[0:10]Hello and welcome to the AI Supercycle Podcast, where we follow the capital flows and physical constraints shaping the largest industrial build-out of our generation. From semiconductors, AI factories and hyperscale data centers, to energy markets and power grids, critical materials, orbital compute, and embodied intelligence. We track where the capital is being deployed, where the binding constraints are emerging, and what it means for traders, investors, and the broader economy. Now I'm a little late with this episode. I spent rather too much time experimenting with voice cloning technology to see whether I could use it to help produce the podcast. It didn't quite go to plan. The results sounded like me, but not quite like me. There were clicks, odd pronunciations, and if I'm being completely honest, I thought the AI version of me was in danger of sending everyone to sleep. I hope the human me doesn't have that effect. So after spending far too long editing a 35-minute recording that was supposed to save me time, I abandoned the experiment. But we got there in the end, and this is the real human me recording on Wednesday, the 8th of July, 2026. I'm Tim Hardwick, founder of Quantum Fields Market Intelligence, and your host for this episode. So with that little production experiment behind us, let's get into today's show.
[1:32]Over the past three episodes, we have worked through three of the binding constraints shaping the AI supercycle. High bandwidth memory in episode 2. Delivered power in episode 3. Critical materials in episode 4. These constraints are real, they are tightening, and they are all terrestrial. Every one of them is a problem of building things on the ground. Fabs need power, power needs a grid, grids need copper, copper needs mines, and mines need permits.
[2:06]The entire supply chain is anchored to the surface of the Earth, and it is running into limits set by physics, geology, and politics. So in this episode, we asked the question that episode 4 set up. I positioned orbital compute as a possible pathway around some of these terrestrial constraints. And on the surface, it is an attractive proposition. No grid, no water, no permitting, unlimited solar energy. But the reality is more nuanced than that. Space does not bypass constraints, it trades them. You escape the power grid, the water consumption, the land permitting and the community opposition. In return, you inherit radiation exposure, latency limitations, launch dependency, urgency, orbital debris risk, and the impossibility of physical maintenance. And the satellites themselves still require copper, rare earths, and semiconductors. The materials constraint from episode 4 follows you into orbit.
[3:05]The supply chain does not end at the launch pad, it extends through it. So the real question is not whether space solves the constraints we have documented, it is whether the orbital set of constraints is more tractable than the terrestrial set, for which workloads, on what timeline, and at what cost. In the previous episode, I introduced the concept called the AI Continuum, the physical stack that runs from underground mines through surface infrastructure and upward into orbit. Every layer depends on the ones below it and constrains the ones above. In this episode, we reach the top of that stack. Orbital compute and data centers in space. This is no longer science fiction. It is an investment thesis with hardware in orbit, billion-dollar capital commitments, and the largest IPO in financial history behind it. I should say at this point that orbital infrastructure is not a new area of interest for me. I've been tracking the idea of space-based data centres since around 2020, including early companies such as Orbit's Edge. Before that, I was following the development of low-Earth orbit communications through OneWeb, Utelsat and Starlink. And of course, I've followed SpaceX since the early launches, like many of us.
[4:24]So this does not feel to me like a completely new story. What has changed is the scale, the capital and the credibility. Several threads I've been following for years are now beginning to converge. Launch economics, satellite connectivity, AI compute, and more recently, the terrestrial constraints that we have been examining throughout this podcast. That convergence is why I think the thesis deserves to be revisited now, not as a science fiction project, and not as a distant technology curiosity, but as something that has moved into the early stages of execution. But experience also tells me that this is exactly the point where we need to be careful and do some more in-depth analysis. I'm not saying space solves the constraints we have already documented. Many of them follow you into orbit. I am saying that serious capital, serious engineering, and serious companies are now pursuing this with a level of commitment that demands analytical attention, and that the intersection of space economics, AI demand, and terrestrial constraints is creating an emerging asset class that many institutional investors are just starting to model. So let's take a closer look for the case of orbital compute. It starts with the constraints that we've already been talking about in this podcast. And let's be specific about how those constraints are compounding, because that is what makes the orbital thesis more than a technology curiosity.
[5:52]On the power side, terrestrial data centres consumed roughly 415 terawatt-hours of electricity in 2024. To put that into perspective, that is more than the entire electricity consumption of the United Kingdom. The IEA projects that figure could double to reach around 950 terawatt-hours by 2030.
[6:15]Grid interconnection queues in the key markets like North Virginia, California and Germany now stretch 7 to 12 years. The hyperscalers are going off-grid with gas turbines and nuclear because the grid cannot keep up. And this is what we covered in detail in episode 3. On the water side, a typical AI data centre consumes millions of gallons of water annually for evaporative cooling. Communities from Virginia to Frankfurt to Amsterdam are pushing back against the draw on resources. Some jurisdictions are now blocking new data centre permits specifically because of water consumption. On the land side, data centre campuses are enormous and getting larger. A gigawatt scale campus requires hundreds of acres, plus the surrounding infrastructure for power, cooling, connectivity and physical security. Permitting and community opposition are slowing construction timelines.
[7:14]And on the materials side, as we covered last week in detail, copper, rare earths, and the entire critical materials supply chain are tightening on geological timescales.
[7:26]Every new data center requires thousands of tons of copper. Every piece of grid infrastructure connecting it requires more. These constraints are not independent, they compound. A copper shortage delays transformers. Transformer delays slow grid connections. Grid delays push hyperscalers off-grid. Off-grid solutions require more land, more water, and more permitting. The terrestrial supply chain is a system of interlocking bottlenecks. And this, for me, is where the orbital thesis becomes more interesting. Not because space suddenly looks easy. It certainly does not. But because the comparison is changing.
[8:06]Orbital compute does not need to become perfect. It only needs to become competitive with the terrestrial system that is in itself becoming more constrained, more expensive, and slower to build. That is an important distinction, because the economics of the thesis depend on both sides of the equation. Now consider what space offers, and I want to be specific here because the orbital compute thesis lives or dies on engineering credibility. When it comes to solar power, not all orbits are equal, and this matters enormously. In what is called a dawn-dusk sun-synchronous orbit, a satellite travels along the Terminator line, which is the boundary between day and night on Earth's surface. In this specific orbit, the satellite avoids Earth's shadow almost entirely. It collects sunlight roughly 99% of the year.
