Space Economy 1.0

Commercialisation of space began in the early '60s, starting with the Telstar 1 satellite which delivered the first television broadcast to Europe in 1962 and Syncom 3 which beamed the 1964 Tokyo Olympics to the US. HBO leased a transponder on Westar 1 to broadcast the "Thrilla in Manila" Muhammad Ali vs. Joe Frazier fight in 1975, but this was still a B2B endeavour, allowing them to send signals to cable TV operators who received them via 3 metre satellite dishes and distributed it on to their customers. Mass adoption of satellite TV to the home was gated by C-band frequencies requiring such large dishes.

By 1994, DirecTV was offering digital signals on higher power radio frequencies (ku-band), which enabled 175 channels delivered to a dish 45cm wide. This increased the space economy by >10x to ~$30 billion in revenues by the end of that decade, with commercial revenues surpassing government space-based spending for the first time in 1997. Changes to regulations opened up GPS signals to the commercial market in the '00s, and satellite internet came online at around the same time. These markets kicked off another decade of rapid growth, and increased space revenues another ~5x to $160 billion (this was partly due to the Satellite Industry Association expanding their definition to include consumer equipment, like GPS chips and receiver dishes). Satellites were few in number - at geostationary Earth orbit (GEO) of 35,786km it only takes three satellites to cover the entire globe. 

The journey from a couple of hundred million in revenue in the early '70s to ~$160 billion by the end of the noughties represents a CAGR of around +20% - good, but not stellar, given the infinite possibilities of space.

Agile Aerospace

From the '50s through the '00s, orbital launches were largely a sovereign entreprise, with national industrial champions heavily subsidised by the state, or commissioned through open-ended 'cost plus' contracts, which rarely came in on time or on budget. The partially re-usable Space Shuttle program ran from 1981-2011 and cost NASA $209 billion, which was spread amongst various aerospace companies which today form parts of Boeing, Lockheed Martin and Northrup Grumman. 

SpaceX was founded in 2002, and achieved the first successful launch of a privately-funded orbital rocket in 2008, with just $182 million in private equity and a milestone-based NASA contract of $278 million, which started in 2006. Since then, SpaceX has achieved the first successful landing of a first stage booster (2015), the first re-flight of a booster and a previously flown Dragon capsule (2017), and the first successful launch and recovery of a fully reusable rocket system in 2024 with their 5th Starship test flight. By the end of 2025, SpaceX was responsible for ~80% of mass-to-orbit globally, which has been achieved with $12 billion in cumulative private equity financing and $22 billion in government grants/contracts - 6x less than the cost of the Space Shuttle program.

The Soviet Union dominated orbital launches in the '70s and '80s, but the US has regained leadership, largely thanks to SpaceX

Mass-to-orbit cost fell rapidly between Vanguard and Saturn V, but stagnated over the lifetime of the Space Shuttle program. SpaceX's partially re-usable rockets re-instated the trend lower

SpaceX is now responsible for 81% of mass-to-orbit launched globally

The success of SpaceX accelerated the expansion of the space economy in a number of ways. Of course more frequent launches and cheaper mass-to-orbit makes it easier to launch satellites but, perhaps more impactfully, SpaceX showed founders and VCs in Silicon Valley that rocketry and space-faring are not solely under the purview of state-level programs. That isn't to say there are no barriers to entry - progress to date has been punctuated with failures, from well-funded, high profile ventures like Virgin Orbit followed by a long tail of less well-known companies. Bezos-backed Blue Origin is two years older than SpaceX, and just recently failed their first commercial satellite launch (though they did succeed in landing their booster on a ship platform at sea), and followed this with another launch soon after which exploded on the launch pad, setting them back at least 12 months. Rocket Lab, founded four years after SpaceX, has become the number two orbital launch start-up, after achieving their first commercial launch in 2018, focusing on smaller, cheaper launches with short lead times.

Beyond rocketry, the general availability of orbital launch capacity enabled a rapidly expanding ecosystem, and new business models that are starting to reach profitable scale. This happened in tandem with the smartphone era of the 2010s which drove rapid iteration and cost reduction in electronic components, empowering start-ups wanting to design highly capable satellites. Commercial off-the-shelf smartphone-grade sensors, MEMS inertial units, GPS chips and lithium-ion cells all became space-qualifiable by around 2012.

If SpaceX is able to achieve rapid re-usability with Starship, they aim to ultimately launch 10,000 Starships per year (about one every hour), equating to two million tonnes of payload into orbit per year. This could bring launch costs down to close to the cost of fuel, and enable a whole new crop of space-based industries and services. 10,000 launches per year is a lot - 60x the 165 launches SpaceX completed in 2025 - but looks less out there in the context of 100,000 commercial flights that take off and land per day. In terms of energy, one Starship launch is about equivalent to 10 Airbus 380s flying for 10 hours. 

