SpaceX’s $26.5 Trillion AI Market: A Leprechaun’s Pot of Gold?

Key Takeaways

• SpaceX’s May 2026 registration statement values its quantifiable total addressable market at $28.5 trillion, of which $26.5 trillion is attributed to artificial intelligence, and the filing connects that opportunity to a plan for orbital data centers that could eventually number in the millions of satellites.
• A headline market size measures the whole AI economy, not the slice a single orbital operator could serve and then win, and the gap between the total market, the serviceable market, and the obtainable market is where most of that $26.5 trillion sits.
• Whether space-based compute captures a meaningful share depends on workload suitability, a growing field of competitors, full cost parity with ground data centers, continued chip and model efficiency gains, data sovereignty pressures, and the social license to operate both on Earth and in orbit.

What SpaceX Told the SEC About Its AI Market

On May 20, 2026, Space Exploration Technologies Corp. filed a Form S-1 registration statement with the United States Securities and Exchange Commission (SEC), the document that precedes an initial public offering (IPO), applying to list its Class A common stock on the Nasdaq Stock Market under the symbol SPCX. Inside that prospectus, the company describes what it calls the largest actionable total addressable market in human history.

The number it puts forward is $28.5 trillion. The breakdown assigns $370 billion to Space, roughly $1.6 trillion to Connectivity, and $26.5 trillion to AI. The AI portion is divided again into $2.4 trillion for AI infrastructure, $760 billion for consumer subscriptions, $600 billion for digital advertising, and $22.7 trillion for enterprise applications. The estimate excludes China and Russia from its global figures, which means the headline number already rests on a partial map of world demand.

The filing then links that opportunity to a specific architecture. SpaceX states that its reusable rockets and satellite manufacturing scale can enable the deployment of massive AI compute satellite constellations, potentially numbering in the millions, with deployment of orbital AI compute satellites beginning as early as 2028. These satellites would sit in Sun-synchronous orbit, draw on solar arrays the company says can generate more than five times the energy per unit area of terrestrial solar, and reject heat through radiative cooling. The long-term goal stated in the prospectus is to deploy 100 gigawatts of annual compute power to orbit, which the company estimates would require thousands of launches per year and the transport of approximately one million metric tons to orbit annually.

The table below reproduces the segment breakdown as stated in the prospectus.

The prospectus itself flags the uncertainty. SpaceX warns readers that estimates of future market opportunity may prove to be inaccurate and that investors should not place undue reliance on statements of expected future market size. It goes further on the orbital plan specifically, conceding that the timeline for deploying 100 gigawatts of annual compute to orbit and the launch cadence required to reach it may be difficult or impossible to determine. That candor is the right starting point for any careful reading of the $26.5 trillion figure.

The Difference Between a Market and a Business

The central question is not whether AI will become a very large market. It probably will. The harder question is whether a broad AI market can be converted into an addressable market for SpaceX, and then narrowed again into a market that orbital data center satellites can actually serve.

A total addressable market is not the same thing as revenue, profit, capital availability, launch feasibility, regulatory approval, customer willingness to buy, or public acceptance. The $26.5 trillion figure folds in enormous downstream categories: enterprise AI applications at $22.7 trillion, plus advertising, subscriptions, and infrastructure. Those categories describe the wider AI economy. Most of that value would not flow automatically to a satellite operator. A company can participate in the infrastructure layer without capturing the value of the applications that run on top of it, in the same way that a company supplying electricity to a film studio does not earn the box office receipts.

Product and market analysts separate that universe into three layers, summarized below.

A $26.5 trillion total addressable market (TAM) may describe the broad AI economy. The serviceable available market (SAM) for orbital AI infrastructure would be far smaller, because it is limited to the workloads that space can actually host. The serviceable obtainable market (SOM) for one company operating orbital data center satellites would be smaller still, once competition, regulation, technical limits, customer preferences, and financing constraints are priced in. The headline number and the business case live in two very different places, and the distance between them is the heart of the matter.