[8:59]But that is one particular orbit. A standard orbit at 550km altitude spends roughly 35-40% of each 90-minute pass in shadow. So the orbit you choose determines the energy you get. This is not a detail. It is a fundamental design decision. In the best case, orbital solar arrays generate roughly 5-8 times more energy annually than identical panels on the ground. Google's SunCatcher study found up to eight times in their modelled configuration.
[9:33]Current space-grade solar cells use advanced multilayer semiconductor designs. They achieve 30-34% efficiency, with the most advanced reaching up to 37%. That is significantly better than terrestrial panels, but it is not the 50% figure you sometimes see quoted. Then there is thermal management. Space is cold, roughly minus 270 degrees Celsius in deep shadow. But a satellite in direct sunlight faces very different conditions. Sun-facing surfaces can reach 850 degrees or more. Thermal management in orbit is not free. It requires heat pipes, radiation panels and active thermal control systems. What it does not require is water. A terrestrial data centre consumes millions of gallons annually for cooling. An orbital data centre consumes none This is a genuine advantage But it is not the same as saying cooling is free The engineering shifts from water and chillers To radiators and thermal control.
[10:37]No land, no grid, no permitting, no community opposition, no water rights. And these are all very real advantages. On paper, the physics offer genuine benefits, but the economics are the binding question. And they are more nuanced than most enthusiastic proponents suggest. And there is a set of constraints that space introduces which do not exist on the ground. Radiation degrades electronics on a timeline measured in years, not decades. Latency between satellites limits the types of workloads you can run. Hardware failures in orbit cannot be repaired. They become debris. The satellites themselves must be manufactured on Earth, launched from Earth, and built from the same constrained materials we discussed last week. Space does not escape the supply chain, it extends it vertically. Let's now take a look at some of the key players. This is not one company's moonshot. At least eight organisations have filed plans. launched hardware or committed capital to orbital compute within the past six months. First, SpaceX and XAI. The merger completed in February at a combined valuation of $1.25 trillion.
[11:49]On June 12, SpaceX listed on the Nasdaq in the largest initial public offering in financial history. It raised $75 billion at a valuation of $1.77 trillion. The stock has been volatile since. It peaked above $225 before pulling back into the mid-150s. The valuation price is in a future that has not yet arrived. The listing prospectus disclosed some remarkable numbers from the AI side. Anthropic has agreed to pay $1.25 billion per month through May 2029 for access to XAI's Colossus Data Center near Memphis.
[12:28]Google has agreed to pay $920 million per month starting in October. Those are terrestrial compute contracts, not orbital ones, but they show the scale of demand for compute. And Musk has cited XAI's surplus capacity as evidence that orbital infrastructure is needed to meet demand that Earth cannot satisfy. SpaceX has filed with the Federal Communications Commission for authority to launch up to 1 million orbital data center satellites. Now, one million is an extraordinary number, and I do not think we should take it at face value as a forecast of what will actually be launched, but I have followed SpaceX for long enough to be cautious about dismissing the underlying direction simply because the opening ambition sounds implausibly large. With SpaceX, the more useful question has often been not whether the original vision is achieved exactly as stated, but how much of the infrastructure required to pursue that vision becomes real along the way. The first-generation spacecraft, called AI-1, has a deployed wingspan of 70 meters, wider than a Boeing 747. It sustains an average compute payload of 125 kilowatts. And underpinning all of this is the Starship program, the only launch vehicle on track to deliver the cost reductions that make orbital compute economics work.
[13:48]SpaceX's own language in the filing was striking. They wrote, Freed from the constraints of terrestrial deployment, within a few years, the lowest cost to generate AI compute will be in space. That is an extraordinary claim, and we'll test this shortly. Second, Google Project Suncatcher, announced in November 2025. Google is pairing its custom AI chips with Planet Lab's satellite expertise. Radiation testing confirmed the chip can withstand five years in low Earth orbit. Lab tests achieved 1.6 terabits per second of data transfer using a single transceiver pair. The architecture envisions clusters of up to 81 satellites flying in formation within a one-kilometer radius. Two prototypes are planned for early 2027. Third, Blue Origin. Jeff Bezos' company revealed two orbital infrastructure programs. The first is TerraWave, a communications constellation of over 5,000 satellites designed as the connectivity backbone for orbital compute. The second is Project Sunrise, proposing over 50,000 compute satellites. Blue Origin's New Glenn rocket provides an alternative to Starship for heavy lift deployment. Satellite deployment begins late 2027.
[15:07]Fourth, StarCloud, a Y-Combinator-backed company that placed the first NVIDIA GPU in space in November 2025. This is the H100, the same chip that powers most terrestrial AI data centers. It became the first company to train a language model in orbit. Starcloud 2, carrying next-generation chips, is scheduled for October this year.
[15:32]Fifth, Axiom Space. They launched the first two dedicated orbital data center nodes to low Earth orbit on January the 11th, riding Kepler Communications' optical relay network. These nodes run cloud computing, AI inference, and cybersecurity applications using Earth-independent processing. And sixth, NVIDIA itself. At its annual technology conference in March, Jensen Huang announced the Space One Vera Rubin module. A dedicated GPU platform designed for space. His words, space computing, the final frontier, has arrived. This is no longer a concept. There is hardware in orbit running production workloads. February 2026 was the first month in history where multiple orbital data center operators ran production compute in space at the same time. For me, this is an important change. For most of the time that I've been tracking this area, The thesis was about what might eventually become possible. We are now at the point where the language has to change. There is hardware in orbit. There are production workloads. There are launch manifests and prototype dates. We are no longer tracking an idea alone. We are tracking execution.
[16:52]And China is not absent. The three-body computing constellation has launched 12 operational satellites and has already run Alibaba's QEN3 model in orbit. This is a separate program with separate technology, separate governance and separate strategic objectives. The orbital compute race is bifurcating along the same geopolitical lines as the semiconductor and materials layers below it. The question is no longer whether orbital data centers will happen, It is which wave of deployment reaches cost parity with ground-based alternatives and who controls the infrastructure when it does, Now let's take a closer look at the economics And this is where the thesis gets tested The economics of orbital compute rest on one variable above all others, Launch cost per kilogram, Everything else flows from it.