The agile aerospace era saw the founding or funding of many of the prominent players in today's space economy

Starlink

Starlink was first announced by Elon Musk in January 2015, which he described as a plan to "rebuild the internet in space". By placing a massive constellation of thousands of low-Earth orbit (LEO) satellites closer to the ground than traditional communications satellites, SpaceX aimed to drastically reduce latency and deliver high-speed broadband anywhere on Earth. The global telecom market is worth ~$2.5 trillion, split between cellular and fixed broadband. As of 2025, there were 8.3 billion 4G connections, 3 billion 5G connections and 1.6 billion fixed broadband connections globally (still, only 68% of the global population are connected to the internet). 

Starlink filed their first application with the FCC in 2016, seeking permission to launch 4,000 satellites. In October 2020 SpaceX launched a public beta testing phase named the "Better Than Nothing Beta"; SpaceX offered expensive phased-array terminals for $499, losing ~$2,500 on each sale. By mid-2021 they had sold 100,000 terminals, and Starlink service officially went live in October of that year. 

Just five years later, Starlink's ~10k active satellites serve over 10 million customers in 155 markets. The business represents more than two-thirds of SpaceX's revenues, with $11.4BN in sales in 2025 at a 63% EBITDA margin (Vodafone's was 39% last year). It is quite an achievement to make it cheaper to serve internet from space than established fibre and cell-tower network infrastructure; this was made possible by Elon Musk's characteristic laser-focus on manufacturing efficiency, scale and verticalisation (including the benefit of orbital launch capacity at non-commercial terms). Starlink is by far the largest satellite constellation in orbit, comprising more than two thirds of all satellites launched per year for each of the last five years.

Starlink has comprised more than two-thirds of all satellites launched into orbit in each of the last 5 years

Starlink has >10k active satellites criss-crossing the planet in LEO, mostly in mid-latitude inclinations. Pre-Starlink there were ~2,000 satellites in orbit in total

As of April 2026, SpaceX is launching 2-3 dedicated Starlink missions per week, with 24-29 Starlink V2 Mini satellites per launch, each with 100Gbps of total throughput capacity amounting to 4.8-8.7Tbps of new capacity every week. For some perspective, 1Tbps is enough bandwidth to stream roughly 200,000 high-definition Netflix shows simultaneously. Every week, SpaceX adds enough capacity to on-board more than a million heavy data users without slowing down the existing network.

At over 1,500kg, Starlink's V3 satellite is more than 2x the mass of the V2 Minis and has 10x the throughput per satellite. Starship can accommodate 60 V3s in its 9 metre bays, and deliver 20x the capacity of a Falcon 9/V2 mission per launch. Were Starship to start launching V3 satellites at the same cadence as Falcon 9/V2 missions, Starlink could double their maximum throughput every 8-10 weeks and reach 10% of the capacity of the global cellular network in 10 years (assuming 75k V3 satellites). 

This is unlikely to happen in practice - the global cellular network is itself growing, and total terrestrial capacity will accelerate with the roll out of 6G. Also, at the time of writing, SpaceX has only filed for approval of ~30k satellites with the FCC, of which only 15k are signed off. The company has commented previously that they their long term target for their satellite internet constellation is around 42k satellites. 

A single V3 satellite is designed to handle 1Tbps of downlink capacity and 160Gbps of uplink. This represents a 10x improvement in downlink and a 24x improvement in uplink over the V2 Minis

V3 satellites will be much larger and heavier, and so require Starship to launch

Starlink's network capacity is still tiny compared to the global cellular network

Starlink is not competitive with terrestrial cellular and fibre networks in dense urban areas, however SpaceX still estimates an addressable market of $1.6 trillion, or about two thirds of the $2.5 trillion overall.

Consumer Internet: Broadband and Mobile
ITU estimate 2.2 billion people globally still aren't connected to the internet. This is weighted toward rural areas, where internet penetration is 58%, versus 85% in urban areas. In emerging markets, lack of connectivity might largely be due to affordability rather than coverage, however, developed markets still have pockets of very poor connectivity in rural areas. As Starlink's network capacity increases, bandwidth and latency are increasingly competitive with terrestrial broadband options.This is aided by the use of ground stations, that effectively side-step last-mile fibre connectivity to remote places, and connect users to the main terrestrial internet backbone via ground stations (SpaceX doubled their ground station network to just over 500 last year). 