Why AI Is Not a Single Workload

Treating “AI” as one block hides the most important detail. Different AI workloads have very different computing, power, latency, networking, and data-location requirements, and only some of them are candidates for orbit.

Large frontier model training sits among the most power-hungry and capital-intensive categories. It needs enormous clusters of processors, high-speed networking, large energy supplies, specialized cooling, and access to massive datasets. High-volume inference, the process by which a trained model produces outputs from new input data, can also become a major power consumer at global scale, especially for complex models serving millions or billions of requests. These are the parts of AI most associated with the surge in data center electricity demand that the International Energy Agency (IEA) projects will more than double to around 945 terawatt-hours by 2030, with consumption from AI-focused facilities growing much faster still.

Many other workloads are far less demanding. Small-model inference, edge AI running on devices, traditional machine learning, data labeling, model evaluation, business analytics, retrieval systems, software orchestration, and routine workflow automation do not require giant new data center campuses. Some of that work runs comfortably on local devices, enterprise servers, regional cloud facilities, or specialized accelerators. Some can be handled by smaller models rather than the largest frontier systems. Each of those shifts narrows the slice of the AI market that could justify orbital infrastructure.

The sharper version of the question is therefore not “how large is AI?” It is: which AI workloads need enormous amounts of centralized compute, are not sensitive to latency, can tolerate space-based networking constraints, are not blocked by data sovereignty rules, and are valuable enough to pay a premium for orbital delivery? That set is much smaller than the broad AI economy. Ground data centers are likely to keep an advantage for work that depends on low latency, physical access to hardware, proximity to users, compliance with national data laws, fast hardware refresh cycles, and tight integration with existing cloud systems.

The Workloads That Might Actually Suit Orbit

A narrower set of categories does look more plausible for space. Delay-tolerant batch processing, certain forms of model training where datasets can be staged efficiently, synthetic data generation, and some space-based sensing and data processing tasks could benefit from orbital solar power and location. Early commercial activity points in that direction. The startup Starcloud launched a satellite carrying an Nvidia H100 graphics processing unit (GPU) in November 2025 and ran a version of Google’s Gemini model in orbit. Google has announced Project Suncatcher, a plan to fly two prototype satellites by early 2027 and study clusters built around its tensor processing units, while Axiom Space and others pursue orbital data center nodes aimed at defense and commercial customers.

Those efforts validate the concept at small scale. They do not, on their own, add up to a trillion-dollar market, let alone a market large enough to support a constellation numbering in the millions. The serviceable market has to be built upward from realistic workload categories, not downward from a headline figure for the entire AI economy. A useful test for any candidate workload is whether the value it produces can survive the round trip: data lifted to orbit, processed under the constraints of a spacecraft, and returned to a customer on the ground, all at a price that beats a terrestrial alternative. Few workloads pass that test today, and the ones that do tend to be the patient, bandwidth-light, compute-heavy jobs rather than the interactive services that dominate current AI usage.

The Competitive Field Forming in Orbit

A serviceable obtainable market assumes competition, and the field of companies pursuing orbital compute is already crowded. That matters directly for the $26.5 trillion claim, because even a generous estimate of the orbital opportunity would be divided among several well-funded players rather than handed to a single operator.

Starcloud, formerly known as Lumen Orbit, has moved fastest in public demonstrations. Its Starcloud-1 satellite reached orbit in November 2025 with an Nvidia H100, became the first to train a model in space, and ran a version of Google’s Gemini, with a follow-on satellite carrying a full GPU cluster planned next and a long-term vision of a multi-gigawatt constellation. Google’s Project Suncatcher takes a different route, pairing its own tensor processing units with a planned 81-satellite cluster roughly one kilometer across, connected by free-space optical links, after reporting that its Trillium-generation chips survived radiation testing meant to mimic low-Earth orbit. Axiom Space has taken yet another path, launching orbital data center nodes on Kepler Communications satellites in late 2025 with a focus on defense and real-time, secure processing.

