What a free rocket wouldn’t fix

Picture an expensive rack of accelerators waiting for permission to turn on. The chips have arrived. The electricity hasn’t. Suddenly, putting the rack above the atmosphere sounds less eccentric.

Louisiana gives that idea a much bigger stage. SpaceX’s announced project opens the possibility of an industrial launch system rather than a scarce flight every so often. More sites could change how quickly launch gets cheaper and how much equipment can reach orbit. The operating ramp still has to happen.

If the rocket were free, would it work?

Under the current accounting, no. Ground compute costs about $11,434 per IT-kilowatt-year. Orbit costs about $13,920 before the rocket gets paid. The gap is already sitting inside the box.S276

Orbital non-IT hardware plus its extra IT and spares burden comes to roughly $4,497 per IT-kilowatt-year. Ground’s entire power-and-facility bill is about $2,701. Orbital operations add another $689.

Sunlight doesn’t settle that invoice.

A pad count is easy to visualize. The less photogenic work happens in a factory, on a balance sheet and in the replacement budget. Louisiana could help make that work possible by providing a much larger route to orbit. Treating the route as the destination would still skip the economics.

That’s the interesting finding. It gives the enthusiast something harder to prove than a cheap launch and gives the skeptic something more useful than an expensive rocket to point at. The opening for orbit is a different cost stack, assembled at industrial scale, competing against persistently scarce terrestrial power.

The physical attraction is straightforward. Orbital solar can collect energy for much more of the year than a ground panel. Starcloud’s design assumptions use roughly 95% availability against 24% for ground solar, producing about 8,322 versus 2,102 kilowatt-hours per peak kilowatt annually. Nearly four times the yield is worth investigating.S204

The temptation is to compare the best feature of orbit with the worst feature of Earth. A satellite gets steady sunlight; a terrestrial project gets a queue. Put complete systems alongside each other instead. The ground operator can change the power source, and the orbital operator has to carry the supporting equipment.

A firm grid connection is the relevant economic competitor, though. Radiators still have to shed the heat. Eclipse management and degradation still need hardware. A splendid solar panel doesn’t automatically make a splendid data center.

The supply pressure on Earth is real enough to justify the investigation. ERCOT’s July 29, 2026 presentation reported about 474 GW of large-load interconnection requests as of June, roughly 90% from data centers. Requests can evaporate. November’s approximately 177 GW response population came from a different information-gathering exercise; subtracting the figures would manufacture a queue collapse.

Nor has cheap launch already arrived as a customer offer. SpaceX’s live sheet, checked November 5, lists $74 million and a maximum fully expendable LEO payload of 22,000 kg. Dividing gives about $3,364/kg. That maximum-payload reference isn’t a reusable Starship invoice for a compute mission.S266

What are we actually paying for?

Useful computing over an asset’s life. That means financing the equipment, keeping capacity available and replacing what fails, as well as supplying electricity.

The base assigns ground power $87/MWh, facility capital $12,500 per IT kilowatt and a power usage effectiveness of 1.35. Annualized IT charges are $1.81/GPU-hour on Earth and $2 in orbit, before the model’s separate 5% ground and 20% orbital provisioning allowances. The $2 figure is a cost input, not a rental price.

Ground IT therefore accounts for about $8,734 of the $11,434 annual bill. Only about $2,701 is power and facility. Describing the whole terrestrial stack as exposed to electricity scarcity would inflate the opportunity rather neatly.

This is why the composition of inflation matters. A higher electricity tariff expands orbit’s opportunity. A higher accelerator bill raises both budgets. Pooling them into a single escalating ground-cost number makes those effects look interchangeable, even though they push the comparison in different directions.

Give the rocket a price of $200 per kilogram and the orbital stack reaches about $15,725 per IT-kilowatt-year. In the units most compute customers recognize, that is $3.42 per available GPU-hour against $2.49 on the ground, at full utilization and the same GPU class.

The strongest fair version of the bull case changes the satellite, not just the rocket: IT and spares at parity with the ground, non-IT hardware at $1,000 per generated kilowatt instead of $5,000, operations halved. That joint recipe lands at $12,125 at a $200 launch price, or $2.64 per GPU-hour. Still above the ground’s $2.49.