Starlink has also started to offer a satellite direct-to-phone service. At the moment, this is via partnerships with existing mobile carriers to supplement coverage dead spots in remote locations (e.g. T-Mobile in the US). This will change with V3 satellites, which will enable direct-to-cell connectivity at 5G speeds. At the moment, Starlink has >650 satellites providing direct-to-cell connectivity to 6 million customers, via 27 mobile network partners. 
 

Starlink uses ground stations to connect user terminals to the global internet backbone, significantly reducing latency 

Starlink's average residential download speed in the UK was 210Mbps in  2025 (+24% YoY) - competitive with 55% of the UK's full-fibre connections

Maritime and Aviation
Starlink covers around 150,000 vessels at sea, and about ~20% of the global merchant fleet above 100 gross tonnes (per Valour Consultancy). Recently announced deals include: Maersk installing Starlink on 330+ owned container vessels; Carnival deploying across its entire 90+ ship global fleet, serving nearly 13 million annual guests and 300,000+ guests and crew on board at any given time; Royal Caribbean across its global fleet of 64 ships; and Norwegian Cruise Line Holdings rolling out across its 29-ship fleet. Use-cases like cruise ships - which require a large number of concurrent users and very high bi-directional bandwidth, require a "Starlink Community Gateway". These cost $1.25 million in hardware/installation costs, followed by $75,000 per month subscriptions. 

Aviation is earlier in its adoption curve, with c.2,000 commercial aircraft contracted, versus a global fleet of over 38,000 (5% penetrated). United announced Starlink across more than 1,000 aircraft; American announced Starlink on more than 500 narrowbodies beginning in Q1 2027; Air France started progressively equipping all aircraft from summer 2025; Qatar completed 54 Boeing 777 installations and is beginning their A350 rollout; Hawaiian completed Starlink across its Airbus fleet; Alaska will bring Starlink to its entire fleet by 2027; Southwest announced Starlink across its network from summer 2026; and WestJet had already equipped its 100th 737 by October 2025.

Starlink's average download speed on airlines is 8x the average of other providers

Entreprise, Defence, and Sovereign Infrastructure
Starlink's entreprise contracts can cover fixed site or mobile businesses on land, with contracts ranging from $250/month to $2,150 per month depending on bandwidth required (Starlink Community Gateways can be used for very high bandwidth requirements). Starlink provides similar connectivity for critically important use cases in government and defence (called StarShield). The economics and specifications of StarShield are confidential, but we can see how effective uninteruptable internet coverage is in modern warfare by observing the role it has played in the Russia/Ukraine conflict and in the Middle East. 

Autonomy & Robotics
Autonomy forms a superset of all the segments above, encompassing consumer, entreprise and civil/defence markets on land, in the air and at sea. 

Global coverage of low latency internet connectivity will become increasingly important as self-driving vehicles and autonomous robots/drones become ubiquitous. While a lot of effort will go into cramming as much offline inference compute as possible into these various form-factors, ultimately the very long tail of unusual situations presented in the real world will necessitate connectivity to request access to more advanced models running in datacentres or human teleoperation. 

As with consumer connectivity, terrestrial 5G networks will be the primary infrastructure for these technologies, especially in urban areas, however satellite internet will provide a critical role outside areas of dense 5G coverage. 

Starlink is targeting 25 million subscribers by the end of the year, but ultimately has sights set on hundreds of millions of subscribers by the late 2030s

The consumer is by far the largest addressable market and fastest growing of these segments, however also the lowest average revenue per user (ARPU). Starlink has been running aggressive promotional deals, offering free equipment and plans (temporarily) as low as $29/month in the US which increased subscribers by 4x at a cost of a -18% YoY decline in ARPU to $81/month. 

While Starlink is increasingly cost-competitive with terrestrial alternatives, there are other obstacles to scaling a satellite internet network aside from simply having to launch a lot of satellites into orbit, chiefly access to spectrum. Taking the US as an example, the FCC/Federal government controls 58% of all spectrum frequencies, which is used by the military, NASA and the FAA, amongst others. 27% has been auctioned off to carriers like AT&T and Verizon, which have exclusive use.  

Spectrum is a finite asset, and much of it is still owned or controlled by governments

Last year, Starlink purchased 50MHz of S-band spectrum and 15MHz of AWS spectrum from Echostar, to advance their direct-to-cell ambitions (EchoStar quite profitably ended up with  2.8% stake in SpaceX as part of the deal). 

