Aaron Burnett: SpaceX's Elon Web Services could hit $100B+ ARR per gigawatt

Elon Web Services emerges as a major new revenue line

The host opened with what he called breaking news: Elon Musk detailed the buildout of "Elon Web Services," SpaceX's AI compute rental business, which the host said hit his feed around 10pm Pacific the night before. He said that if all the capacity being built were rented out, it could generate "north of a hundred billion in ARR by the end of the year." Aaron Burnett, the guest, said the pace of deployment was "on pace for kind of what they had said originally, maybe slightly better," putting the buildout in a "two to two and a half" gigawatt range by year-end, referencing SpaceX's earlier earnings-call guidance of roughly two gigawatts.

Burnett distinguished training from inference: Elon has estimated about 10% of compute goes to training new models, with 90% used for inference, i.e., serving queries to end users. He said hyperscaler clients are reportedly paying "somewhere between 30 and 50 billion dollars per gigawatt," with the host speculating the fourth major client could be Microsoft given a reported spend "north of a billion a month," though Burnett said he hadn't speculated much on which company it is, also floating Meta as a possibility given its compute needs.

On the question of an AI bubble, Burnett said he does not see inference demand slowing, arguing that as chip efficiency (like the upcoming Vera Rubin generation) increases token density, "the demand is still meeting that efficiency." The host cited unverified word from a Silicon Valley source that Elon had "secured 30 to 40% of the Vera Rubin supply," which Burnett said "makes sense" given that coherent, ready-to-use SpaceX infrastructure lets Nvidia's chips reach higher utilization than competitors.

The four-legged revenue stool: launch, comms, AI compute, and defense

The host outlined SpaceX's business as a "four-legged stool": rocket launches, global communications/internet (including direct-to-cell), AI compute, and government/defense spending including Starlink for military use and Star Shield. He suggested that if Starlink captures "30 to 40% of the global cell phone market," that alone could be "a $5 trillion market cap business." Burnett noted SpaceX's earnings calls generally group revenue into launch/defense, connectivity, and AI/enterprise, and said his own models weight Starlink heavily for near-term IRR while AI compute is viewed as the larger, longer-term opportunity.

The host also referenced Tesla's "Grok" integration and the "Optimus" TAM figure of "$28.5 trillion" from a prior model, tying robotics into the broader convergence thesis discussed later in the show.

Starship reusability and the falling cost-per-kilogram curve

Burnett explained the cost curve to orbit as an asymptote driven first by reusability and then by manufacturing scale. He said a vehicle starting at "$250 million" is cut in cost each time it's reused — halved after one reuse, thirded after two — and that current boosters have "already used them 30 to 40 times." He said the curve could approach "$10 to $20 per kilogram," with the first phase of cost reduction from reuse and the second from building many vehicles, potentially dropping the initial build cost to "maybe 50 or even less million dollars."

On competition, Burnett said catching up in hardware is far harder than in software, estimating rivals are "five years in the best case" and "more like 10 or beyond" in the worst case behind SpaceX, given its launch and satellite counts. He argued there should be an "execution premium" priced into space companies based on demonstrated launch history, noting SpaceX is second only to Russia in historical launches despite Russia's 70-year head start.

Space-based data centers: physics, cooling, and Kessler syndrome

Asked whether orbital data centers are viable by 2027, Burnett said the core challenge is thermal dissipation via radiators rather than solar power. He said satellites are moving from "tens of kilowatts" to "100 plus kilowatts, 200 kilowatts" of power, requiring much larger radiators positioned to face the coldest parts of space rather than the sun, using sun-synchronous orbit. He said the new "AI1" satellite will have "substantially larger" radiator and solar panel area than the current Starlink V3 satellite and roughly twice the mass, though it fits in the same launch volume. He also said solar panels in space can produce "five times more energy" than on the ground because there is no atmosphere to filter sunlight.

On Kessler syndrome, raised by a viewer, Burnett called it "a valid concern" but compared existing orbital debris to an ocean garbage patch — present but diluted across a very large area — and noted active debris-management and collision-avoidance engineering solutions exist but have not been seriously funded. Another commenter noted satellites already perform collision avoidance "sometimes three per day per satellite," which went unchallenged.

Do frontier AI labs' thin moats threaten SpaceX's compute business?

The host raised concern about the "fungibility" of frontier AI models like Anthropic's and OpenAI's, noting a claim from an all-in podcast guest that model weights "can fit on a USB thumb drive." Burnett agreed these labs lack a strong moat but argued that even if they falter, Grok itself is "a cheap alternative" and threat to them, and that his own team's token usage — "a billion tokens a week" for himself alone — suggests demand keeps rising via agentic use. He estimated that if SpaceX monetizes Vera Rubin capacity internally via Grok at even 50% utilization, it could generate "25 to 50 million per rack," equating to "100 to 200 billion per gigawatt," higher than the 30-50 billion external rate.

The host separately noted that 80% of global token usage comes from open-source, largely Chinese models rather than OpenAI or Anthropic, and cited Palantir's Alex Karp as warning enterprises against handing data to frontier labs, arguing this favors "sovereign AI" buildouts that would still need SpaceX-style compute.

Mach 33's MCP tool and AI-agent research usage

Burnett described his firm's "Mach 33 MCP," a model context protocol connector letting AI agents like Grok query his research directly rather than requiring users to read reports. He said fewer than 1% of his newsletter's users have set it up, but those who have show "15 times more usage per user" than traditional readers. He framed this as evidence that agentic AI usage, not just human reading, is driving token demand upward "in one direction."

Pricing models point to sharply higher valuations by 2028-2030

The host ran through several of his own valuation models, describing them as deliberately conservative ("sandbagged"). An early sum-of-the-parts model put SpaceX at "$2.8 trillion," implying a "$210" stock price; a more detailed version reached "$1,383" by 2030. He cited Morgan Stanley's forecast of "$3.5 trillion" revenue by 2040, and Elon's own counter-claim of reaching that figure by 2033, alongside talk of "100 billion ARR by December." He also cited Semianalysis's compute-only target of "$457" by Q4 2027, and his own "Falcon math" showing per-rocket build cost of "800k" against lifetime return of "13.3 million."

For Starship specifically, using what he called "200 launch math" (30,000 tons to orbit), he modeled lifetime revenue of "$999 billion" from 200 launches, which at a 20x sales multiple implies a $20 trillion valuation and a "$1,500" stock price, or "$40 trillion" and "$3,000" at 40x — noting Elon has talked of "10,000 launches by early 2030s." His Monte Carlo blend of these models produced a sandbagged target of "$527" by 2028, described as "3.61x from where we are today," and "$92" by 2030, described as "6.18x." Burnett said these figures were "right in line" with his own models, though he suggested the host's Starlink assumptions may be conservative, noting his own team's estimates run higher than the host's cited "170 billion range by 2030" for Starlink.

Burnett summarized his framework as treating a roughly "1.75 to two trillion" valuation as a predictable "Starlink bond," with everything above that representing open-ended growth optionality. Neither speaker offered a single, reconciled target figure, and the two agreed to compare and stress-test their full models in a follow-up episode the next week.