier model runs. ODC satellites are expensive (more expensive than Starlink satellites until the late 2030s, when they approach the roughly $1 million per unit expected of Starlink V3), and around 2033 the model starts drawing on the ODC facility to fill launch capacity stranded by Starlink's maturation. The cumulative draw is about $50 billion between 2033 and 2038, split $40 billion for satellites and $10 billion for TerraFab. Against $2.7 trillion of total invested capital, that is not very much.
The recent news that the IPO raise will be $75 billion, all primary, forced a revision. The team had assumed a meaningful secondary component (an intern did not hesitate to correct the record), so the model now carries a much larger near-term cash pile, and the TerraFab facility draws only about $10 to $12 billion where it previously drew $20 billion. The roughly $20 billion bridge IPO loan, notably, emerges from the model organically without being hardcoded.
One sharp audience question probed whether freed-up 2030s launch capacity might find higher-IRR payloads than SpaceX's own AI satellites. The team conceded this touches one of the hardest problems in the build: customer launch IRR in the early 2030s is in fact very high, higher than launching ODCs initially. The catch is market size. The AI TAM dwarfs the customer launch TAM, so a two-layer allocation was needed, and a one-off launch cannot compete with years of high-margin recurring revenue. The team flagged customer launch market sizing as the model's current soft spot, with sub-segments like point-to-point and Starfall still owed proper treatment in future iterations.
SpaceX is the space industry
A question about whether new categories will keep opening up that only SpaceX can access prompted the episode's broadest thesis. SpaceX holds a 5-to-10-year lead on launch capability and operational scale, will always prioritize its own payloads, and exerts so much gravity that everyone else fights for whatever it chooses not to do. The evidence offered: StarCloud was widely dismissed until Musk declared SpaceX would do orbital data centers, which changed the trajectory of the entire segment overnight, and the Mars-to-Moon pivot just shifted opportunity windows for whole categories of technology companies by perhaps a decade. Neocloud investors got a taste of this when SpaceX announced its multi-year deals with Google and Anthropic. The practical takeaway for investors: without a strong SpaceX thesis, it is hard to claim a space thesis at all, because every other company operates inside SpaceX's orbit. Unmodeled categories like in-space manufacturing and asteroid mining remain a cherry on top, deliberately excluded as another layer of conservatism.
On Starship as a space station: plausible only in the early scarce years for premium scientific use, but generally a poor use of an asset designed for reuse. A manufacturing hub that goes up, pulls fiber, and comes back down struck the hosts as the better version of that idea.
Audience science corner
Two orbital mechanics questions got satisfying answers. ODCs sit in polar, dawn-dusk sun-synchronous orbits because that narrow band keeps the solar arrays in perpetual sunlight, avoiding the mass penalty of batteries, and operators climb to higher altitudes within the band to avoid collisions, which explains the recent rush of FCC filings staking claims to that thin slice of real estate. And the design detail that recently clicked for the team: the radiators on the AI satellite (the constellation now nicknamed the Halo) are extremely thin, with only the thin edge facing the sun, so the broad faces stay perpendicular to solar heating and dissipate maximum heat to deep space.
Parting thought
The model is a newborn, not a finished product. The team plans to keep iterating, welcomes feedback and critiques, and signed off looking ahead to the next 24 hours of SpaceX news.