Other players are starting to enter the satellite internet market, most notably Amazon with their Leo constellation, which they recently bolstered through the acquisition of Globalstar for $11BN. This included 24 satellites in orbit, as well as ground stations and spectrum licenses to accelerate the build out of their coverage (Amazon recently won a contract with Delta Airlines). AST SpaceMobile is taking a different route, building satellites with very large phased arrays (the antennas on its latest BlueBirds unfold to over 220 square metres) designed to connect directly to unmodified smartphones at broadband speeds. Rather than competing with carriers, AST partners with them, with agreements covering AT&T, Verizon, Vodafone, Rakuten and others reaching ~3 billion subscribers, positioning itself as wholesale space-based capacity for the existing cellular industry.

SpaceX puts Starlink's total addressable market at $1.6 trillion by 2040 ($870 billion for broadband/$740 billion for mobile). As a purely illustrative example, 100 million subscribers, and a blended ARPU decline towards $50/month over the coming years gets to $60 billion in revenues (6x FY25) and $20 billion in FCF per year (not a forecast).

Earth Observation

Earth observation for military purposes pre-dates satellite communications by a hundred years - it started with tethered balloons in the mid-19th century, and camera-equipped aircraft played important roles in WWI and WWII. The precursor to modern space-based earth observation was the United States' CORONA program, which began with its first test launch in February 1959. Designed to photograph "denied areas" during the Cold War, these satellites took high-resolution photographs and dropped the film canisters back to Earth via parachute to be retrieved mid-air by aircraft. The Soviet Union launched its own photo-reconnaissance program, Zenit, in 1961. The 1960s and 1970s saw militaries transition from recoverable film capsules to digital transmissions and multispectral sensors that could capture images at night or through cloud cover. As with communications, satellites in the space economy 1.0 era were large, expensive, and few in number.

In 2011, three ex-NASA engineers founded Planet Labs (originally called Cosmogia) and by April 2013 had launched two demonstration CubeSats into orbit, designed to validate using off-the-shelf electronics for earth observation. By February 2017, they had 149 satellites in operation, and were collecting daily scans of the Earth's entire landmass at 3 metre resolution. These 'Doves' were the size of a shoebox, weighed 5kg and only cost $50,000 each. At around this time Maxar was the gold standard for commercial earth observation, with 10x better resolution of 30cm, but their satellites were in the larger, costlier and fewer class. 

Planet's Dove satellites had 150x less mass and cost 1,000x less than Maxar's high resolution satellites

Planet's 'agile aerospace' approach allowed them to quickly launch by far the largest Earth observation constellation into orbit

Dove satellites encircle the Earth in sun-synchronous orbit like a bracelet, scanning the entire landmass and some maritime regions once per day

The complexities of collecting high-frequency, high-resolution image data of the earth's 149 million square kilometre landmass would be intractable without AI. Traditional machine learning has been embedded in earth observation for well over a decade, helping apply heavy pre-processing to raw satellite imagery to correct for images taken at different angles (orthorectification) and to correct GPS positioning data to align images of the same area (co-registration). In addition, AI is used to combine images taken in many different spectral bands, helping capture more details and overcome cloud cover.

But even after all of the pre-processing, there remains an overwhelming amount of information. US agencies have thousands of personnel, and tens of thousands more private contractors involved with daily satellite imagery analysis. AI has been able to help with this too, with image classification models like convolutional neural networks (CNNs) which can be trained on large quantities of satellite imagery to identify specific features like buildings or ships.

It was clear to all involved that the combination of regular and comprehensive earth imagery with AI unlocked a much larger opportunity set than the prior decades of earth observation, which required tasking images of specific points of interest and the heavy involvement of expert analysts. In security & defence use cases, militaries could now track changes in naval ship building (including submarines at port), missile silo construction or troop build-ups anywhere on earth. But beyond the military, all kinds of other valuable use cases could be envisaged:
 

The volatile geopolitics of the last few years has seen many of these segments overlap - one recent example was the Russian invasion of Ukraine, during which companies like Maxar and Planet Labs provided invaluable situational awareness data to US and European governments, but also to journalists, helping to lift the fog of war and cut through social media disinformation. As the conflict dragged on, its implications for food security became critically important, given Ukraine's status as the "bread basket of Europe". NASA and Planet Labs combined optical and radar data with machine learning to enable daily field tracking to continuously observe planting, growing and harvesting phases, as well as classifying the type of crops across 5 million fields. High-frequency satellite data has been similarly employed to evaluate the impact on maritime traffic through the Strait of Hormuz, as well as assess the damage to critical energy infrastructure. 

