Episode four led with a framework rather than the news, because the framework turned out to be the cleanest lens for reading the week's headlines. Vlad walked through the team's first stab at modeling the AI revenue stack, the layer cake that sits on top of the infrastructure work already published in the recent ODC report. The headline number: roughly $41.6 billion of revenue per gigawatt of compute, split across five layers.
Energy is the smallest slice at about 1.7%, even though it remains the binding physical constraint on everything else and the reason orbital data centers exist at all. Chips run around 16%. Infrastructure (the data center on Earth, the satellite bus in orbit) sits at roughly 20 to 22%. Models command the largest share of revenue, but the team expects them to carry the thinnest margins once next week's cost-side analysis lands, since training has to be amortized out of inference revenue. Orchestration, the layer that turns a pattern matcher into a product, sits on top.
Vlad and Aaron spent real time on the model-vs-orchestration line, and were upfront that it is more gradient than border. The pure model is a probabilistic pattern matcher. Orchestration is everything that wraps it: skills, plugins, computer access, multi-agent workflows, the harness that lets a stochastic engine play nicely inside the deterministic software world. Vlad's personal example landed the point: he uses Claude not because the base model is best, but because the orchestration around it is. Closing that gap is exactly what XAI is now racing to do, with Grok integrations rolling out and skills teased.
The framework also caveats itself. It is terrestrial-orbital agnostic, since chips and models and orchestration look the same on either side of the Karman line. It is also inference-focused, which is the relevant lens for ODCs in the near term.
How the news reads through the stack
SpaceX AI as a consolidated brand. XAI is being formally folded into SpaceX under the SpaceX AI label. Aaron read this as a sub-divisional rename rather than a top-level rebrand of the parent company, but the signal is consistent with the stack picture: SpaceX wants to occupy the whole column, from energy at the bottom to orchestration at the top.
Three business models, run in parallel. Model A is the full stack play, where SpaceX owns everything from orbital energy through Grok and, eventually, its own orchestration layer. Model B is the infrastructure-only play, exactly what was announced with Anthropic on Colossus last week, capturing roughly 40% of the stack. Model C is the hybrid, which is where the Cursor relationship sits today: SpaceX owns most of the column but lets a third-party orchestrator sit on top, capturing perhaps 80 to 90%. The team is not surprised SpaceX is running all three at once. It buys scaling flexibility, and the long-run mix will likely drift the way Falcon 9 did, where internal Starlink launches climbed from roughly half of cadence to about three quarters as the full-stack option matured. Whether Grok eventually closes the orchestration gap organically or acquires Cursor outright is the open question on Model A.
Google's launch rumors. Reports surfaced this week that Google is in talks to launch ODC hardware via SpaceX. The articles were vague, and the news may simply be a rewrite of the existing Project Suncatcher prototype with Planet Labs. But the strategic logic is real: Google owns the chips (TPUs) and the model (Gemini), and it lacks the infrastructure and energy layers in orbit. If they want to scale, they need a bus and a launcher. The most likely partners are SpaceX itself, StarCloud, or K2, given Planet Labs is an Earth-observation company without obvious experience building high-power data-center buses. The tell, if Google is serious, will be FCC filings for sun-synchronous orbit. The upside for them is real: more efficient chips translate to more revenue per kilogram lifted, which the team's rough math pegs at $5 to $15 billion of additional revenue per gigawatt.
Aetherflux is now Cowboy Space. A $275 million Series B led by Index Ventures at a $2 billion valuation, paired with a pivot from space-based solar to orbital data centers. The architecture is novel (the rocket's second stage becomes the ODC, with radially deployed solar arrays), but a 2028 first-launch target on a still-to-be-built rocket invites healthy skepticism. The implied vertical integration down to propulsion runs the opposite direction from where Google appears to be heading, and the first test in 2027 is a power-beaming demonstration that has little to do with ODCs.
Starship V3 next Tuesday. May 19 is the test launch, and it matters more than the IPO narrative implies. ODCs will not move the income statement soon. Starlink will, and the V3 constellation depends on Starship V3 as its workhorse, with bandwidth expected to grow 50 to 100x by 2030. Every slip in this launch slips that ramp. The vehicle has been stripped down toward a payload mass fraction worthy of operational flights, which is the right direction. Raptor 3 performance is the variable to watch.
EchoStar approved. The FCC voted through the EchoStar spectrum transaction. More spectrum sales are likely, since spectrum tends to gravitate toward whichever operator can actually pump traffic through it. Amazon may be the more aggressive buyer from here, given SpaceX already holds a deeper inventory.