Smaller and more specialized entrants round out the picture. Cowboy Space Corporation, formerly Aetherflux, founded by former Robinhood co-founder Baiju Bhatt, has targeted an orbital data center satellite for early 2027, and Lonestar Data Holdings is pursuing data storage on the lunar surface. The same reporting notes that the chief executive of OpenAI has explored a partnership with or acquisition of a rocket maker, a sign that demand for launch capacity from AI firms is not unique to SpaceX. A European feasibility effort known as ASCEND has studied whether orbital data centers could support the continent’s climate and sovereignty goals, and Blue Origin has signaled interest in the category as well.

The table below summarizes the main efforts.

SpaceX holds real advantages over this field, chiefly its launch cadence, its satellite manufacturing base, and the Starlink network it could use to move data to and from orbit. Those advantages strengthen its position within whatever serviceable market exists. They do not enlarge that market, and they do not change the fact that the obtainable share is split among rivals, several of whom are backed by the same chipmakers and cloud providers that would otherwise be customers.

The Cost Parity Problem

The financial case has to clear a demanding bar: the cost of orbital infrastructure would need to be on par with, or better than, terrestrial infrastructure for the relevant workloads. That does not mean matching only the cost of electricity, or only the cost of launching hardware. It means matching the full delivered cost of computation, including capital spending, operating costs, refresh cycles, reliability, financing, insurance, data transport, maintenance, customer acquisition, and risk.

Ground data centers are expensive, but they benefit from mature supply chains, ground access, utility-scale power contracts, equipment replacement options, service crews, competitive fiber networks, established financing models, and decades of operational learning. They are also getting faster and cheaper to build. SpaceX itself reports bringing its first COLOSSUS compute cluster online in 122 days and the first COLOSSUS II cluster in 91 days, against an industry benchmark of roughly two years for a 100-megawatt greenfield facility, at construction costs it describes as well below industry norms. Every gain on the ground raises the bar that orbit has to clear.

For parity to become plausible, several parameters would all have to move at once, as set out below.

Launch cost is the parameter SpaceX has done the most to move. According to figures the company cites from NASA, the first version of Falcon 9 cut launch cost to about $2,700 per kilogram, roughly 85 percent below the historical average of $18,500 per kilogram, and the company states that fully reusable Starship is designed to reduce the cost to reach orbit by 99 percent or more relative to that historical average. Even a reduction of that size has to hold at extreme cadence. Deploying 100 gigawatts of compute per year, with each satellite carrying more than 100 kilowatts of compute per metric ton, would require thousands of launches per year and about one million metric tons to orbit annually. That is a flight rate without precedent, and per-kilogram economics that look attractive in a brochure can erode quickly once range scheduling, refurbishment, insurance, and failure rates are layered onto a campaign of that scale.

Manufacturing is the next constraint. SpaceX argues it can produce AI compute satellites at the scale of automotive manufacturing, which would be a first for the industry. Reaching that scale at high yield, with space-grade reliability and the latest processors, is a manufacturing problem as hard as the launch problem, and the cost only works if the production line never stalls.

The third constraint is the most subtle: the useful life of each satellite. In a ground data center, operators swap GPUs, servers, memory, cooling equipment, and networking gear as technology advances, often on a two-to-three-year cycle. A satellite cannot be upgraded the same way. Its lifetime has to be long enough to amortize the capital tied up in the spacecraft, yet the computing hardware inside it risks becoming economically obsolete long before the bus around it wears out. If the satellites are treated as disposable, replacement becomes a recurring launch and manufacturing burden. If they are designed to be serviced, then robotic servicing vehicles, docking interfaces, and spare-parts logistics become additional infrastructure that must be built, launched, and paid for. Either path has to be priced into the real cost of orbital computation, and neither has a terrestrial equivalent that is anywhere near as expensive.

A breakthrough in one of these parameters is not enough if the rest of the stack stays high. Cost parity would require improvement across launch, manufacturing, satellite lifetime, computing efficiency, power mass, radiator mass, network throughput, ground infrastructure, failure rates, replacement cadence, insurance, and regulatory certainty together. The orbital stack starts well above terrestrial economics on most of those lines, and the company’s own numbers show how far there is to go: its AI segment recorded a loss from operations of about $6.4 billion in 2025 and consumed $7.7 billion of capital spending in the first quarter of 2026 alone, all of it on the comparatively easier terrestrial side of the business.