NASA's estimates for Ukraine's 2022 harvest proved to be significantly more accurate than early wartime forecasts, which initially warned that up to 20-30% of winter crops might go unharvested. Instead, NASA’s satellite-based analysis revealed that 94% of the winter crop was actually harvested across the country, even in conflict zones

Three years on, NASA Harvest provides reliably accurate estimates for Ukrainian crops that are crucial for monitoring global food security

The ultimate vision of Planet Labs is to create the Bloomberg Terminal for Earth data - creating a queryable planet, with every feature on its surface indexed and categorised with the help of AI. But while the vision is clear, there have been technical barriers to fully unlocking the value. 

Chief among them is that image classification, while more or less a solved problem through the use of CNNs back in 2022, lacked generality. You could train one on container ships, or aircraft, or buildings, but for each of these you would lean heavily on human data annotation to train the models. Military customers have the resources to curate datasets for model training on the features relevant to them and design monitoring systems to make use of the models, but what about the local council that wants to monitor illegal construction projects, or the portfolio manager that wants to count cars in Walmart car parks over Thanksgiving?

This lack of generality and associated technical barriers to unlocking the full value of planetary data have been substantially addressed by generative AI and multi-modal foundation models. Meta has open-sourced incredibly accurate and general segmentation models on the permissive Apache 2.0 licence, the latest being Meta Segment Anything Model 2 (SAM2). And multi-modal foundation models from the frontier labs have innate world knowledge that can understand queries in natural language. Furthermore, their strength in coding enables them to spin up custom software for specific use cases and create easy dashboards for monitoring by non-technical users. 

Foundation model segmentation turns archival satellite data into queryable, trackable, vector-ready insights

On the hardware side, bandwidth has always been a challenge for earth observation. Satellites capture multi-spectral images down to 30cm resolutions, and are capable of generating hundreds of terabytes of raw data every day. This is 4-5x more than available downlink capacity, made worse by orbital mechanics resulting in the satellites only being in a position to broadcast to ground stations about 10% of the time.

Earlier this year, Planet Labs successfully used an NVIDIA Jetson Orin module to run an AI image processing model in space, aboard one of their Pelican satellites. By processing images locally, Planet can reduce the amount of data that needs to be transmitted to Earth by several orders of magnitude, and reduce time to answer from 2-4 hours to minutes. 

Imagery captured by Pelican-4 in March 2026 over Alice Springs, Australia, demonstrating the first successful deployment and execution of AI-driven object detection directly onboard Planet spacecraft

Against the backdrop of all of this exciting technological advancement, the EO market is surprisingly small, with estimates ranging from $5-10 billion depending on what you include in it. By comparison, major financial data companies, such as S&P Global, Moody's, FactSet, MSCI and Bloomberg have combined revenues of well over $40 billion. There are relatively few players in the data layer, given the barriers to actually building, launching and operating constellations of satellites in LEO, and a much longer tail of fragmented analytics companies processing the data and selling it on. 

In recent years, growth has largely been driven by the Defence & Intelligence side of things (to use Planet's categorisation). This isn't surprising given the major geopolitical conflicts in Ukraine and the Middle East, as well as the realisation in Europe that more investment must be made in domestic defence capability. As well as multiple federal agencies in the US (NRO, NASA, NOAA, NGA and the US Navy), Planet has multi-year deals with the EU and NATO, as well as directly with the government of the UK, Sweden, the Czech Republic and Greece. Several of these have a total contract value exceeding nine figures. While not all D&I relationships are publicly disclosed, it is clear there is considerable room to expand this list, given notable absences of countries in parts of the world with elevated geopolitical risk. 

Commercial adoption is still very early - over time, we expect the opportunity to be larger than military and non-military government combined

Planet Labs' revenue growth has re-accelerated since it became clearer that the AI inflection created a similar inflection for EO. Bloomberg consensus forecasts +24% top line growth CAGR from 2019-2030e doubles revenue every three years

Planet Labs is used as a case study in this section because its unique dataset gives the company a differentiated position. Its daily planetary scans over the past eight years have created a database containing thousands of images of almost every location on Earth - this temporal depth sets Planet apart from EO peers and, when combined with AI, has the potential to unlock entirely new use cases and sources of value. But looking at the industry more broadly, we expect the combination of EO and AI to enable a significant expansion of TAMs across sector verticals that are not yet reflected in growth forecasts at company-level. Planet's CEO, Will Marshall, pegs the company's total addressable market at $75-100 billion - ~10x larger than the current EO market and~300x larger than Planet's revenues last year. This sits somewhat in the middle of the peer group, with BlackSky's $40 billion at the lower end and Satellogic's $140 billion at the upper.

