14 September 2026
Heard In AI

Why AI in orbit needs a launch industry, not just a cheaper rocket

Ramez Naam puts himself between those who dismiss orbital data centers and those expecting an imminent boom. He estimates that launch prices need to fall to roughly a quarter to a tenth of current levels for space-based AI to compete on cost. But cheaper flights leave a separate hurdle: building and launching enough hardware, with permission and reliability to keep flying.

A briefing reports one development when it happens. We correct or clarify it later; a new development gets a new briefing. How our formats work

Elon Musk’s pitch for data centers in orbit gets a split reaction, Ramez Naam observed on Moonshots with Peter Diamandis. Some people say it will never work; others anticipate a terawatt of computing in space. Naam put himself elsewhere: “I’m somewhere in the middle.”

His case turns on two different achievements. Launching hardware must become much cheaper. Then someone must manufacture and lift enough of it, often enough, to build a substantial computing industry in orbit. Solving the first problem does not automatically solve the second.

A data center houses computers that train and run AI systems, among other tasks. Those computers need electricity and a way to shed heat. SpaceX’s proposed orbital version would use solar power, radiators to release heat into space, and laser communications between satellites and the Starlink network. It could bypass a terrestrial grid connection—but it would acquire a dependence on rockets.

The price that would make it competitive

Naam estimates that space-based AI could become cost competitive when launch prices fall to roughly a quarter to a tenth of current levels. That is a back-of-the-envelope threshold, he said, rather than a demonstrated cost: “Nobody knows for sure.”

Cost is not the only reason to go. Naam also sees orbital computing as a hedge against constraints on the ground. If computing demand keeps rising while grid queues, permitting disputes and local opposition prevent construction, paying more to build in space could become worthwhile.

But escaping a data center’s planning dispute does not mean escaping regulation. Naam immediately raised the permissions needed for the launch industry that would serve it: “I think we are not fully internalizing what the scale of this is or the permitting and regulatory challenges with doing a launch at that volume.”

The mass problem

SpaceX’s June 5, 2026 prospectus gives that scale a physical meaning. The company’s long-term ambition is to add 100 gigawatts of orbital computing capacity annually. Its assumptions pair roughly 100 kilowatts of computing power per metric ton with about one million metric tons delivered to orbit each year.

Gigawatts here describe the power scale of the computing hardware, not a direct measure of how much useful AI work it performs. One gigawatt is one million kilowatts. At the prospectus’s assumed power-to-mass ratio, even that smaller deployment would mean about 10,000 metric tons in orbit.

Those tons are not simply boxes of chips. The proposed satellites need electronics able to tolerate radiation and integrated systems for moving and releasing heat. SpaceX describes radiators, vapor chambers and active cooling loops, along with orbits chosen for near-continuous sunlight. The company says its annual deployment ambition requires thousands of launches and a substantial expansion of satellite manufacturing. It also says smaller deployment rates could be commercially useful.

The podcast’s launch counts approached the problem through different assumptions. Naam estimated that adding 10 gigawatts a year would require 1,500 to 2,000 Starship launches annually—several flights every day. He did not spell out the payload mass behind that particular estimate in the exchange.

The hosts offered a separate scenario: with 20 to 30 computing satellites aboard each Starship, they put a 100-gigawatt annual deployment at roughly 30,000 launches a year. That is a flight roughly every quarter-hour. Their count depends on how much computing capacity those satellites carry; it is not interchangeable with Naam’s smaller deployment estimate or the prospectus’s mass-based assumptions.

In each case, the job extends well beyond producing a cheaper rocket. Factories must turn out the computing satellites, and a launch fleet must repeatedly carry them into orbit.

Airline operations, or a grounded fleet

The hosts’ answer was to change the operating model. Viewed as a conventional rocket program, tens of thousands of launches look prohibitive. Viewed as an airline-style service, frequent departures are the point: recover a vehicle, refuel it and fly it again.

They pointed to Starship’s intended capture-and-reuse design as a route toward that model. They also asked whether smaller satellites could eventually increase the computing capacity carried on each flight.

Naam’s reservations concerned permission and reliability as well as scale. He was unsure whether authorities would allow more than about 200 Starship launches a year. That was his expression of uncertainty, not an established regulatory ceiling. Even that rate, using his assumption of 100 tons per flight, would deliver 20,000 tons annually.

He also raised the risk that a failed launch or explosion could ground operations under Federal Aviation Administration oversight. A schedule built around repeated reuse needs more than vehicles that can fly again: it needs dependable turnaround and permission to keep the fleet operating. An interruption matters much more when a deployment plan assumes departures throughout every day.

Why AI demand could help pay for Mars

Naam nevertheless sees a business reason to pursue orbital computing. Elon Musk wants to go to Mars, and making that affordable requires driving Starship’s launch costs toward the marginal cost—the extra expense of flying one more mission.

Frequent flights would spread research, development and infrastructure spending across many launches. That means building tens or perhaps hundreds of vehicles and finding enough customers to keep them flying. In Naam’s view, communications demand through Starlink is not sufficient to finance that scale, and Mars does not yet provide a business model.

“So this is a gift to SpaceX that we have this AI demand,” he said.

His argument is conditional: if AI demand continues growing and construction on Earth remains bottlenecked, orbital computing could become a paying customer for the launch cadence Mars ambitions require. Starship would enable the computing business, while the computing business could help finance Starship’s expansion.

Nor did the panel treat success in space as requiring failure on Earth. In the hosts’ optimistic scenario, frequent rocket departures could coexist with a thriving terrestrial data-center build-out.

Naam’s near-term test was much smaller than SpaceX’s long-term annual ambition. “I think by 2030, if SpaceX has a single gigawatt in space, I will be very impressed.”

Share this article

Go to the original

Sources & further reading

  1. 01

Connected ideas and articles

From the conversation

Podcast episodes

Article history

Updates to this article

Tags

Why Ramez Naam changed his mind about computers on ocean buoys

Ramez Naam passed on Panthalassa’s early Bitcoin-mining pitch, then invested twice in 2026 at much higher valuations. The company now proposes wave-powered AI computing, cooled by seawater and connected by satellite. Its $140 million Series B is intended to support an Oregon pilot factory and northern-Pacific pilots; cheap electricity and longer-lived chips remain prospective benefits.

4 min read

AI money can help nuclear scale—but not guarantee faster power

Ramez Naam sees AI demand as a powerful source of nuclear financing, but doubts new small reactors can supply electricity within the five-year window he considers reasonably predictable for investment. His argument turns on what can be delivered sooner—and whether repeated construction and factory production can make later plants cheaper.

6 min read

Why flexible AI data centers could get power sooner

Expensive AI chips can sit idle while data centers wait for grid connections. On Moonshots, energy investor Ramez Naam argued that accepting less grid power during peak demand could shorten that wait. Workload scheduling and batteries offer two ways to do it, but national estimates of spare capacity are not promises of power at a particular site.

7 min read

Before a grand AI treaty, Graylin wants hotlines and shared safety tests

Alvin Wang Graylin proposes a practical starting point for US–China AI cooperation: an emergency hotline, shared safety tests and an agreement to keep talking. Speaking personally on Moonshots, ahead of a September 24 dialogue described in the episode, he connects those steps to a larger bargain—financing AI deployment abroad while spreading agreed safety standards.

6 min read