Power, Heat, and Data Transport in Orbit

The power equation is more favorable in space than on the ground, but it is not free. Space-based systems can access solar energy, and dawn-dusk Sun-synchronous orbits can keep panels in near-continuous light that delivers up to eight times the energy of ground installations, a figure broadly consistent with the more than five times that SpaceX cites. The catch is mass. Each satellite has to carry solar arrays, power electronics, storage, thermal systems, pointing systems, and the structure needed to hold all of it together and aim it at the Sun. Every kilogram tied to power generation has to be manufactured, launched, operated, protected, and eventually disposed of, so abundant sunlight does not translate into free power once the full hardware chain is counted.

Thermal management may become the hardest financial constraint of all, and it is the one that scales worst with ambition. Computation produces heat, and in the vacuum of space heat can only be shed by radiating it away. There is no air to carry it off and no water loop to a cooling tower. Rejecting the heat from high-density compute therefore requires large radiator surfaces, added mass, and careful spacecraft design, which is why Google’s own Project Suncatcher research highlights the need for advanced, preferably passive heat transport to move large heat loads to dedicated radiator surfaces. The more compute packed into a satellite, the more radiator it needs, and radiators are mass that competes with the very payload they exist to serve.

Data transport is the third layer of the cost stack. An orbital data center is useless if it cannot move data in and out at commercially attractive rates. That means high-capacity links between satellites, to ground gateways, and possibly to customer networks, built from optical terminals, phased arrays, spectrum coordination, ground stations, routing systems, cybersecurity, redundancy, and latency management. SpaceX’s strongest card here is Starlink, which it says would let data from compute satellites reach ground stations anywhere on Earth. That is a real advantage, but it is also a real cost, because the relay capacity consumed by moving compute traffic is capacity not sold to broadband and mobile customers.

Maintenance, insurance, and financing complete the picture. SpaceX plans to test compute hardware extensively before launch and to use Starlink fleet management software to reallocate traffic around failed hardware rather than repair it in place. That mitigates downtime but does not eliminate the cost of failures, since a dead processor in orbit is a stranded asset. Beyond the engineering, orbital systems add launch failure, satellite failure, collision risk, debris risk, reentry liability, solar storm exposure, cybersecurity exposure, and geopolitical risk to the list of things investors and insurers must price. Even if every piece of hardware works as designed, the financial model still has to support a competitive cost of capital against ground data centers that carry none of those orbital risks.

Efficiency Gains That Could Shrink the Target

Future data center demand is not fixed, and several technology trends could lower it independently of any public opposition. The IEA notes that power consumption per AI task is declining rapidly, with efficiency improving at a rate unprecedented in energy history, even as total consumption rises because more people use AI and more demanding applications such as agents spread.

Hardware is the first lever. The industry is moving well beyond general-purpose GPUs toward specialized accelerators, tensor and neural processing units, application-specific chips, wafer-scale processors, advanced packaging, chiplet designs, high-bandwidth memory, optical interconnects, and liquid and immersion cooling. Smaller transistor geometries, better lithography, and 3D stacking continue to improve performance per watt. Each of these changes the amount of electricity and physical infrastructure needed to deliver a given amount of AI output, and they do so on a faster cycle than any satellite can be replaced.

Model architecture is the second lever, and it has moved quickly. Mixture-of-experts designs activate only a fraction of a model for each query, as with DeepSeek R1, which uses roughly 37 billion of its 671 billion parameters per token. Quantization, pruning, distillation, caching, retrieval, and routing tasks to appropriately sized models all reduce wasted computation, and more processing is moving onto phones, laptops, vehicles, and enterprise edge systems. The effect on price has been steep. One analysis of pricing across vendors finds that, controlling for benchmark performance, token prices have been falling by a factor of ten or more per year, and the chief executive of OpenAI has described DeepSeek’s model as running 20 to 50 times cheaper than a comparable system.