[17:45]Google's own feasibility study, published in November 2025, concluded that if launch costs to low Earth orbit reach $200 per kilogram, orbital data center satellites could become cost-competitive with terrestrial energy costs. They projected this could occur around 2035.
[18:05]Independent analysis suggests full cost parity, accounting for satellite manufacturing, integration and ground infrastructure, would require costs closer to $100 per kilogram. Where are we today? Well, that depends on whose estimates you trust, and SpaceX has not published an official figure. What we can say is that the cost needs to come down by at least a factor of 10 from current levels. So can it happen?
[18:33]SpaceX's track record on cost reduction is extraordinary. Falcon 9 reduced launch costs by roughly a factor of 10, compared with legacy vehicles. Starship, if fully reusable at high flight rates, could theoretically deliver another order of magnitude reduction. But the words could theoretically and has demonstrated are different things. Starship has not yet achieved full reusability. The economics are directional, not proven. There is a structural advantage that partially offsets the launch cost gap, but it needs careful qualification. Proponents cite orbital solar energy costs as low as half a cent per kilowatt hour. That figure represents the marginal cost of collecting sunlight once the array is in orbit. It does not include the cost of launching the array, manufacturing the satellite or replacing it after five years. The fully loaded cost is significantly higher. But even on a fully loaded basis, it compares favourably with terrestrial data centre electricity at 3 to 8 cents per kilowatt hour, provided launch costs fall to the levels we discussed.
[19:46]The energy advantage is real. It is just not as dramatic as the marginal cost figures suggests. Zero water costs, zero land costs, zero grid connection costs. Those operational savings are genuine and they compound over the satellite's lifespan. When you model the total cost of ownership over a five-year satellite life cycle, the calculation is more balanced than either side suggests. The launch cost is a large upfront capital expenditure that terrestrial data centres do not face, The energy and water savings are recurring operational advantages that compound over that lifespan The question is where the crossover point sits And that depends almost entirely on what happens to launch costs, But there are structural disadvantages too Satellite hardware in low Earth orbit faces continuous radiation exposure that degrades electronics over time. Current estimates put the usable lifespan of compute hardware at roughly 5 years, depending on altitude, shielding, and chip architecture. That means continuous replacement cycles. Every 5 years, you launch a new generation. This is a feature if launch costs fall, because you continuously upgrade to the latest chips.
[21:05]It is a problem if launch costs stall, because you are perpetually rebuilding your infrastructure. And it creates a waste problem that is currently unsolvable. These are not recyclable, they burn up on re-entry. There is also the question of which workloads suit orbital compute. Training large language models requires massive data synchronization between processors. Latency between satellites makes this very difficult. Inference is a different story.
[21:36]This is the process of running a trained model to generate outputs. It can be split across independent processors far more easily and is much less sensitive to latency. The near-term commercial opportunity is likely to be inference, edge processing, and specific applications like satellite imagery analysis, not training, The in-orbit data center market is projected to reach roughly $1.8 billion by, 2029 and $39 billion by 2035, Even the most aggressive projections show space-based compute handling a single-digit percentage of total global AI workloads before 2035. This is a supplement to terrestrial infrastructure, not a replacement. The timeline really does matter. And the timeline is measured in years, not quarters.
[22:29]Now let's give the other side its due, because a good analytical framework does not just make the case. It tests it, and the skeptics of orbital compute are not uninformed. Some of them are very well informed indeed. The strongest critique came from SoftBank's Masayoshi Sun, who told shareholders in June that chip costs, launch prices, and the latency between satellites makes SpaceX's orbital compute vision irrelevant in the decisive years of the AI race. His argument deserves careful consideration, because it is not about whether orbital compute will eventually work. It is about whether it will work soon enough to matter.
[23:10]Let's now talk about latency. Latency is the most fundamental technical challenge. Distributed AI training requires massive data synchronization between processors. Now on the ground, that synchronization happens over high bandwidth interconnects within a single data center, where the processors are meters apart. In orbit, the satellites are hundreds or thousands of kilometers apart. The data links between them cannot yet carry enough information fast enough. And the satellites are constantly moving relative to each other. Google's own Suncatcher study acknowledged that large-scale training across orbital satellites would require data links supporting tens of terabits per second, achievable only with satellites flying within a kilometre of each other. That is a fundamentally different architecture from what SpaceX has proposed. The combination makes distributed training in orbit essentially unworkable with current technology. Which is why, as we discussed, the near-term opportunity is inference rather than training. But even for inference, there is a latency question. Every query has to travel up into orbit and every answer has to come back down. That round-trip adds milliseconds that are fine for batch processing or satellite imagery analysis but problematic for real-time applications like conversational AI. Orbital compute will not serve every workload. The question is whether the workloads it can serve are large enough to justify the economics.
[24:39]Space debris is another serious concern. There are currently thousands of active satellites in orbit, with SpaceX alone accounting for over 9,000 of those. Adding a million data centers would increase the orbital population by a hundredfold. Collision risk and orbital congestion are not hypothetical.
[24:58]There is a recognised scenario, sometimes called the Kessler effect, where one collision generates debris that causes further collisions, which generate further debris, in a cascading chain. That risk gets worse with every satellite launched. Also, maintenance is effectively impossible. On the ground, if a server fails, a technician replaces it. In orbit, if a compute node fails, it becomes debris. There is no repair infrastructure for satellites at this scale. Every failure is a write-off. And then there is the single point of dependency problem. Starship is the only launch vehicle on track to deliver the cost reductions orbital compute requires. Blue Origin's New Glenn is operational but has not yet demonstrated the cost of economics needed. If starship development stalls, or if costs fail to reach the projected levels, the entire thesis is delayed by years, if not decades, There is no plan B for heavy lift at these economics, The orbital compute thesis is, in practical terms, a starship thesis, And the sceptics are not wrong about the challenges The question is whether they are right about the timeline and whether the severity of terrestrial constraints changes the calculation faster than anyone expects.
