SpaceX - Powering AI from Orbit, Space Solar, and the Musk Stack (Pt.2)

SpaceX - Powering AI from Orbit, Space Solar, and the Musk Stack (Pt.2)

AI compute has outgrown Earth’s power, cooling, and land walls — Starship and advanced solar cell chemistry now make the orbital Musk Stack economically inevitable.

This article is part 2 in a series of 6. You can find the rest of the articles here:

P1 - P3 - P4 - P5 - P6

Summary

• The bet is that AI compute has outgrown Earth — power, cooling, and land are all hitting walls — and 1.2 million satellites can host it in orbit instead.

  
  

• Starship makes launch tractable, but cell chemistry is the real economic lever — HJT and perovskite tandems are what actually collapse the cost stack.

  
  

• The constellation is a three-tier fabric — VLEO for inference, LEO for training, MEO for storage — wired together by lasers into one distributed supercomputer.

  
  

• Space doesn't make cooling free, it makes it structural: the same panel that catches sunlight on one side dumps waste heat on the other.

  
  

• Vacuum has no chromatic dispersion, which rewrites the photonics supply chain — most links can use $10 lasers instead of $1,000 coherent modules.

  
  

• The endgame is a closed loop: xAI plans, Optimus builds, Tesla fabs, SpaceX launches — with the Moon eventually taking over from Earth as the deployment base.

Space AI Data Center Architecture

If Part 1 made the case for why space solar is the path forward, Part 2 turns to the harder question: what does it actually look like in orbit? The architecture Musk is assembling is not a gradual extension of today's satellite industry — it is a step-change in scale, and the numbers only make sense once you see how the pieces fit together.

Start with the filing itself. SpaceX has applied to the FCC to launch over 1.2 million satellites — more than every satellite humanity has ever put into orbit, combined. These satellites will form into a three-tier constellation designed to function as a single distributed compute fabric.

The first wave will be variants of the Starlink V3 platform already in production: roughly 2 tons each, 30 kW of solar capacity, and around 500 m² of panel area per satellite. Familiar hardware, unfamiliar purpose.

Three Orbital Tiers

Each tier is optimized for a different workload, with altitude trading off against latency, coverage, and thermal headroom.

• VLEO (500 km) — 800,000 satellites. The edge layer. Each satellite delivers roughly 1,000 TOPS (Tera Operations Per Second - FP8 but could change) with ~20 ms latency to ground users, handling real-time inference for everything from autonomous vehicles to consumer AI assistants.

  
  

• LEO (1,000 km) — 300,000 satellites. The regional layer. At ~5,000 TOPS per satellite, this tier carries the bulk of model training and mid-tier compute.

  
  

• MEO (2,000 km) — 118,000 satellites. The cloud layer. At ~20,000 TOPS per satellite, MEO handles long-running workloads and data storage, where latency matters less than density and stability.

  
  

Underneath all three tiers, the platform is the same. Each Starlink V3 satellite weighs roughly 2 tons, and nearly half that mass — about 850 kg — is dedicated to the compute payload itself: radiation-hardened AI silicon designed to survive cosmic ray strikes that would corrupt ordinary chips, liquid cooling loops to shed heat in an environment with no air for convection, and high-efficiency power conversion electronics that translate the solar array's raw output into the precise voltages the processors need. These aren't communications satellites with some compute bolted on. They're flying data centers, with the bus built around the silicon rather than the other way around.

The compute itself is a space-rated variant of Dojo 2 — Tesla's second-generation AI training chip, redesigned for the orbital environment — with 12 to 24 chips per satellite depending on the tier. Stitching them together are inter-satellite laser links, the connective tissue of the constellation, now running at 10 Gbps per channel (on Earth 8-16 channels per lane but in space there is no limit). Across thousands of simultaneous links, aggregate throughput climbs into the terabits per second, enough to let geographically scattered satellites behave as a single coherent compute fabric rather than a million isolated nodes.

And the launch math is what makes any of this tractable. At 2 tons per satellite, a single Starship flight — with its 100–150 ton payload capacity to LEO — can deliver 50 to 75 satellites at once. That cadence is the only reason a 1.2 million-satellite constellation is even on the table; on any prior launch system, the deployment timeline would stretch past the useful life of the hardware itself.

Deployment Timeline

The nine-year deployment window targets 10% of the constellation in the first three years, 50% by year six, and 100% by year nine.