These trends do not guarantee that total demand will fall. Cheaper computation tends to stimulate more usage, a pattern often called the rebound effect, so efficiency and growth can rise together. What the trends do is complicate any straight-line forecast that assumes AI power demand will keep climbing without large efficiency improvements, and they weaken the specific argument that only orbit can supply the energy the future will need. If the cost of inference keeps dropping and training keeps getting more efficient, the premium that customers would pay for orbital delivery shrinks, and the business case for an extreme orbital buildout grows harder to defend.

Quantum computing is sometimes raised in this context, but it deserves careful handling. It is better understood as a future specialized capability for optimization, simulation, cryptography, and materials science than as a near-term replacement for the conventional compute used to train and serve AI models. For the next decade it is unlikely to displace the large clusters that orbital data centers would target. Its relevance here is mostly as one more reminder that long-term compute demand is shaped by technology shifts rather than by today’s scaling assumptions.

Demand That May Never Be Serviceable

Some demand will exist but remain out of reach, which further separates the available market from the obtainable one. Many governments may not want strategic AI infrastructure, national datasets, defense applications, or sensitive enterprise workloads dependent on a United States company or a United States-controlled orbital system. The pull toward domestic capability is already visible. An Accenture study found that European organizations are placing greater emphasis on controlling their own data and infrastructure, accelerating demand for sovereign AI, while McKinsey research reports that 44 percent of European technology leaders cited data security concerns as a reason not to use public cloud. France alone has announced around 109 billion euros in AI infrastructure investment, and the European Union’s AI Act adds a layer of lifecycle governance that favors local hosting and control. Europe, China, India, the Gulf states, and others may all prefer domestic or allied infrastructure for reasons tied to data sovereignty, industrial policy, and national security. That fragments the global market and reduces the portion available to any single operator, a point underlined by the prospectus already excluding China and Russia from its own figures.

A second source of unreachable demand is self-supply. The largest AI customers are precisely the firms most able to build their own infrastructure rather than rent capacity from an orbital operator, and several of them, including the companies behind Starcloud and Project Suncatcher, are building competing space hardware of their own. Capacity that a customer builds for itself is capacity SpaceX cannot sell.

Business-cycle risk compounds both problems. The AI buildout is consuming extraordinary capital, with five large technology companies spending more than $400 billion in 2025 and that figure set to rise by a further 75 percent in 2026. If part of that surge proves to be a bubble, a correction could leave the market with excess ground capacity, distressed assets, cancelled expansion plans, and weaker pricing power, making an expensive orbital architecture harder to justify at exactly the moment financing becomes scarce.

The Social License Problem Moves to Orbit

There is a paradox in using opposition to ground data centers as an argument for space. Resistance on Earth is real and growing. Data Center Watch, a research project run by the firm 10a Labs, estimates that around $156 billion in data center projects were blocked or delayed by local opposition in 2025, with 142 activist groups operating across more than 20 states and the backlash crossing party lines. More than 230 state and local environmental groups signed a letter to Congress in December 2025 demanding a national moratorium on new construction. Public sentiment reinforces the picture: a December 2025 YouGov survey found that most Americans use AI but still do not trust it, with skepticism highest in finance and healthcare. The opposition appears among students, faculty, local communities, environmental groups, and the general public, which suggests it is not confined to one region or one political camp.

Moving the infrastructure off Earth does not make the objections disappear. It relocates them. A constellation numbering in the millions raises questions about orbital debris, collision risk, astronomy interference, reentry disposal, spectrum coordination, launch emissions, national security, and international governance. Astronomers have already documented the strain. A study published in Nature found that 4.3 percent of Hubble Space Telescope images taken between 2018 and 2021 already carried satellite trails, and modeling suggests that if the satellite population grows to the roughly 56,000 projected by the end of the decade, it could contaminate close to 40 percent of Hubble’s images and the large majority of images from several other observatories. A separate astronomy committee notes that companies have announced plans for up to 400,000 satellites by 2030 and that more than one million debris fragments larger than one centimeter may already orbit Earth, warning of collision cascades, sometimes called Kessler syndrome, that could render whole orbital regions unusable.