The comparison also cuts against an easy scarcity story. More expensive chips hit both venues. With the larger orbital provisioning allowance, common chip inflation hurts orbit more. Launching a processor cheaply doesn’t reduce what its replacement costs.

Give the rocket a price.

The exhibit beside this block is live. Change the launch price, the firm power tariff or the orbital cost stack and watch the two columns move. The ceiling P* is the most orbit could pay for launch and still match the ground.

How cheap does the ride have to get?

In this base case, no nonnegative launch price works. We can make that useful by solving for the most orbit can afford while matching ground. Call it P*, the break-even launch price.

Derivation kept in one place: Break-even launch price and Capital recovery factor, in Mechanisms.

K recovers capital over life L at financing rate r. Ground electricity is e; F is facility capital. For orbit, h is non-IT hardware and m is mass per generated kilowatt; η converts generated power to IT power. O covers operations. The survival provision b adjusts non-IT capacity; a and g cover separate IT provisioning. Each C is dollars per useful IT-kilowatt-year, while P is dollars per kilogram and e is dollars per megawatt-hour.S276

At the base, A is about 9.027 kg per IT-kilowatt-year, and P* comes to about −$275 per kilogram. A negative number means there isn’t room for a launch bill. It also tells us where to look next.

So what would it actually take?

Removing the modeled IT-charge and spares disadvantage lifts P* by $254/kg, to about −$21.

Cutting non-IT hardware from $5,000 to $1,000 per generated kilowatt independently adds $195/kg. Halving operations adds about $38/kg. Doubling ground power and facility costs adds $299/kg. Those are separate levers against the same base.

Combine IT parity with the hardware reduction and P* reaches about +$174/kg. Launch matters again, but $200 is still too expensive for that combination.

The payoff is being able to name the saving that changes the answer. A sufficiently expensive ground alternative could compensate for slower hardware progress. Cheaper finance could help too. The conditional bull case concentrates on industrial cost reductions against persistent power scarcity; the arithmetic doesn’t decree a single route to parity.

How much manufacturing does “industrial scale” actually mean?

The $5,000 hardware input is assumed, and its proposed reduction is huge. At an assumed 20% hardware-cost reduction per doubling, an 80% decline takes about 7.21 doublings, or 148 times cumulative production.

That’s a factory program, not a catchy launch milestone.

Count compute-satellite production, rather than every flight carrying something else. Think of a factory learning to make the same difficult assembly more cheaply. A different rocket payload doesn’t teach that factory much. A redesign might bring a sharp improvement, but it also changes the product being measured.

Solar offers encouragement with limits. IRENA’s 2010–2024 endpoints show installed PV cost down about 87%, and its electricity cost down about 90%. They show that large industrial declines happen. They don’t establish the same learning rate for a radiation-qualified compute satellite.

The base capital-recovery factor is 0.298 in orbit versus 0.134 for the ground facility. Orbit uses 15% finance over five years; ground uses 10.3% over fifteen. That’s a consequential financing and lifetime asymmetry.

At 12% over seven years, orbit’s factor falls to 0.219. In a separate assumed capital treatment, retaining five-year IT refresh gives P* = −$176; granting seven-year useful IT life lifts it to +$122. At 10% with seven-year bus and IT life, the threshold reaches +$222.

A processor can survive physically while becoming economically tired. Those positive results need useful accelerator life and financing to cooperate. IT parity and cheaper IT financing overlap; adding their savings together would count part of the improvement twice.

The energy advantage survives all this accounting. The question is how much complete infrastructure it saves, and whether Earth can provide a workable alternative.

Launch has already fallen a long way. In 2018 dollars, the Shuttle era sat near $54,500 per kilogram, the expendable average near $18,500, Falcon 9 near $2,720 and Falcon Heavy near $1,410. That history is why the idea refuses to die.S201

But the current posted reference is $3,364/kg for a fully expendable Falcon 9, and $7,000/kg for sun-synchronous rideshare. Model inputs used elsewhere, from $1,400 to $3,600, are assumptions rather than quotes.