[26:18]Before we get to the base case, there is one more dimension worth addressing. Orbital Compute introduces governance questions that do not exist for terrestrial infrastructure, and they are questions that nobody has answered yet. Terrestrial data centres operate within national jurisdictions. They are subject to data sovereignty laws, environmental regulations, and tax regimes. Data centres in orbit sit in a regulatory void. The Outer Space Treaty, written in 1967, governs activities in space, but it was not written for commercial compute infrastructure. It does not contemplate a million satellite data center constellation processing AI workloads for customers in hundreds of countries. This creates both opportunity and risk. Opportunity for companies seeking to operate beyond terrestrial regulatory constraints. Risk for governments seeking to maintain oversight of AI systems that may increasingly operate beyond their reach. If you can train a model in orbit, whose rules apply? The country that launched the satellite? The country that owns the company? The country whose citizens' data is being processed? Nobody knows, because nobody has had to answer the question before.
[27:34]We have already noted that China's three-body computing constellation is operational. The orbital layer of the AI supercycle is bifurcating along the same geopolitical fault lines as every other layer. Two technology stacks, two governance frameworks, two competing visions of how AI infrastructure should be controlled. That pattern is now extending from the ground into space. And there is also the SpaceX concentration question. If orbital computes scales and if SpaceX controls the dominant launch vehicle, the dominant communications network, and a leading AI model through XAI, then a single company occupies a position of strategic leverage across the entire orbital infrastructure stack that has no precedent.
[28:21]As we discussed, the Anthropic and Google contracts disclosed in the listing prospectus demonstrate the scale of commercial relationships that could extend from terrestrial to orbital infrastructure. The regulatory framework for orbital compute does not yet exist. That is either a feature or a risk, depending on your perspective and your position in the value chain. This is also why I find the subject so difficult to reduce to a simple bullish or bearish view. There are too many parts of the system moving at different speeds. The launch technology is advancing. The compute hardware is advancing. The communications architecture is advancing, the regulation is barely beginning, and the economics depend on the cost reductions that have not yet been demonstrated.
[29:07]So rather than trying to predict a single outcome, I think the more useful approach is to identify what would have to become true for the thesis to accelerate, and what evidence would tell us that this is failing.
[29:22]Now for the base case, and this is where we take a position. My base case on orbital compute over the next three to five years is that the thesis is directionally correct, but the timeline is longer than most enthusiastic proponents suggest. The physics are compelling. The economics are conditional on launch cost reductions that have not yet been demonstrated. And of course, we have all of the engineering challenges. Space is hard.
[29:50]The investable opportunity in the near term is not orbital compute itself It is the infrastructure that enables it Launch services, satellite manufacturing, space-grade semiconductors, optical communications, These are the picks and shovels of the orbital compute thesis And several of them are already generating revenue, SpaceX, at its current valuation, is pricing in orbital compute as a meaningful future revenue stream, Whether that is justified depends on Starship economics. If Starship achieves full reusability at high flight rates and drives launch costs down dramatically by 2028, the thesis accelerates. If it does not, Orbital Compute remains a decade-plus horizon. The framework assigns a probability of approximately 40% that Orbital Compute reaches commercial cost parity, with terrestrial alternatives for specific workloads within five years. This is likely for inference rather than training. The probability rises to roughly 70% over a 10-year horizon. The principal risk is starship execution. The principal accelerant is terrestrial constraint severity. If the power, water and materials constraints documented in previous episodes tighten faster than expected, the economic case for orbital compute strengthens by default because the comparison set gets more expensive.
[31:17]We are actively tracking the data that will tell us whether this thesis is accelerating or stalling. Our Global Launch Intelligence Database logs every orbital launch this year. The mission objectives, the economics, the success rates, and the forward manifest. 150 orbital launches so far in 2026. That cadence is itself a data point. Beyond launch tracking, the indicators that matter most are Starship reusability milestones, the trajectory of launch costs, Google Suncatcher prototype results in early 2027, StarCloud 2 performance data later this year, and any regulatory framework for orbital compute that emerges from international bodies or national authorities. This is the base case as it reads today. If the data changes enough to warrant it, we'll update the view. So that is the orbital compute thesis, the newest layer of the AI continuum and the most speculative.
[32:15]We started this podcast series underground with the copper and uranium that come out of the earth. We moved up to the surface, to the power grids and the data centres being built on it. And now we have reached Orbit, where a growing number of companies believe that the future of AI compute will literally look down on the constraints below. The AI continuum runs from the mine to the antenna, from geological time to orbital mechanics, from the oldest constraints to the newest possibilities. And at every layer, the same question applies. Can supply keep pace with demand? For memory, the answer is not yet. For power, the answer is not fast enough. For critical materials, the answer is not on these timescales. For orbital compute, the answer is we just don't know yet. But serious people are spending serious money to find out.
[33:10]And perhaps that is what I find most interesting about this part of the AI supercycle. I've been following space-based data center concepts for a good few years now, and the wider orbital infrastructure story for much longer. For most of that time, these were separate threads. Launch economics was one story. Low Earth orbit communications was another. And AI compute is yet another story. And the terrestrial constraints we have spent the last few episodes examining were yet another story. What has changed is that those stories are beginning to converge. This is why I'm paying a lot of attention right now. Not because I think one million data center satellites are about to appear overhead, they are not. Not because I think terrestrial data centers are going away, they are not. And not because every ambitious projection in a SpaceX filing should be taken at face value, it should not. I'm paying attention because the evidence has changed. There is hardware in orbit, there are production workloads, there are dedicated chips and communications architectures being developed, there are prototype dates on the calendar, and there is now serious capital behind the thesis.
[34:24]After following this area for several years, this is the point at which an interesting technology story becomes something more. It becomes a thesis that needs to be tracked, tested and updated as the evidence comes in. I still don't know how this will resolve, and I don't think anybody does. But we know what to watch. Launch costs, starship reusability, satellite lifespans, inter-satellite bandwidth, the economics of inference, the sun-catcher prototypes, and, perhaps most importantly, whether the constraints on Earth continue to tighten quickly enough to make the alternatives in orbit more valuable. That is the work from here. What I find most striking about the orbital compute thesis is not actually the technology. It is what the technology reveals about the scale of the problem it is trying to solve. When the proposed solution to your infrastructure constraint is to leave the planet, it, that tells you something about how severe the constraint has become. The AI supercycle is not just the largest infrastructure build-out of our generation, it may be the first one that spans two environments, the surface of the earth and the space above it. And I suspect we're going to be returning to this subject quite a few times over the coming years.