SpaceX’s own operations show how quickly this scales. The company reports operating about 9,600 Starlink satellites, which it says account for roughly 75 percent of all active maneuverable satellites and perform more than 1,000 automated collision-avoidance maneuvers per day. Adding millions of larger, power-hungry compute satellites to that environment compounds every concern that megaconstellations already raise. The opposition that now appears at county zoning boards and campus protests could simply move to space regulators, astronomers, defense agencies, insurers, and foreign governments, none of whom SpaceX controls.

For the next decade, public resistance looks more like a serious headwind than a single decisive obstacle. The likelier showstopper is the combined weight of economics, launch cadence, manufacturing rate, orbital safety, energy and thermal management, data relay architecture, regulatory approval, customer adoption, competition, workload suitability, computing efficiency, semiconductor progress, and uncertainty over future demand. A negative shift in opinion would make every one of those harder to overcome, reducing political tolerance for risk, making regulators more cautious, and making investors more skeptical of grand market projections.

Summary

SpaceX’s $28.5 trillion total addressable market, and the $26.5 trillion AI component inside it, may be defensible as a broad statement about the future AI economy. It is a much weaker guide to what a satellite operator can capture through space-based data centers. The economically serviceable market for orbital AI infrastructure is smaller, slower to emerge, and far more constrained than the headline suggests, because it depends on a narrow set of suitable workloads, a field of capable competitors already forming in orbit, full cost parity across an entire orbital stack, continued demand that efficiency gains and sovereignty rules do not erode, and a social license that holds on Earth and in orbit alike. None of that makes orbital data centers impossible, and SpaceX brings advantages in launch, manufacturing, and connectivity to the contest. It does mean the TAM figure answers the easy question about the size of the AI economy while leaving the harder questions, about what is serviceable, what is obtainable, who else will win part of it, and who will permit, finance, insure, and pay for the physical infrastructure, largely open.

Appendix: Useful Books Available on Amazon

• Liftoff by Eric Berger traces the early history of SpaceX and the engineering culture behind its launch business.
• Reentry by Eric Berger continues the story through the Falcon 9 and reusable rocket era.
• Elon Musk by Walter Isaacson offers a detailed biography of the company’s founder and his approach to risk.
• The Space Economy by Chad Anderson examines how investors evaluate commercial space markets.
• Chip War by Chris Miller explains the semiconductor supply chain that underpins AI compute.
• The Coming Wave by Mustafa Suleyman explores the scale and governance challenges of advanced AI.
• Crossing the Chasm by Geoffrey Moore is a standard reference on how new markets are sized and adopted.

Appendix: Top Questions Answered in This Article

What exactly is the $26.5 trillion figure SpaceX uses?

It is the AI portion of a $28.5 trillion total addressable market that SpaceX stated in its May 2026 Form S-1 filing with the SEC. The AI portion combines $2.4 trillion of infrastructure, $760 billion of consumer subscriptions, $600 billion of advertising, and $22.7 trillion of enterprise applications, and the estimate leaves out China and Russia.

Does a large total addressable market mean SpaceX will earn that much?

No. A total addressable market measures the entire universe of possible demand. It is not revenue, profit, or the share a single company can serve and win. Most of the $26.5 trillion describes applications and services that would not flow to a satellite operator even if the broader market reaches that size.

What do TAM, SAM, and SOM mean?

TAM is the total addressable market, the full universe of demand. SAM is the serviceable available market, the part a specific product can realistically serve. SOM is the serviceable obtainable market, the part one company can actually win after competition, pricing, regulation, and financing are taken into account.

Which AI workloads could actually run in orbit?

The better candidates are delay-tolerant batch processing, some training where data can be staged in advance, synthetic data generation, and certain space-based sensing tasks. Latency-sensitive work, workloads bound by national data laws, and tasks needing frequent hardware access are better suited to ground facilities.