It is likely that SpaceX will initially deploy Tesla's Dojo and AI5 chips for early space missions. Dojo 2 may ultimately be superseded by AI6, as current Dojo chips suffer from fundamental design architecture flaws. Both Dojo 2 and AI5 are intermediate products with significant design improvements still needed. In the initial two years — combined with immature HJT and not-yet-ready perovskite technology — SpaceX's AI DC launches will likely be experimental proof-of-concept missions only.

Tesla is restarting its Buffalo, New York factory for solar panel manufacturing. The critical milestone is 2028: the deadline Musk has set for both Tesla and SpaceX to each deliver 100 GW of annual solar cell production capacity. The two-year timeline is the shortest achievable window for both companies to stand up solar production and for Chinese partners to mature HJT and perovskite technology. Per our understanding, even with 100% acceleration in investment and labor effort, the fastest HJT can reach commercial maturity is approximately two years.

The ultimate goal is limitless energy capacity for AI data centers in space. The cost to build a 40 MW data center in space could be as low as $8.2 million, compared to $167 million on Earth — with the latter requiring 2–3 years of supply chain bottlenecks at minimum.

The AI Feedback Loop and Lunar Base

It may feel surreal to discuss these topics, but we believe strongly that they are not vapor — there is no foundational science bottleneck, only engineering grind.

Ultimately, Tesla and SpaceX should converge into a unified industrial ecosystem: SpaceX unlocks space solar for effectively unlimited energy supply, xAI unlocks truthful AI and the intelligence to plan and instruct digitally, Tesla Optimus humanoids execute physically, and Tesla AI chips supply the compute.

Until recently, the missing piece was semiconductor fabrication — for both logic and memory — at a moment when the industry is experiencing the largest undersupply of cutting-edge capacity in its history. Fabs are notoriously conservative in expanding capacity because of the industry's brutal cyclicality, while AI infrastructure builders are ultra-optimistic due to overwhelming demand. The unveiling of Tesla's Terafab plans in March 2026 closes that gap, at least in ambition: rather than wait for TSMC or Samsung to build out capacity at their own pace, Tesla is bringing fabrication in-house, applying the same first-principles, vertically-integrated playbook that reshaped automotive and launch. Tesla's existing relationship with Samsung — which locked in Tesla as a core customer for eight years, reportedly selling chips below cost of goods sold — gives the project a running start on process know-how and supply chain. Again, this is an engineering problem, not a science problem.

The deeper challenge is finding an alternative path to cutting-edge chips with 10x better performance-per-cost beyond traditional lithography. On this front, we have yet to see a successful startup demonstration, and Terafab does not by itself solve it.

If Tesla or a successor company can close the loop, we reach an age of abundance — an economy measured in output rather than traditional currency. xAI builds ever-better models to improve SpaceX, Optimus, the fabs, and the models themselves. SpaceX supports xAI with energy, launch capacity, and satellite platforms. Tesla provides AI chips and physical agents for xAI to execute in the real world. If this positive feedback loop closes and accelerates, AI can improve and scale autonomously. At that point, beyond perhaps rare elements, AI can drive production costs toward zero because energy and labor are no longer scarce for most goods — especially those required for human survival.

For the intermediate future, a lunar base is almost certainly necessary. Shipping finished satellites from Earth means fighting through a thick atmosphere and climbing out of a deep gravity well — both enormous energy costs that show up directly in the price per kilogram to orbit. The Moon has neither problem: no atmospheric drag, and one-sixth the surface gravity. Once humanoid robots can operate factories on the lunar surface and launch finished satellites via electromagnetic mass drivers — essentially giant railguns that accelerate payloads to orbital velocity without rockets — the Earth-launch model stops making economic sense for bulk constellation deployment.

SpaceX has accordingly shifted its near-term priorities away from landing humans on Mars and building a sustaining Martian city, toward supporting xAI's space data center buildout. That includes an early lunar base as proof of concept, using lunar regolith — the layer of fine dust and broken rock blanketing the Moon's surface, rich in silicon, aluminum, iron, and oxygen — to manufacture solar panels locally rather than ferrying them up from Earth.

Solar Cost Analysis

Before walking through the numbers, it's worth briefly recapping the two threads from Part 1 that this section pulls together: why space solar delivers roughly 10x more useful power per watt installed, and why the cell chemistry roadmap matters so much for getting there.

