The second episode of our Private Podcast was billed as a walkthrough of our new 80-page Orbital Data Center (ODC) report, the product of seven months of diligence, more than a dozen analyses, and a fresh subsystem-level cost model. It delivered exactly that, with Aaron and Vlad threading the technical, economic, and capital-markets implications into a single thesis: by the end of the decade, the cost of compute in space crosses the cost of compute on the ground, and the company best positioned to exploit that crossover is the one already building the launch supply to do it.
Vlad opened in the voice of Douglas Adams, an LLM-generated pastiche he couldn't resist reading aloud: "the problem with building data centers being that they require electricity, the problem with electricity being that it has to come from somewhere, and the problem with somewhere being that it is almost always somewhere else, connected to the place you actually needed the electricity by a series of wires, transformers, substations, environmental review boards, regional transmission organizations, cost allocation disputes, and at least one retired school teacher named Margaret." Margaret became our running mascot for the terrestrial power wall.
Two News Items Worth Tracking
Before getting to the report, we flagged two developments. Blue Origin's third New Glenn launch has been formally classified by the FAA as a mishap, with one upper-stage engine producing insufficient thrust and the AST SpaceMobile payload failing to reach its intended orbit. Our read: a one-to-three month grounding is the modal outcome, with tail risk longer. AST recovers via insurance on this single launch, but the larger story is the cascading delay to constellation deployment for both AST and Amazon Leo, since New Glenn was the heavy-lift workhorse outside SpaceX's Starlink-saturated manifest.
The second item is today's FCC vote on Equivalent Power Flux Density. The headline number being repeated in coverage, a 7x increase in capacity, is being widely misinterpreted. It is not a 7x increase in bandwidth. It is permission to put down 7x more power per FCC-defined cell, and because beams cannot overlap on the same frequency, the effect is concentrated in suburban density bands. Dense urban areas remain beam-limited; rural areas don't need the additional capacity. Suburban customers are where SpaceX picks up new addressable subscribers. We will publish a quantitative read in next week's analysis.
The Triangular Bottleneck
The physics chapter of our report frames orbital compute economics as a three-way constraint: compute, solar generation, and thermal rejection. Improving any one without the other two yields nothing. A breakthrough in solar cell efficiency just turns the satellite into a giant radiator, because in vacuum, energy in must equal energy out. This is the central pushback against the most common naive intuition, that space is cold, so cooling must be easy. It is cold, and that helps, because radiator efficiency scales with temperature differential. But there is no convective dissipation, no atmosphere doing the work, only radiative emission off panels you have to mass-budget for.
A consequence we emphasized: orbital compute satellites are architecturally distinct from comms satellites. Roughly 40% of a SATCOM bus mass goes to the communications payload and the precise pointing it requires. A compute satellite, by contrast, does not need to be precisely Earth-pointed (its laser links handle constellation routing), which lets it run thinner solar panels, more flex, and a more aggressive thermodynamic profile. The recently surfaced renderings of SpaceX's AI satellite, with deployable radiators and oversized arrays, look completely different from a Starlink. They are optimized, in our framing, "to turn energy from the sun into intelligence."
The Cost Crossover: 2029 to 2032
The economic core of the report is a comparison between three orbital-compute capex trajectories and a band of terrestrial estimates, indexed across years because each scenario assumes a different chip generation and launch-cost regime. The bear case starts now, with today's chips lifted on expendable Starship at $500 per kilogram. The base case starts in 2028, with an AI5-class chip and $300 per kilogram launch. The bull case starts in 2030, with a Dojo 3 or Terafab-manufactured chip and $100 per kilogram. Terrestrial is anchored to publicly stated cost-per-gigawatt figures from Jensen Huang, Stargate-implied numbers, and Bernstein, escalated at a historically validated 7% per year on the non-chip portion.
We disclosed a material correction live. Earlier versions of this comparison stacked upfront capex against upfront capex without normalizing for useful life. Satellites get five years of service; terrestrial data centers get twenty. After Vlad's initial panic, when he thought our model was 4x off, we realized that roughly 40% of capex on either side is chips, which refresh on a similar five-year cadence regardless of where they sit. The actual error was closer to 2x on the terrestrial side, and the corrected numbers push the crossover into a 2029 to 2032 window in the base case. Our color: this is still the conservative read. A 20-year terrestrial life with no retrofit assumes none of the cooling and density upgrades AI workloads will demand, and assumes Margaret stays quiet, which she will not.
The single biggest variable, we agreed, is when the base and bull scenarios actually begin. The horse is reusable launch: we are confident in Starship's timeline and not concerned about that side of the equation. The cart is chips at scale, manufactured by Terafab (the SpaceX/Tesla joint venture), and sold at cost rather than at NVIDIA-style margin. Aaron pressed Vlad on attribution: how much of the cost decline comes from launch versus chips? Not much from launch, was the answer. Satellite cost dominates, and 30 to 40% of satellite cost is chips. Terafab is the lever.
What This Does to the IPO Narrative
The reason this matters now, and not only in 2030, is the SpaceX IPO. Vlad walked through our cash-flow projection: every dollar of Starlink profit recycles into rockets and satellites, and that flywheel saturates around 2030 at somewhere between 40,000 and 100,000 V3 satellites, when the marginal ROI of another Starlink drops below the threshold for redirecting capacity. By that point SpaceX has both a cash position and a launch supply far in excess of what Starlink alone can absorb, and that is the same year the orbital and terrestrial cost lines cross.
Two lazy framings of orbital compute are circulating, both wrong. The first writes it off entirely on the grounds that revenue won't show up in the near term. The second projects hundreds of billions of orbital compute revenue by 2030 because they assume launches start tomorrow. Our view sits in the middle: orbital compute revenue is small relative to Starlink until roughly 2030, but the strategic value to the IPO is not near-term revenue. It is preservation of the growth multiple. A mature Starlink in 2030 is a staple business with a multiple in the high teens. Our earlier modeling pegged Starlink alone at roughly $3.5 trillion at maturity. The orbital data center is what keeps the back half of the 2030s a growth story rather than a cash-cow story.
Aaron added the AI-stack angle. Hyperscalers today appear to make $12 to $17 billion per gigawatt of AI capacity per year. Vertically integrated players capturing inference revenue, what OpenAI looks like, may be closer to $20 to $25 billion per gigawatt. The Cursor acquisition is the strongest public signal we see that SpaceX is targeting that integrated layer, not the GPU-rental layer. If terrestrial Cursor-driven revenue scales aggressively through 2028, the orbital data center is what allows that growth curve to continue once terrestrial power becomes the binding constraint.
The closing observation, almost an aside in the Q&A: by 2033, the largest single slice of Starship launch demand in our model is not Starlink, and not orbital compute. It is depot launches, five tanker flights for every Moon-bound payload, pointing at the next flywheel after this one.
The full report is available to our institutional clients and LPs. A public summary is forthcoming. Next week's analysis will quantify the EPFD impact on suburban Starlink economics.