Who else is building orbital data centers?

Starcloud launched a GPU satellite in November 2025, Google is pursuing Project Suncatcher with tensor processing units, Axiom Space has deployed orbital data center nodes, and Aetherflux and Lonestar are pursuing their own orbital and lunar projects. A European study called ASCEND and interest from Blue Origin and OpenAI add to the field, which means any orbital opportunity would be shared.

When does SpaceX plan to deploy orbital AI compute?

The filing states that deployment of orbital AI compute satellites could begin as early as 2028, with a long-term goal of placing 100 gigawatts of annual compute power in orbit. SpaceX acknowledges that the timeline and the launch cadence required may be difficult or impossible to determine.

Is solar power in space effectively free?

No. Sunlight is abundant in the right orbit, but each satellite must carry solar arrays, storage, power electronics, pointing systems, and the structure to support them. Every kilogram has to be built, launched, operated, and disposed of, so the delivered cost of power is real.

Why is cooling such a difficult issue in space?

Computation generates heat, and in a vacuum that heat can only be shed by radiating it away, with no air or water to carry it off. That requires large radiator surfaces and added mass, which is one of the harder cost and engineering constraints for high-density compute in orbit.

Could efficiency gains shrink the market for orbital data centers?

Yes. Power consumption and inference cost per AI task are falling quickly, token prices have dropped by a factor of ten or more at comparable performance, and architectures such as mixture-of-experts activate only part of a model per query. If inference and training keep getting cheaper, the premium customers would pay for orbital AI compute capacity shrinks.

How does data sovereignty affect the opportunity?

Many governments and enterprises want AI infrastructure under domestic control. Demand for sovereign AI is rising in Europe and elsewhere, which fragments the global market and reduces the share that a United States-controlled orbital system could realistically serve.

What happens to cost parity if launch costs fall sharply?

Lower launch costs help, but they are only one line in the cost stack. Parity also depends on satellite manufacturing cost, useful life, power and radiator mass, data transport, maintenance, insurance, and financing. A breakthrough in launch alone does not deliver parity if the rest of the stack stays expensive.

Would moving data centers to space end the public backlash?

Not necessarily. It relocates the objections. A constellation numbering in the millions raises concerns about debris, collision risk, astronomy interference, reentry, and spectrum, shifting opposition from local zoning boards to space regulators, astronomers, insurers, and foreign governments.

Appendix: Glossary of Key Terms

Total Addressable Market (TAM): The full universe of demand for a product or service if every possible customer bought it.

Serviceable Available Market (SAM): The portion of a total market that a company’s actual product or service could realistically serve.

Serviceable Obtainable Market (SOM): The portion of the available market that a single company could plausibly capture after competition, pricing, and other constraints.

Form S-1: The registration statement a company files with the SEC before an initial public offering, describing its business, financials, and risks.

Initial Public Offering (IPO): The first sale of a private company’s shares to public investors.

Low-Earth Orbit (LEO): A region of space relatively close to Earth’s surface where most communications satellites operate.

Sun-synchronous Orbit: A polar orbit in which a satellite passes over each point at the same local solar time, allowing consistent solar exposure, with dawn-dusk variants offering near-continuous sunlight.

Training: The process of building an AI model by adjusting its internal values using large datasets.

Inference: The process by which a trained AI model produces outputs from new input data.

Mixture-of-Experts: A model architecture that activates only a fraction of its parameters for each query, reducing the compute used per response.

Free-Space Optical Link: A high-capacity laser connection used to transfer data between satellites without cables.

Radiative Cooling: A method of shedding heat by radiating it into space, used because there is no air in orbit to carry heat away.

Rebound Effect: The tendency for efficiency gains to lower cost and stimulate more usage, offsetting part of the expected reduction in total demand.

Kessler Syndrome: A scenario in which collisions between orbiting objects create debris that triggers further collisions, potentially making some orbits unusable.

Sovereign AI: The capability of a country to develop and run AI using local infrastructure, data, and models to reduce dependence on foreign providers.