The 10x figure compounds two distinct gains. On the generation side, a panel in high Earth orbit produces about 5x as much energy as the same panel on the ground — partly because there is no day-night cycle to interrupt it, partly because there is no atmosphere absorbing or scattering incoming sunlight, and partly because orbital deployments can use multi-junction cell architectures that capture a broader slice of the solar spectrum than terrestrial single-junction silicon. On the system side, a further ~2x falls out of the fact that continuous orbital sunlight eliminates the need for batteries and overbuild capacity, both of which dominate the cost stack of any serious terrestrial solar installation. Multiply the two together and you get the roughly 10x advantage in effective cost per watt delivered.

The cell chemistry roadmap matters because today's space-grade panels are built on gallium arsenide (GaAs) — a material with excellent efficiency and radiation tolerance, but one that is fundamentally too expensive to scale to gigawatt-class constellations. GaAs panels can run upward of $50–100 per watt at space-grade specifications, which is fine for a handful of geostationary communications satellites but ruinous for a million-satellite buildout. The path forward, as outlined in Part 1, runs through heterojunction (HJT) silicon cells in the near term and perovskite-on-HJT tandem cells in the longer term. HJT delivers GaAs-class performance at roughly one-tenth the cost; perovskite tandems push efficiency higher still while remaining compatible with the ultra-thin, rollable form factors that orbital deployment demands. The catch is durability — perovskite cells have not yet been fully validated even under terrestrial conditions, let alone the high-radiation environment of space.

With those two pieces in mind, the cost trajectory looks like this:

China's current solar cell cost sits at roughly $0.25 per watt. Space deployment converts that into roughly 5x more delivered power per watt installed, and removing batteries and overbuild adds another ~2x in effective system-level savings. Stacked together, the same panels China is producing today would cost the equivalent of about $0.025 per watt on a delivered basis if you could simply lift them to orbit for free. That is the upper bound — the worst-case starting point assuming no progress in cell technology at all.

Existing panels, however, are too thick and too heavy to make orbital deployment economical even with Starship-class launch costs. This is where HJT and perovskite become essential rather than merely attractive. HJT alone can drive terrestrial costs down to roughly $0.06 per watt while enabling ultra-thin glass (UTG) substrates that are both lightweight and rollable — exactly the form factor needed for high-area, low-mass orbital arrays. Apply the same 10x space-deployment multiplier to mature HJT, and the effective delivered cost falls to about $0.006 per watt before launch. Perovskite-on-HJT tandems, once durability is solved, push the number lower still.

The annual production math makes the stakes concrete. To build out 100 GW of solar capacity per year — roughly the scale required to power the constellation described earlier — costs would range from $25 billion using today's Chinese cells, to around $6 billion with mature HJT, to roughly $600 million with mature perovskite tandems. The spread between the high and low ends is more than 40x, which is why the cell chemistry roadmap is not a side detail but the central economic lever of the whole program.

The conclusion is straightforward. Once HJT matures, or even once UTG substrates alone become production-ready, SpaceX can unlock effectively limitless power generation at a cost that no terrestrial energy source — whether nuclear, gas, or ground-based solar — can come close to matching, and without the regulatory, land-use, and grid-interconnect frictions that constrain every Earth-bound option.

Cooling: Radiating Heat into the Void

Cooling is one of the few areas where the conventional intuition about space — that it's cold, therefore easy to cool things in — gets the physics exactly backwards.

Space is indeed a near-perfect vacuum at temperatures approaching absolute zero. But "cold" in the everyday sense requires a medium to carry heat away, and a vacuum has none. On Earth, every server rack sits bathed in air, and air does an enormous amount of invisible thermal work: it absorbs heat from hot surfaces, rises, mixes, and is replaced by cooler air, spreading thermal load across the entire room before any active cooling system has to intervene. Remove the air and that mechanism vanishes entirely. A chip in orbit cannot lose heat by convection because there is nothing to convect into.

That does not make terrestrial cooling cheap, however — it just makes it possible by default. Data center operators on Earth pay a substantial premium for the privilege of having air around. Cooling systems must be sized for peak summer thermal loads, which typically means overbuilding capacity by around 40% relative to average demand. Power generation has to be overbuilt in turn, both to absorb the cooling overhead itself and to maintain headroom for generator maintenance and grid variability. The net effect is that a meaningful fraction of any hyperscale data center's capital and operating cost is spent moving heat from one place to another.