Behind-the-meter power is a better scarcity test than attaching an invented dollar premium to a queue. The sourced model range is $110–170/MWh all-in. At $120 and $150, the unchanged orbital stack still has negative break-even launch prices, about −$232 and −$193/kg. The joint case reaches +$168 and +$206. Expensive ground power helps once the satellite bill comes down.

So the ceiling the joint recipe can afford sits between $128/kg at $87 power and $206/kg at $150 power. Against today’s posted reference, that is a gap of sixteen to twenty-six times, before anyone has built the cheaper satellite.

Texas’s October FAA notice described 25 currently authorized Starbase launches annually. KSC LC-39A’s 44-launch environmental envelope and SLC-37’s conditional 76-launch case sit at different legal stages. Adding them doesn’t create 145 currently licensed flights.

Ground has meaningful hurdles of its own. TCEQ’s October 19 letter confirms a permit pause for infrastructure directly supporting data centers. New York’s July 14 order puts certain DEC permits on hold for data centers that can consume at least 50 MW. Neither statement amounts to a nationwide construction ban.

How many rockets make a gigawatt?

Roughly 93–202 full payload flights for a provisioned IT gigawatt, using the model’s 13.94–30.26 kg per IT kilowatt and an assumed 150-tonne payload. The survival provision is already included once. Mass remains a satellite proxy pending a measured compute design.

Under the assumed ramp, 1,000 flights annually could deliver about 4.96–10.76 provisioned IT GW each year if every payload were compute. Real allocation competes with other missions; retirement and replacement consume capacity too.

A useful capacity comparison asks how much compute could be delivered after allocation, utilization and retirement. A business still needs a customer willing to pay enough for that service. Keeping those questions separate prevents an impressive payload manifest from quietly turning into an order book.

Is the model smuggling in its own answer?

It could be. The negative base rests heavily on the $2 versus $1.81 annual IT charge and the provisioning gap. Together they explain 92.35% of the pre-launch dollar deficit. Anyone challenging those inputs has identified a load-bearing part of the argument.

That cuts both ways. Favorable hardware assumptions shouldn’t receive a free pass, and unfavorable replacement assumptions shouldn’t acquire authority merely because they produce a cautious conclusion. Compare useful work at matched availability. A ground interruption that a technician fixes and a permanently lost orbital device are different events.

Google’s radiation test gives that challenge substance. A tested TPU showed no hard total-dose failures through 15 krad against a stated mission dose around 750 rad. Memory stress irregularities appeared near 2 krad. That supports investigating a smaller qualification penalty; it doesn’t measure fleet spares, financing or useful economic life.

History supplies a useful warning about customers. Iridium began commercial voice service in November 1998 and filed bankruptcy petitions in August 1999. During the fiber overbuild, excess long-haul capacity coexisted with last-mile shortages.

Building the trunk didn’t connect the customer.

Ground can improve too. Efficient chips, better power procurement and faster commissioning could shrink the scarcity premium before orbit matures.

What would change my mind?

A reconciled cost stack and delivered service would do it. Here are dated evidence gates, rather than permission to roll the deadline forward whenever another exciting announcement arrives.

By the end of 2027, check Louisiana’s construction, financing and project-specific review record. Require actual upper-stage reflight; catching a vehicle alone doesn’t establish reuse. Seek a delivered launch offer with payload, orbit and recovery terms, alongside a reproducible satellite bill of materials and fully annualized IT replacement costs. Give ground its test too: compare projects that can actually deliver power, with their complete costs and usable-service dates.

By the end of 2029, seek a full year of thermal telemetry and separately reported failures. Keep the proposed minimum of 100 GPU-years and one flight-year, while reporting silent corruption and permanent loss separately. A small sample can expose a bad system more easily than certify a great one.

By the end of 2030, require paid launch invoices, compute-only manufacturing mass and observed hardware costs. Assess workloads on useful output and availability. A favorable launch curve cannot substitute for measured satellite learning.

A failed test should change the scenario, rather than prompt a search for a more flattering comparison. Publish the missing cost or disappointing operating result alongside the successful launch. Otherwise the exercise becomes a moving finish line, with the model jogging politely behind it.