The economic value that could be unlocked by earth observation is larger still - in a report published in 2024, the World Economic Forum estimated that EO data created $266 billion in economic value in 2023. Agriculture, mining and oil & gas account for nearly three quarters of this, with the latter two industries much higher adopters, due to agriculture being more fragmented. By their estimates, a +33pts rise in adoption across sector verticals could generate $702 billion in economic value by 2030. 

The World Economic Forum view the 2023-30 period as the steepest part of the S-curve for EO adoption, with every +1ppt increase in adoption generating ~$10 billion in economic value

By their methodology, EO could generate over $700 billion in economic value by 2030, with 94% of this potential accruing to six sectors

Looking horizontally across sectors, there is over half a trillion dollars in commercially measurable value to address in precision agriculture, supply chain monitoring and vulnerability analysis

Datacentres in Space

Most people today would find the concept of orbital datacentres (ODCs) quite fanciful, but the idea has gained serious momentum over the last six months. Massive terrestrial AI clusters are increasingly facing local opposition, with growing perception amongst the broader public that they consume a lot of water and result in higher household energy prices. While neither of these concerns are necessarily true, polling shows seven in ten Americans oppose the construction of new datacentres in their neighbourhoods, and local politicians are starting to take notice. Maine, Georgia, Oklahoma, South Carolina and Vermont have all either passed, or are in the final stages of approving, pauses on new datacentre construction to allow time for environmental and infrastructure impact studies. Virginia (home to 35% of global hyperscale datacentres), Maryland, Ohio and Wisconsin have all passed "ratepayer protection" laws or strict siting requirements that make new projects significantly more difficult and expensive. And Michigan, Minnesota, New York and South Dakota are rolling back tax incentives that were originally designed to attract datacentre investment. 

Aside from the state-level regulatory and permitting headwinds, there are all kinds of physical bottlenecks throttling the build-out. Lead times for custom-engineered, high-voltage power transformers have stretched to 24–36 months, with some specialised units reaching 48 months, switchgear lead times are 12-18 months and diesel and gas backup generators are at 12-18 months. Lead times for the chillers and heat exchangers necessary for datacentre cooling are at 12-15 months. Longest of all is the wait time for new grid connections, which average 5 years (7 years in Northern Virginia). On the labour side, iRecruit estimate the industry is facing a deficit of nearly 500,000 tradesmen (especially electricians and plumbers). AI datacentres require 4,000-5,000 workers on site at peak construction. 

All of this is causing considerable cost inflation across all aspects of construction required to get to a fully energised warm shell in which to house server racks. This won't go away, with a synthesis of forecasts from various real estate consultants producing an estimated annual inflation rate of +7% over the next decade.

Time spent in the US grid interconnection queue has risen 3x in the last 15 years

ODCs don't have any of these problems, though they do have some of their own. We'll start with the advantages:

There are a few challenges with designing compute racks for space, and, while they are engineering-related rather than constraints set by the laws of physics, they are still quite hard problems to solve (cheaply):

Cooling
While space is cold, and there is no need for expensive liquid-cooling loops and chillers, heat must be radiated passively through radiators. This is governed by the Stefan-Boltzmann law, by which higher temperature exponentially increases the radiant energy emitted per unit area, meaning the hotter something is, the smaller the surface area of radiator needed to cool it. Elon Musk has commented that "If you increase operating temperature by +20% in degrees Kelvin, you can cut radiator mass in half". Radiators can easily take up 30-50% of a satellite's mass budget, and this gets worse with higher energy density (which is the direction datacentre rack designs are trending). Big reductions in radiator surface area could be achieved by designing chips that can run hotter - Blackwell GPUs currently run at 75-85°C and get throttled above 92°C. Running chips too hot shortens their life and increases the risk of burn out. 

Radiation
There is much more radiation outside of the protection of Earth's magnetosphere, though there have also been a lot of satellites in orbit for a long time, so the effect on semiconductors is quite well understood. As part of Project Suncatcher, Google researchers subjected TPUs to a 67 MeV proton beam using a cyclotron accelerator. The tests simulated a 5-year sun-synchronous Low Earth Orbit (LEO) mission, which corresponds to a cumulative total ionizing dose of about 750 rad(Si). The results showed that the TPUs survived a 15krad dose - roughly 20x the expected dose over 5 years of operation - without any permanent hardware failures. Starcloud did similar testing on GPUs with both proton and heavy ion beams, and have had an operational H100 in orbit for several months now. There is a separate problem of "bit flips" - where a high-energy particle changes a 1 to a 0 or vice versa within the chip. This actually doesn't impact AI workloads in GPUs too much, as they are fundamentally probabilistic and can deal with some noise. High Bandwidth Memory (HBM) is significantly more sensitive to radiation than GPU core logic primarily due to its physical architecture and density. Their "critical charge" - the amount of electrical energy required to hold a binary 1 or a 0 - is far lower. Because this energy threshold is so low even a weak cosmic ray or an alpha particle possesses enough energy to overcome the critical charge and flip the bit. This problem can be solved with some shielding, though at a cost to the mass budget. 