To put numbers on it: a modern 1 GW AI cluster servicing roughly 330,000 NVIDIA GB300 GPUs and 165,000 Grace CPUs draws about 627 MW for the compute silicon itself (GB300s at 1.4 kW each, Grace CPUs at 500 W each). The remaining ~373 MW — nearly 40% of total facility power — goes to cooling systems, power conversion losses, networking, storage, and the rest of the supporting infrastructure. That overhead is the cost of operating in an atmosphere.

In space, the physics inverts. With no air, cooling is not done by blowing fluid past hot surfaces but by radiative heat rejection — the same mechanism by which a hot stove glows or the Earth itself sheds heat to space at night. Every object warmer than absolute zero (0 kelvin, or −273°C — the theoretical floor of temperature, at which all atomic motion effectively stops) emits thermal radiation as infrared photons, and in a vacuum that radiation is the only way heat can leave. The advantage is the size of the temperature gap: a satellite radiator running at room temperature is shedding heat into a background of roughly 3 kelvin — just three degrees above absolute zero — and the larger that gap, the faster heat radiates away. For practical purposes, deep space is an infinite, near-perfect heat sink.

The catch is that radiation, even into an ideal sink, moves heat far less densely than convection does. A square meter of fan-cooled heatsink can shed several kilowatts because air is constantly flowing past it, physically carrying heat away by the kilogram. A square meter of radiator in orbit can only shed a few hundred watts, because photons are a much thinner heat-carrying medium than moving fluid. The cold background helps, but it cannot make up the gap entirely — so orbital cooling pays for itself in surface area instead. To shed the same wattage a fan-cooled rack handles in a few square meters, an orbital radiator might need tens or hundreds of square meters.

The architectural implication is elegant: each satellite is essentially a large flat panel, with its sun-facing side generating power and its shaded side radiating waste heat into the dark. Inside the satellite, liquid cooling loops circulate coolant through the dense AI silicon and carry the heat out to the radiator surface, where it leaves the spacecraft as infrared light. The same panel geometry that makes solar generation efficient in orbit also makes thermal rejection efficient — one of the quiet structural reasons the architecture works.

Whether orbital cooling is ultimately cheaper than its terrestrial counterpart depends on factors that are still being worked out: radiator mass per watt rejected, coolant loop reliability over multi-year missions, and the thermal behavior of high-density chip packages in microgravity. But the qualitative case is strong. Terrestrial data centers pay a 40% overhead for cooling and power conversion that they cannot avoid; orbital ones replace it with surface area — and because that surface area is the same flat panel already serving as the satellite's solar collector and structural chassis, it costs almost nothing extra to provide. Starship lifts it cheaply, and the satellite gets its radiator, its power source, and its skeleton in a single piece of hardware.

Launch Frequency

Elon Musk has predicted that within approximately five years (roughly 2030–2031), AI computing deployments in space will be launching at a rate of 300–500+ GW per year, with a theoretical ceiling of 1 TW per year before launch cadence and logistical constraints bind.

The rough arithmetic: 100 GW requires approximately 11,000 Starship launches per year — roughly one Starship launch per hour.

The rough arithmetic: 100 GW per year requires roughly 5,000–11,000 Starship launches annually — up to one launch per hour at peak cadence. From a launch-capacity perspective, this looks increasingly feasible given Starship’s reusability progress. The more immediate bottlenecks are likely solar panel production, radiation-hardened chips, and overall manufacturing scale.

On the fuel side, each Starship launch consumes ~1,000 tonnes of methane. At these volumes, natural gas and LOX demand would rise noticeably but would still represent only a low-single-digit percentage of annual U.S. production — easily absorbed by shale supply and SpaceX’s own on-site LNG capabilities. Even at a theoretical 1 TW annual deployment, the impact would be more material, though logistical and cadence constraints would likely bind well before fuel supply becomes a real issue.

Longer term, material suppliers for Starship structures and Raptor engines — including specialty alloys and certain rare earth elements — should see clearer, sustained tailwinds from the required vehicle production ramp.

Musk expects that using a lunar factory to build satellites and an electromagnetic mass driver to send them to orbit could ultimately deliver up to 1 PW of deployment capacity per year.