[35:43]Okay, so before I go, just a brief production note. This episode arrived later than planned because of my experiment with an AI version of myself. The irony is not lost on me that an episode about putting AI into space was delayed because I couldn't get AI to sound convincingly like me. And perhaps there is a lesson in there somewhere. Anyway, this is the human version, and actually, I think it's probably better that way. Thank you for sticking with me. You can find all of our work at qfmi.substack.com. The Market Pulse and In The Spotlight articles are free and always will be. The weekly outlook, the weekend debrief, and the monthly strategic research report sit behind a paid subscription. Subscribe and you get the full picture. Thank you for listening to the AI Supercycle Podcast. See you next week.
[1:32]Over the past three episodes, we have worked through three of the binding constraints shaping the AI supercycle. High bandwidth memory in episode 2. Delivered power in episode 3. Critical materials in episode 4. These constraints are real, they are tightening, and they are all terrestrial. Every one of them is a problem of building things on the ground. Fabs need power, power needs a grid, grids need copper, copper needs mines, and mines need permits.
[2:06]The entire supply chain is anchored to the surface of the Earth, and it is running into limits set by physics, geology, and politics. So in this episode, we asked the question that episode 4 set up. I positioned orbital compute as a possible pathway around some of these terrestrial constraints. And on the surface, it is an attractive proposition. No grid, no water, no permitting, unlimited solar energy. But the reality is more nuanced than that. Space does not bypass constraints, it trades them. You escape the power grid, the water consumption, the land permitting and the community opposition. In return, you inherit radiation exposure, latency limitations, launch dependency, urgency, orbital debris risk, and the impossibility of physical maintenance. And the satellites themselves still require copper, rare earths, and semiconductors. The materials constraint from episode 4 follows you into orbit.
[3:05]The supply chain does not end at the launch pad, it extends through it. So the real question is not whether space solves the constraints we have documented, it is whether the orbital set of constraints is more tractable than the terrestrial set, for which workloads, on what timeline, and at what cost. In the previous episode, I introduced the concept called the AI Continuum, the physical stack that runs from underground mines through surface infrastructure and upward into orbit. Every layer depends on the ones below it and constrains the ones above. In this episode, we reach the top of that stack. Orbital compute and data centers in space. This is no longer science fiction. It is an investment thesis with hardware in orbit, billion-dollar capital commitments, and the largest IPO in financial history behind it. I should say at this point that orbital infrastructure is not a new area of interest for me. I've been tracking the idea of space-based data centres since around 2020, including early companies such as Orbit's Edge. Before that, I was following the development of low-Earth orbit communications through OneWeb, Utelsat and Starlink. And of course, I've followed SpaceX since the early launches, like many of us.
[4:24]So this does not feel to me like a completely new story. What has changed is the scale, the capital and the credibility. Several threads I've been following for years are now beginning to converge. Launch economics, satellite connectivity, AI compute, and more recently, the terrestrial constraints that we have been examining throughout this podcast. That convergence is why I think the thesis deserves to be revisited now, not as a science fiction project, and not as a distant technology curiosity, but as something that has moved into the early stages of execution. But experience also tells me that this is exactly the point where we need to be careful and do some more in-depth analysis. I'm not saying space solves the constraints we have already documented. Many of them follow you into orbit. I am saying that serious capital, serious engineering, and serious companies are now pursuing this with a level of commitment that demands analytical attention, and that the intersection of space economics, AI demand, and terrestrial constraints is creating an emerging asset class that many institutional investors are just starting to model. So let's take a closer look for the case of orbital compute. It starts with the constraints that we've already been talking about in this podcast. And let's be specific about how those constraints are compounding, because that is what makes the orbital thesis more than a technology curiosity.
[5:52]On the power side, terrestrial data centres consumed roughly 415 terawatt-hours of electricity in 2024. To put that into perspective, that is more than the entire electricity consumption of the United Kingdom. The IEA projects that figure could double to reach around 950 terawatt-hours by 2030.
[6:15]Grid interconnection queues in the key markets like North Virginia, California and Germany now stretch 7 to 12 years. The hyperscalers are going off-grid with gas turbines and nuclear because the grid cannot keep up. And this is what we covered in detail in episode 3. On the water side, a typical AI data centre consumes millions of gallons of water annually for evaporative cooling. Communities from Virginia to Frankfurt to Amsterdam are pushing back against the draw on resources. Some jurisdictions are now blocking new data centre permits specifically because of water consumption. On the land side, data centre campuses are enormous and getting larger. A gigawatt scale campus requires hundreds of acres, plus the surrounding infrastructure for power, cooling, connectivity and physical security. Permitting and community opposition are slowing construction timelines.
[7:14]And on the materials side, as we covered last week in detail, copper, rare earths, and the entire critical materials supply chain are tightening on geological timescales.
[7:26]Every new data center requires thousands of tons of copper. Every piece of grid infrastructure connecting it requires more. These constraints are not independent, they compound. A copper shortage delays transformers. Transformer delays slow grid connections. Grid delays push hyperscalers off-grid. Off-grid solutions require more land, more water, and more permitting. The terrestrial supply chain is a system of interlocking bottlenecks. And this, for me, is where the orbital thesis becomes more interesting. Not because space suddenly looks easy. It certainly does not. But because the comparison is changing.
[8:06]Orbital compute does not need to become perfect. It only needs to become competitive with the terrestrial system that is in itself becoming more constrained, more expensive, and slower to build. That is an important distinction, because the economics of the thesis depend on both sides of the equation. Now consider what space offers, and I want to be specific here because the orbital compute thesis lives or dies on engineering credibility. When it comes to solar power, not all orbits are equal, and this matters enormously. In what is called a dawn-dusk sun-synchronous orbit, a satellite travels along the Terminator line, which is the boundary between day and night on Earth's surface. In this specific orbit, the satellite avoids Earth's shadow almost entirely. It collects sunlight roughly 99% of the year.
[8:59]But that is one particular orbit. A standard orbit at 550km altitude spends roughly 35-40% of each 90-minute pass in shadow. So the orbit you choose determines the energy you get. This is not a detail. It is a fundamental design decision. In the best case, orbital solar arrays generate roughly 5-8 times more energy annually than identical panels on the ground. Google's SunCatcher study found up to eight times in their modelled configuration.