Bandwidth
We covered this a bit in the Earth Observation section - LEO satellites may only be in a position to broadcast information down to terrestrial ground stations about 10% of the time. When it comes to inference, this might be ok, as once the computation has been completed, the bitstream to send down to the user is quite small. Lasers can be used for satellite-to-satellite communication (Starlink already does this). Training clusters in orbit would be much more challenging, as large constellations of satellites would need to fly in formation. Google's Project Suncatcher pre-print envisages clusters of 81 satellites orbiting within a 1km radius, linked by optics running at tens of terabits per second (they have demonstrated 1.6Tbps in the lab), with the first two TPU-equipped satellites launching by early 2027 (to be built and operated by Planet Labs). A key challenge will be maintaining the formation which will be affected by the non-sphericity of the Earth's gravitational field and slight atmospheric drag, even at a 650km orbit. This will require argon thrusters with enough argon reaction mass to keep each individual satellite aligned correctly for the five year useful life. 

Maintenance
The answer to maintenance is to not do any. Satellites must be designed for graceful degradation and ultimately treated as disposable - deorbited and replaced rather than repaired. Hyperscalers already depreciate GPUs over roughly five years, conveniently similar to the lifespan of a satellite in very low Earth orbit. Starlink already operates this model at scale - hundreds of satellites are deorbited and replaced by newer versions every year. The highest failure rates for GPUs happen in the burn-in phase right at the beginning, which could take place on Earth (though semianalysis still estimate a 3-6% failure rate over their useful lives, so some over-provisioning would be required). 

It should be noted that the tractability of these various engineering challenges is an area of active debate. It does seem possible though - SpaceX released a reference design for their first AI compute satellite this week. It's 150kW peak payload roughly equates to an NVIDIA B300 NVL72 rack, and SpaceX think they can achieve 70kW/tonne in terms of mass. Starship could potentially stow ~40 of these satellites, or 6MW of AI compute. On these numbers, you could put a GW of AI compute into orbit per year with a Starship launch every other day. 

SpaceX recently announced a reference design for their first orbital compute satellite

We asked Claude Fable to reverse engineer the mass of the AI1 satellite's subsystems based on SpaceX's figures, and their 70kW/tonne target - solar PV and radiators make up about 50% of the satellite's mass

The question of cost is even more contested than the question of engineering. The SpaceX/Starcloud camp think ODCs will become cost-competitive towards the end of the 2020s, while others, such as Semianalysis are forecasting ~2039. Google/Planet Labs sit somewhere in the middle ("within ten years"). Even the longer timelines aren't that long compared to major engineering programs - it took Lockheed Martin 26 years to develop the F35 and NASA's Artemis program - the second phase of which just successfully sent astronauts on a trip around the Moon and back - was first established in 2017. 

On Starcloud's/SpaceX's numbers, Orbital datacentres could become cost-competitive with terrestrial ones in 2028, but this largely depends on Starship delivering on projected declines in mass to orbit and some heroic learning rates when it comes to designing and manufacturing bespoke orbital compute racks

The main ODC players so far are Google (in partnership with Planet Labs), Blue Origin/Amazon (in partnership with Starcloud), and, of course, SpaceX. SpaceX is by far the most ambitious, having already filed an application with the FCC to operate an "Orbital Data Center System" comprised of up to a million satellites. Both AWS and Starcloud have also filed with the FCC, but for much smaller constellations of up to 51,600 and 88,000 satellites respectively. 

Aside from the obvious advantages of being vertically integrated with mass-to-orbit capability, SpaceX's merger with xAI has resulted in the most vertically integrated company in what Jensen Huang calls the "five-layer cake" of AI. The chip design layer is borrowed from Tesla, who have experience designing highly efficient inference accelerators for Tesla's autonomous driving system, and, longer term, the two Musk companies will work together to build their own chip fabs. This project has already broken ground, and will initially licence Intel's technology for manufacturing, design and advanced packaging. The Terafab has the ultimate goal of being able to produce a terawatt of AI chips per year (1,000GW) - a monumental quantity by any measure. To put in perspective, around 20GW of AI compute will come online in the US this year, and perhaps 100GW cumulatively over 2025-30. 