Tesla plans to start producing its next-generation AI5 chip in Q2 2027 and the more advanced AI6 chip in Q2 2028. AI6 is expected to serve as the core chip for the orbital Space Data Centers. Its timeline aligns with HJT maturation and the 100 GW capacity buildout.

AI6 will be Tesla’s first chip designed for both training and running AI models across all architectures — not just vision systems for cars and robots. However, for maximum efficiency when training the very largest models, Tesla will still need specialized training hardware. This is likely why Elon has restarted the Dojo project.

Earlier Dojo versions relied on highly custom designs (RISC-V, software-defined dataflow, extreme parallelization, and vertical DRAM) that proved difficult to scale. The new Dojo 3 will instead be built around large clusters of AI6 chips using a more standard NPU-style design. This improves compatibility with the industry-standard tensor-core approach used by NVIDIA GPUs, Google TPUs, and other leading AI hardware.

Communication: Architecture, Physics, and Technology

What It Is

Inter-satellite communication is fiber optics without the fiber — laser beams shot through vacuum between satellites. SpaceX calls these Optical Inter-Satellite Links (OISLs), and they are the nervous system of the entire Space DC architecture. Without them, each satellite is an isolated computer. With them, a million satellites become one coherent supercomputer.

Starlink already operates 9,000+ laser terminals in orbit — three per V2 Mini satellite — each running at up to 200 Gbps. This is already the largest coherent optical network ever built. For the Space DC to function — especially for training workloads requiring massive data synchronization — links must push 400 Gbps to 1+ Tbps between satellites. The FCC filing for 1.2 million satellites implies 4–8 laser terminals per satellite depending on orbit tier, link direction, and mesh density, putting the total terminal count in the range of 5–10 million units over the nine-year deployment window. That number shatters every existing aerospace optical supply chain and forces the photonics industry to restructure around semiconductor-class volume economics.

How It Works

The easiest way to picture an optical inter-satellite link is as a flashlight beam fired between two moving targets — except the flashlight is a laser, the targets are satellites a thousand kilometers apart, and the beam has to stay locked on a window the width of a coin while both ends are travelling at orbital velocity.

The process starts inside the transmitting satellite. A small laser diode produces a beam of infrared light at a wavelength of 1,550 nanometres — the same wavelength that carries internet traffic through undersea fiber-optic cables. The reason for the borrow is practical: decades of telecom investment have made the lasers, detectors, and amplifiers at this wavelength cheap, reliable, and well understood. Space hardware almost never gets to inherit a mature supply chain like this, and the OISL designers have wisely taken the gift.

Next, the data is written onto the beam. A component called a modulator rapidly varies the laser's brightness or phase — flickering it on and off, in effect — at fifty to a hundred billion times per second. Each flicker is a bit, and the pattern of flickers is the data stream. This is the same trick used in terrestrial fiber: light is just a faster, cleaner medium for the same on-off encoding that runs through every copper wire on Earth.

For shorter hops between neighboring satellites, the laser's natural output is bright enough. But for longer links — particularly the vertical hops between satellites in different orbital shells, which can span over a thousand kilometers — the beam needs a boost. An erbium-doped fiber amplifier (essentially a short length of specially treated fiber that pumps energy into the passing light) raises the signal from milliwatts to roughly a watt. Without that amplification, only a handful of photons would survive the journey, which is not enough to recover a clean signal at the other end.

The amplified beam then exits the satellite through a small telescope, typically 5 to 15 centimeters across. The telescope's job is to collimate the light — squeezing the diverging beam into a tight, near-parallel cone so that as little energy as possible is wasted spreading out into empty space. On the receiving satellite, an identical telescope works in reverse, gathering the arriving photons and focusing them onto a highly sensitive detector, which converts the light back into an electrical signal. From there, the data enters the satellite's onboard network and is handed off to the AI compute payload or relayed onward.

So far this is essentially a fiber-optic link with the fiber replaced by vacuum, and none of it is the hard part. The hard part is aim.

Two satellites in low Earth orbit can be closing on each other at 15 kilometers per second — roughly 20 times faster than a rifle bullet. The laser beam itself, despite being collimated, still spreads slightly with distance: about 20 microradians of divergence, which works out to a beam roughly 20 meters wide after a thousand kilometers of travel. The receiving telescope, by contrast, is a target less than a microradian wide from the transmitter's perspective — the angular equivalent of trying to land a laser pointer on a coin from several kilometers away, while both you and the coin are moving.

