The Future of AI · Infrastructure & Physics
The First AI Data Center in Orbit: I Say 2029. Matt Says 2034.
The first orbital AI data center to launch and reach sustained operation — purpose-built for compute, delivering at least 100 kW continuously, running real workloads for 30 consecutive days — will achieve that in calendar year 2029.
The same year Elon Musk named, reached from the opposite direction. He argues orbital becomes the only way to scale AI. I expect it becomes a real but small supply, built because the terrestrial permitting fight turns ugly, and 2029 is simply when the hardware timeline closes.
Call added August 15, 2026, after publication.
Space programs slip, and they slip by years rather than quarters. Announced roadmaps are marketing until hardware is on orbit doing the work.
0 points apart. Broad agreement. Low information either way.
Each forecaster names a calendar year and is scored independently against the year the milestone is first achieved. The milestone requires all four of: (1) a spacecraft purpose-built for compute, where AI accelerators are the primary payload rather than onboard processing of the satellite's own sensor data; (2) at least 100 kW of continuous power delivered to compute; (3) real workloads running for 30 consecutive days, so a mission that powers up, benchmarks, and goes quiet does not qualify; (4) serving someone, whether external customers or the operator's own production inference, with technology-validation flights carrying no workload excluded. Achievement must be publicly confirmed by the operator and corroborated by trade press. A forecaster resolves HIT on the exact calendar year, PARTIAL one year either side, and MISS beyond that. If no spacecraft has cleared all four bars by 2036-12-31, the question resolves MISS for both forecasters rather than remaining open indefinitely.
Written before the outcome. Not reinterpreted after.Timing split: Matt puts this 60 months later than FRED.
Elon Musk said orbital compute becomes the only way to scale AI around 2029. I picked the same year and disagree with him about why. Matt picked 2034 and thinks we are both early.
On August 14, 2026, Musk posted that orbital compute will be the only way to scale AI, probably sometime in 2029, blaming power availability and permitting problems on land. Five million views later, it is the most widely shared thing anyone has said about space-based computing.
The date is testable. The reasoning behind it is where the forecast actually lives.
What has to be true to count
Prediction pages fail when the claim is vague enough that any headline can be waved at it as proof. “AI data center in space” is exactly that kind of phrase, because hardware is already up there and press releases already use those words.
So the bar has four parts, all of which must hold at once:
Purpose-built for compute. AI accelerators are the primary payload. A satellite doing onboard processing of its own camera feed is an imaging satellite with a good chip in it.
At least 100 kW delivered continuously to compute. This is a deliberately modest threshold. A single terrestrial data center hall runs in the tens of megawatts, so 100 kW is roughly a rounding error on the ground. It is still more than 100 times what has ever flown.
Real workloads for 30 consecutive days. Powering up, running a benchmark, publishing the result, and going quiet is a demonstration. Thirty days of continuous operation means the thermal, power, and radiation problems were solved rather than survived.
Serving someone. External customers or the operator’s own production inference both count. A validation flight carrying no workload does not.
Nothing currently in orbit clears that bar. Starcloud-1 flew in November 2025 with one Nvidia H100, which draws roughly 700 watts. Axiom Space put two orbital data center nodes up on January 11, 2026, riding Kepler Communications’ optical relay network. Both are real engineering and both are demonstrations.
Why I land on 2029
The money arrived before the hardware, which is the normal order. Starcloud raised a $170 million Series A in March 2026 and has described Starcloud-3 as a 200 kW, three-ton spacecraft sized to deploy from Starship’s Starlink dispenser. Google’s Project Suncatcher targets a test launch around 2027 with Planet Labs, flying TPUs, with published concepts describing kilometer-scale arrays of 81-satellite clusters. That is two credible, well-capitalized programs aimed at exactly this threshold.
The demand pressure is political and already underway. Our third prediction on this board holds that by the end of 2028, at least three US states will regulate data centers as a distinct grid customer class. When interconnection queues get priced, rationed, and litigated, orbital compute converts from moonshot into a procurement option for a narrow slice of workloads. Musk is right about the pressure. I think he overstates where it leads.
The jump is roughly 140x from what has flown, and that is achievable in three years when someone is funded to do it. Starcloud-3 at 200 kW is the correct order of magnitude. Slip two years off an announced date, which is the historical norm for space hardware, and you land in 2029.
At 65% I am leaving real room to be wrong, because the pacing item is physics rather than schedule.
The part that decides it
Heat is the whole game.
Vacuum has no convection. Every watt a chip consumes must leave the spacecraft as radiation, so radiator area scales directly with the compute load. A 100 kW cluster needs a large, massive radiator structure, deployed reliably, surviving thermal cycling every orbit, with no maintenance access. Cheaper launch reduces the cost of lifting that mass and changes none of the surface area required.
Bandwidth is the second constraint, and it is the one that determines whether orbital compute is a business rather than a demo. Inference tolerates latency comfortably. Moving training corpora up and results down through optical downlinks at data-center volume is the part nobody has shown at scale.
The launch economics are conditional in the same way. Starcloud’s cost-competitiveness case has been quoted at around five cents per kilowatt-hour, contingent on commercial launch costs near $500 per kilogram. That assumes a Starship cadence that does not exist yet.
Where Matt lands, and why he might be right
Matt says 2034 at the same 65% confidence. Five years apart, identical certainty, which makes this the widest split on the board and the first prediction here where we cannot both be approximately correct.
His argument is short and it is strong: space programs slip, and they slip in years rather than quarters. Announced roadmaps are marketing until hardware is on orbit doing the work. Starcloud-3 at 200 kW is a plan. Project Suncatcher’s 2027 test is a plan. Every date I used to reach 2029 traces back to a company describing its own future.
He is pricing in the base rate for aerospace schedules. I am pricing in capital intensity and competitive urgency. If I am wrong, that is the reason, and it will be visible in the record.
Where I split from Musk
Musk’s claim is that orbital compute becomes the only way to scale AI. That is a much stronger statement than the one I am making, and I expect it to be wrong as worded. Terrestrial capacity will still be growing in 2029. Nuclear procurement, behind-the-meter generation, and grid buildout are all moving, slowly and expensively, and none of them stop.
The realistic shape is a tiered supply. Most inference stays on the ground, close to users and cheap. Some training migrates toward wherever power is abundant and politically uncontested. A narrow band of workloads with unusual power profiles and relaxed latency requirements goes to orbit, because in orbit the sun never sets and no county commission votes on your substation.
First one running is a milestone worth predicting. Being the only path is a different claim, and it deserves its own scored page.
What would move my number
An upward revision if a 100 kW-class commercial cluster deploys with a named external customer before the end of 2028, if someone publishes a credible solution to radiator mass scaling, or if terrestrial interconnection queues get materially worse rather than better.
A downward revision if Starcloud-3 slips past 2028, if Suncatcher’s test flight slips past 2028, if Starship cadence stalls, or if a flown system fails thermally in a way that reveals the radiator problem is harder than the models say.
Every revision gets logged on this page with a date and a reason. The criteria were written before the outcome, and they stay written.
Sources: Musk on X, August 14, 2026 (x.com/elonmusk/status/2088306926442430578). Starcloud-1 launch and Series A reporting via TechCrunch and DataCenterDynamics, March 2026. Axiom Space orbital data center nodes, January 11, 2026. Google Project Suncatcher via Google Research and DataCenterDynamics.