This scale of ambition is hard for us to compute, but if we bring things closer to earth again, there is a fairly reasonable argument to be made that ODCs will become steadily cheaper over the next ten years, while TDCs will become more expensive, and, potentially, more unpopular. 

SpaceX is the most vertically integrated in the AI stack, followed by Google. Amazon, Oracle and Microsoft are all missing the most lucrative model layer

Mach33 research draws a parallel between terrestrial datacentres (TDCs) and orbital datacentres (ODCs) and this famous chart - TDCs look likely to follow the path of regulated, people intensive sectors, with prices rising over time, while ODCs can benefit from technological innovation, driving costs lower over time

If orbital datacentres prove viable on a 5–15 year horizon - which remains an open engineering and economic question - we can start to think about this theme from an investment perspective. The good news is that much of the domain knowledge acquired during the current AI infrastructure boom will be useful for investors in the potential ODC boom of the 2030s. Maybe GPUs/XPUs and memory silicon get a makeover to harden them against the g-force of rocket launches (perhaps NVIDIA can repurpose their GeForce trademark) or against long-term exposure to radiation, but they will remain fairly close cousins to the chips powering AI datacentres on Earth. The optics theme we are just embarking on will continue to provide an important role, not least for their lower thermal overhead vs. copper, and lasers will form the backbone of satellite to satellite communication. And the GaN/SiC components that are featuring in new NVIDIA 800V DC system designs will be equally useful in AI satellites, given the need for power and thermal efficiency, and that solar PV supplies DC power natively. 

Moving Heavy Industry Off-planet

The Moon's escape velocity is one-fifth of Earth's, and kinetic energy scales with the square of velocity, so the energy cost of launching a kilogram from the Moon's surface is roughly 4% of lifting it off Earth, Crucially, the Moon has no atmosphere. There is no drag to burn up a hypersonic payload at ground level, and no weather, which is what makes electromagnetic launch feasible. A mass driver converts cheap solar electricity directly into orbital delivery, at a marginal cost per kilogram that chemical rocketry can never approach. And the raw material is already there - lunar regolith is roughly 40% oxygen by mass, alongside silicon, aluminium, iron and titanium - all useful for solar panels, structures and propellant oxidiser.

Blue Origin's Blue Alchemist program uses molten regolith electrolysis to produce solar cells, transmission wire, silicon and oxygen from (simulated) lunar regolith, and has passed its critical design review; a full end-to-end autonomous demonstration - regolith in, working solar cells out, inside giant vacuum chambers, without human intervention - is scheduled for 2026, in partnership with NASA.

None of this makes economic sense without orbital datacentres to drive the launch cadences and mass-to-orbit cost declines. And it isn't technically feasible without a ChatGPT moment for physical AI — humanoid robots, or some other generally capable form-factor, operating autonomously in vacuum on nothing but solar power. Physical AI completely disrupts the off-world labour model by removing the immense cost of human life-support systems, creating a purely capital-driven, infinitely scalable workforce.

Ad Astra

Robert Heinlein is quoted as saying “once you get to Earth orbit, you’re halfway to anywhere in the solar system". The 400km journey to get there is extraordinarily hard: 92% of Starship's 5kt mass at launch is propellant - using chemical rockets to reach orbit would be infeasible were Earth's gravity just 1.5x stronger.

Humanity's first foray to orbit drew its impetus from the West vs. the USSR dynamic of the Cold War. When this motivator faded, so did much of the government backing, and humanity's extra-terrestrial ambitions had only economic incentives to fall back on. In the decades leading up to the 2010s, these were insufficient to drive the exponential flywheel of innovation and scaling that we have seen in other areas of technology.

As AI infrastructure scales on Earth, so too will regulatory and physical constraints. Investing in the necessary launch capacity and engineering to develop AI infrastructure in orbit will take tens of billions -sums that are very large in the context of Space Economy 1.0, but modest compared to terrestrial datacentre spending and already part-justified by newer, rapidly growing LEO industries such as satellite connectivity and earth observation. Orbital datacentres may ultimately attract trillions in capital, computation capacity which, in turn, could generate tens of trillions of downstream value to the global economy. This investment will drive down launch costs and advance technologies that enable further expansion, whether it be creating a lunar industrial base, or even more ambitious goals beyond Earth's gravity well. In AI, we finally have a techno-capital loop powerful enough to drive the next phase of expansion into the solar system.

AI has the potential to create tens of trillions in economic value - a revenue base large enough to fund expansion into the solar system

Acknowledgements and Further Reading