[9:33]Current space-grade solar cells use advanced multilayer semiconductor designs. They achieve 30-34% efficiency, with the most advanced reaching up to 37%. That is significantly better than terrestrial panels, but it is not the 50% figure you sometimes see quoted. Then there is thermal management. Space is cold, roughly minus 270 degrees Celsius in deep shadow. But a satellite in direct sunlight faces very different conditions. Sun-facing surfaces can reach 850 degrees or more. Thermal management in orbit is not free. It requires heat pipes, radiation panels and active thermal control systems. What it does not require is water. A terrestrial data centre consumes millions of gallons annually for cooling. An orbital data centre consumes none This is a genuine advantage But it is not the same as saying cooling is free The engineering shifts from water and chillers To radiators and thermal control.
[10:37]No land, no grid, no permitting, no community opposition, no water rights. And these are all very real advantages. On paper, the physics offer genuine benefits, but the economics are the binding question. And they are more nuanced than most enthusiastic proponents suggest. And there is a set of constraints that space introduces which do not exist on the ground. Radiation degrades electronics on a timeline measured in years, not decades. Latency between satellites limits the types of workloads you can run. Hardware failures in orbit cannot be repaired. They become debris. The satellites themselves must be manufactured on Earth, launched from Earth, and built from the same constrained materials we discussed last week. Space does not escape the supply chain, it extends it vertically. Let's now take a look at some of the key players. This is not one company's moonshot. At least eight organisations have filed plans. launched hardware or committed capital to orbital compute within the past six months. First, SpaceX and XAI. The merger completed in February at a combined valuation of $1.25 trillion.
[11:49]On June 12, SpaceX listed on the Nasdaq in the largest initial public offering in financial history. It raised $75 billion at a valuation of $1.77 trillion. The stock has been volatile since. It peaked above $225 before pulling back into the mid-150s. The valuation price is in a future that has not yet arrived. The listing prospectus disclosed some remarkable numbers from the AI side. Anthropic has agreed to pay $1.25 billion per month through May 2029 for access to XAI's Colossus Data Center near Memphis.
[12:28]Google has agreed to pay $920 million per month starting in October. Those are terrestrial compute contracts, not orbital ones, but they show the scale of demand for compute. And Musk has cited XAI's surplus capacity as evidence that orbital infrastructure is needed to meet demand that Earth cannot satisfy. SpaceX has filed with the Federal Communications Commission for authority to launch up to 1 million orbital data center satellites. Now, one million is an extraordinary number, and I do not think we should take it at face value as a forecast of what will actually be launched, but I have followed SpaceX for long enough to be cautious about dismissing the underlying direction simply because the opening ambition sounds implausibly large. With SpaceX, the more useful question has often been not whether the original vision is achieved exactly as stated, but how much of the infrastructure required to pursue that vision becomes real along the way. The first-generation spacecraft, called AI-1, has a deployed wingspan of 70 meters, wider than a Boeing 747. It sustains an average compute payload of 125 kilowatts. And underpinning all of this is the Starship program, the only launch vehicle on track to deliver the cost reductions that make orbital compute economics work.
[13:48]SpaceX's own language in the filing was striking. They wrote, Freed from the constraints of terrestrial deployment, within a few years, the lowest cost to generate AI compute will be in space. That is an extraordinary claim, and we'll test this shortly. Second, Google Project Suncatcher, announced in November 2025. Google is pairing its custom AI chips with Planet Lab's satellite expertise. Radiation testing confirmed the chip can withstand five years in low Earth orbit. Lab tests achieved 1.6 terabits per second of data transfer using a single transceiver pair. The architecture envisions clusters of up to 81 satellites flying in formation within a one-kilometer radius. Two prototypes are planned for early 2027. Third, Blue Origin. Jeff Bezos' company revealed two orbital infrastructure programs. The first is TerraWave, a communications constellation of over 5,000 satellites designed as the connectivity backbone for orbital compute. The second is Project Sunrise, proposing over 50,000 compute satellites. Blue Origin's New Glenn rocket provides an alternative to Starship for heavy lift deployment. Satellite deployment begins late 2027.
[15:07]Fourth, StarCloud, a Y-Combinator-backed company that placed the first NVIDIA GPU in space in November 2025. This is the H100, the same chip that powers most terrestrial AI data centers. It became the first company to train a language model in orbit. Starcloud 2, carrying next-generation chips, is scheduled for October this year.
[15:32]Fifth, Axiom Space. They launched the first two dedicated orbital data center nodes to low Earth orbit on January the 11th, riding Kepler Communications' optical relay network. These nodes run cloud computing, AI inference, and cybersecurity applications using Earth-independent processing. And sixth, NVIDIA itself. At its annual technology conference in March, Jensen Huang announced the Space One Vera Rubin module. A dedicated GPU platform designed for space. His words, space computing, the final frontier, has arrived. This is no longer a concept. There is hardware in orbit running production workloads. February 2026 was the first month in history where multiple orbital data center operators ran production compute in space at the same time. For me, this is an important change. For most of the time that I've been tracking this area, The thesis was about what might eventually become possible. We are now at the point where the language has to change. There is hardware in orbit. There are production workloads. There are launch manifests and prototype dates. We are no longer tracking an idea alone. We are tracking execution.
[16:52]And China is not absent. The three-body computing constellation has launched 12 operational satellites and has already run Alibaba's QEN3 model in orbit. This is a separate program with separate technology, separate governance and separate strategic objectives. The orbital compute race is bifurcating along the same geopolitical lines as the semiconductor and materials layers below it. The question is no longer whether orbital data centers will happen, It is which wave of deployment reaches cost parity with ground-based alternatives and who controls the infrastructure when it does, Now let's take a closer look at the economics And this is where the thesis gets tested The economics of orbital compute rest on one variable above all others, Launch cost per kilogram, Everything else flows from it.
[17:45]Google's own feasibility study, published in November 2025, concluded that if launch costs to low Earth orbit reach $200 per kilogram, orbital data center satellites could become cost-competitive with terrestrial energy costs. They projected this could occur around 2035.
[18:05]Independent analysis suggests full cost parity, accounting for satellite manufacturing, integration and ground infrastructure, would require costs closer to $100 per kilogram. Where are we today? Well, that depends on whose estimates you trust, and SpaceX has not published an official figure. What we can say is that the cost needs to come down by at least a factor of 10 from current levels. So can it happen?
[18:33]SpaceX's track record on cost reduction is extraordinary. Falcon 9 reduced launch costs by roughly a factor of 10, compared with legacy vehicles. Starship, if fully reusable at high flight rates, could theoretically deliver another order of magnitude reduction. But the words could theoretically and has demonstrated are different things. Starship has not yet achieved full reusability. The economics are directional, not proven. There is a structural advantage that partially offsets the launch cost gap, but it needs careful qualification. Proponents cite orbital solar energy costs as low as half a cent per kilowatt hour. That figure represents the marginal cost of collecting sunlight once the array is in orbit. It does not include the cost of launching the array, manufacturing the satellite or replacing it after five years. The fully loaded cost is significantly higher. But even on a fully loaded basis, it compares favourably with terrestrial data centre electricity at 3 to 8 cents per kilowatt hour, provided launch costs fall to the levels we discussed.
[19:46]The energy advantage is real. It is just not as dramatic as the marginal cost figures suggests. Zero water costs, zero land costs, zero grid connection costs. Those operational savings are genuine and they compound over the satellite's lifespan. When you model the total cost of ownership over a five-year satellite life cycle, the calculation is more balanced than either side suggests. The launch cost is a large upfront capital expenditure that terrestrial data centres do not face, The energy and water savings are recurring operational advantages that compound over that lifespan The question is where the crossover point sits And that depends almost entirely on what happens to launch costs, But there are structural disadvantages too Satellite hardware in low Earth orbit faces continuous radiation exposure that degrades electronics over time. Current estimates put the usable lifespan of compute hardware at roughly 5 years, depending on altitude, shielding, and chip architecture. That means continuous replacement cycles. Every 5 years, you launch a new generation. This is a feature if launch costs fall, because you continuously upgrade to the latest chips.
[21:05]It is a problem if launch costs stall, because you are perpetually rebuilding your infrastructure. And it creates a waste problem that is currently unsolvable. These are not recyclable, they burn up on re-entry. There is also the question of which workloads suit orbital compute. Training large language models requires massive data synchronization between processors. Latency between satellites makes this very difficult. Inference is a different story.
[21:36]This is the process of running a trained model to generate outputs. It can be split across independent processors far more easily and is much less sensitive to latency. The near-term commercial opportunity is likely to be inference, edge processing, and specific applications like satellite imagery analysis, not training, The in-orbit data center market is projected to reach roughly $1.8 billion by, 2029 and $39 billion by 2035, Even the most aggressive projections show space-based compute handling a single-digit percentage of total global AI workloads before 2035. This is a supplement to terrestrial infrastructure, not a replacement. The timeline really does matter. And the timeline is measured in years, not quarters.
[22:29]Now let's give the other side its due, because a good analytical framework does not just make the case. It tests it, and the skeptics of orbital compute are not uninformed. Some of them are very well informed indeed. The strongest critique came from SoftBank's Masayoshi Sun, who told shareholders in June that chip costs, launch prices, and the latency between satellites makes SpaceX's orbital compute vision irrelevant in the decisive years of the AI race. His argument deserves careful consideration, because it is not about whether orbital compute will eventually work. It is about whether it will work soon enough to matter.
[23:10]Let's now talk about latency. Latency is the most fundamental technical challenge. Distributed AI training requires massive data synchronization between processors. Now on the ground, that synchronization happens over high bandwidth interconnects within a single data center, where the processors are meters apart. In orbit, the satellites are hundreds or thousands of kilometers apart. The data links between them cannot yet carry enough information fast enough. And the satellites are constantly moving relative to each other. Google's own Suncatcher study acknowledged that large-scale training across orbital satellites would require data links supporting tens of terabits per second, achievable only with satellites flying within a kilometre of each other. That is a fundamentally different architecture from what SpaceX has proposed. The combination makes distributed training in orbit essentially unworkable with current technology. Which is why, as we discussed, the near-term opportunity is inference rather than training. But even for inference, there is a latency question. Every query has to travel up into orbit and every answer has to come back down. That round-trip adds milliseconds that are fine for batch processing or satellite imagery analysis but problematic for real-time applications like conversational AI. Orbital compute will not serve every workload. The question is whether the workloads it can serve are large enough to justify the economics.
[24:39]Space debris is another serious concern. There are currently thousands of active satellites in orbit, with SpaceX alone accounting for over 9,000 of those. Adding a million data centers would increase the orbital population by a hundredfold. Collision risk and orbital congestion are not hypothetical.
[24:58]There is a recognised scenario, sometimes called the Kessler effect, where one collision generates debris that causes further collisions, which generate further debris, in a cascading chain. That risk gets worse with every satellite launched. Also, maintenance is effectively impossible. On the ground, if a server fails, a technician replaces it. In orbit, if a compute node fails, it becomes debris. There is no repair infrastructure for satellites at this scale. Every failure is a write-off. And then there is the single point of dependency problem. Starship is the only launch vehicle on track to deliver the cost reductions orbital compute requires. Blue Origin's New Glenn is operational but has not yet demonstrated the cost of economics needed. If starship development stalls, or if costs fail to reach the projected levels, the entire thesis is delayed by years, if not decades, There is no plan B for heavy lift at these economics, The orbital compute thesis is, in practical terms, a starship thesis, And the sceptics are not wrong about the challenges The question is whether they are right about the timeline and whether the severity of terrestrial constraints changes the calculation faster than anyone expects.
[26:18]Before we get to the base case, there is one more dimension worth addressing. Orbital Compute introduces governance questions that do not exist for terrestrial infrastructure, and they are questions that nobody has answered yet. Terrestrial data centres operate within national jurisdictions. They are subject to data sovereignty laws, environmental regulations, and tax regimes. Data centres in orbit sit in a regulatory void. The Outer Space Treaty, written in 1967, governs activities in space, but it was not written for commercial compute infrastructure. It does not contemplate a million satellite data center constellation processing AI workloads for customers in hundreds of countries. This creates both opportunity and risk. Opportunity for companies seeking to operate beyond terrestrial regulatory constraints. Risk for governments seeking to maintain oversight of AI systems that may increasingly operate beyond their reach. If you can train a model in orbit, whose rules apply? The country that launched the satellite? The country that owns the company? The country whose citizens' data is being processed? Nobody knows, because nobody has had to answer the question before.
[27:34]We have already noted that China's three-body computing constellation is operational. The orbital layer of the AI supercycle is bifurcating along the same geopolitical fault lines as every other layer. Two technology stacks, two governance frameworks, two competing visions of how AI infrastructure should be controlled. That pattern is now extending from the ground into space. And there is also the SpaceX concentration question. If orbital computes scales and if SpaceX controls the dominant launch vehicle, the dominant communications network, and a leading AI model through XAI, then a single company occupies a position of strategic leverage across the entire orbital infrastructure stack that has no precedent.
[28:21]As we discussed, the Anthropic and Google contracts disclosed in the listing prospectus demonstrate the scale of commercial relationships that could extend from terrestrial to orbital infrastructure. The regulatory framework for orbital compute does not yet exist. That is either a feature or a risk, depending on your perspective and your position in the value chain. This is also why I find the subject so difficult to reduce to a simple bullish or bearish view. There are too many parts of the system moving at different speeds. The launch technology is advancing. The compute hardware is advancing. The communications architecture is advancing, the regulation is barely beginning, and the economics depend on the cost reductions that have not yet been demonstrated.
[29:07]So rather than trying to predict a single outcome, I think the more useful approach is to identify what would have to become true for the thesis to accelerate, and what evidence would tell us that this is failing.
[29:22]Now for the base case, and this is where we take a position. My base case on orbital compute over the next three to five years is that the thesis is directionally correct, but the timeline is longer than most enthusiastic proponents suggest. The physics are compelling. The economics are conditional on launch cost reductions that have not yet been demonstrated. And of course, we have all of the engineering challenges. Space is hard.
[29:50]The investable opportunity in the near term is not orbital compute itself It is the infrastructure that enables it Launch services, satellite manufacturing, space-grade semiconductors, optical communications, These are the picks and shovels of the orbital compute thesis And several of them are already generating revenue, SpaceX, at its current valuation, is pricing in orbital compute as a meaningful future revenue stream, Whether that is justified depends on Starship economics. If Starship achieves full reusability at high flight rates and drives launch costs down dramatically by 2028, the thesis accelerates. If it does not, Orbital Compute remains a decade-plus horizon. The framework assigns a probability of approximately 40% that Orbital Compute reaches commercial cost parity, with terrestrial alternatives for specific workloads within five years. This is likely for inference rather than training. The probability rises to roughly 70% over a 10-year horizon. The principal risk is starship execution. The principal accelerant is terrestrial constraint severity. If the power, water and materials constraints documented in previous episodes tighten faster than expected, the economic case for orbital compute strengthens by default because the comparison set gets more expensive.
[31:17]We are actively tracking the data that will tell us whether this thesis is accelerating or stalling. Our Global Launch Intelligence Database logs every orbital launch this year. The mission objectives, the economics, the success rates, and the forward manifest. 150 orbital launches so far in 2026. That cadence is itself a data point. Beyond launch tracking, the indicators that matter most are Starship reusability milestones, the trajectory of launch costs, Google Suncatcher prototype results in early 2027, StarCloud 2 performance data later this year, and any regulatory framework for orbital compute that emerges from international bodies or national authorities. This is the base case as it reads today. If the data changes enough to warrant it, we'll update the view. So that is the orbital compute thesis, the newest layer of the AI continuum and the most speculative.
[32:15]We started this podcast series underground with the copper and uranium that come out of the earth. We moved up to the surface, to the power grids and the data centres being built on it. And now we have reached Orbit, where a growing number of companies believe that the future of AI compute will literally look down on the constraints below. The AI continuum runs from the mine to the antenna, from geological time to orbital mechanics, from the oldest constraints to the newest possibilities. And at every layer, the same question applies. Can supply keep pace with demand? For memory, the answer is not yet. For power, the answer is not fast enough. For critical materials, the answer is not on these timescales. For orbital compute, the answer is we just don't know yet. But serious people are spending serious money to find out.
[33:10]And perhaps that is what I find most interesting about this part of the AI supercycle. I've been following space-based data center concepts for a good few years now, and the wider orbital infrastructure story for much longer. For most of that time, these were separate threads. Launch economics was one story. Low Earth orbit communications was another. And AI compute is yet another story. And the terrestrial constraints we have spent the last few episodes examining were yet another story. What has changed is that those stories are beginning to converge. This is why I'm paying a lot of attention right now. Not because I think one million data center satellites are about to appear overhead, they are not. Not because I think terrestrial data centers are going away, they are not. And not because every ambitious projection in a SpaceX filing should be taken at face value, it should not. I'm paying attention because the evidence has changed. There is hardware in orbit, there are production workloads, there are dedicated chips and communications architectures being developed, there are prototype dates on the calendar, and there is now serious capital behind the thesis.
[34:24]After following this area for several years, this is the point at which an interesting technology story becomes something more. It becomes a thesis that needs to be tracked, tested and updated as the evidence comes in. I still don't know how this will resolve, and I don't think anybody does. But we know what to watch. Launch costs, starship reusability, satellite lifespans, inter-satellite bandwidth, the economics of inference, the sun-catcher prototypes, and, perhaps most importantly, whether the constraints on Earth continue to tighten quickly enough to make the alternatives in orbit more valuable. That is the work from here. What I find most striking about the orbital compute thesis is not actually the technology. It is what the technology reveals about the scale of the problem it is trying to solve. When the proposed solution to your infrastructure constraint is to leave the planet, it, that tells you something about how severe the constraint has become. The AI supercycle is not just the largest infrastructure build-out of our generation, it may be the first one that spans two environments, the surface of the earth and the space above it. And I suspect we're going to be returning to this subject quite a few times over the coming years.
[35:43]Okay, so before I go, just a brief production note. This episode arrived later than planned because of my experiment with an AI version of myself. The irony is not lost on me that an episode about putting AI into space was delayed because I couldn't get AI to sound convincingly like me. And perhaps there is a lesson in there somewhere. Anyway, this is the human version, and actually, I think it's probably better that way. Thank you for sticking with me. You can find all of our work at qfmi.substack.com. The Market Pulse and In The Spotlight articles are free and always will be. The weekly outlook, the weekend debrief, and the monthly strategic research report sit behind a paid subscription. Subscribe and you get the full picture. Thank you for listening to the AI Supercycle Podcast. See you next week.