To make this work, each terminal uses a fast-steering mirror: a tiny, precisely actuated mirror that adjusts the beam's direction more than a thousand times per second, with sub-microradian accuracy. It is supported by predictive software that doesn't aim where the receiving satellite is, but where it will be a few milliseconds from now, accounting for orbital motion, vibration from the satellite's reaction wheels, and slight thermal warping of the structure as it moves between sunlight and shadow. This entire subsystem is called pointing, acquisition, and tracking (PAT), and it is the reason optical inter-satellite links remained a laboratory curiosity for forty years. The lasers were never the problem. Aiming them was. SpaceX's contribution was less a scientific breakthrough than an industrial one: building PAT systems cheaply enough, and in large enough numbers, to put thousands of them in orbit and have them actually work.

The Zero-Dispersion Advantage

There is one piece of physics that, more than any other, shapes how the optical supply chain for a space-based data center will look over the next decade. It's called zero dispersion: in a vacuum, all wavelengths of light travel at exactly the same speed.

To see why this matters, it helps to understand what goes wrong on Earth. When a laser beam carries data through a glass fiber-optic cable, the beam is not made up of a single pure wavelength but a small spread of them — a tight band of colors clustered around 1,550 nm. Glass slows different wavelengths by very slightly different amounts, an effect called chromatic dispersion. Over short distances this is negligible, but over hundreds of kilometers the slower wavelengths fall behind the faster ones, and the crisp on-off pulses the modulator wrote onto the beam start to smear into one another. By the time the signal reaches the far end, the bits have run together like wet ink, and the receiver can no longer cleanly tell a "1" from a "0."

Terrestrial telecom solves this with coherent detection: a sophisticated, power-hungry technique that uses high-speed digital signal processors to mathematically un-smear the signal at the receiving end. Coherent modules are the workhorses of long-haul fiber, but they are expensive — roughly $1,000 per unit — and they consume around 30 watts each.

In a vacuum, none of this happens. The refractive index of empty space is exactly 1 at every wavelength, by definition, so there is no smearing to undo. A clean pulse leaves the transmitter and arrives as a clean pulse at the receiver, no matter how far it has travelled. This means satellites can use a far simpler and cheaper class of transmitter known as an EML — an electro-absorption modulated laser, which combines the laser source and the modulator onto a single fingernail-sized chip of indium phosphide. An EML costs around $10, draws roughly 300 milliwatts, and pairs with a basic direct-detection receiver that simply measures whether light is arriving or not. No coherent DSP, no un-smearing, no $1,000 module.

This matters at constellation scale. The bulk of the SpaceX architecture is built around tight clusters of satellites flying in close formation, where neighboring satellites are typically less than 160 kilometers apart. At those distances, the cheap EML-and-direct-detection combination is more than sufficient, and the vast majority of links in the constellation — millions of them — can use it. The expensive coherent gear is reserved for the much smaller number of vertical links that span between different orbital shells, where distances stretch from 500 to 2,000 kilometers and signal levels drop to the point that the heavier hardware earns its keep.

The implication for the photonics industry is significant. A space-based data center is not just a bigger market for the same components that go into terrestrial fiber networks; it is a market that runs on different components entirely, weighted heavily toward the cheap end of the catalogue. The companies that win in this supply chain will be those that can manufacture EMLs at semiconductor-class volumes, not those that dominate today's coherent-module business.

Why Lasers Beat Radio

Before laser links matured, satellites talked to each other (and to the ground) using radio-frequency signals, typically in the Ka-band — the same general slice of spectrum used by satellite TV and some 5G networks. RF still has a role, especially for satellite-to-ground links where lasers struggle with clouds and atmospheric turbulence. But for satellite-to-satellite communication inside a dense constellation, optical links beat radio on essentially every dimension that matters.

The first advantage is raw bandwidth. A modern OISL terminal moves data at hundreds of gigabits per second and is on a clear path to a terabit per second; a Ka-band crosslink, by comparison, tops out one to two orders of magnitude lower. For a space-based data center training large AI models, where satellites must constantly synchronize enormous volumes of intermediate results, this difference is the gap between a workable system and one that grinds to a halt waiting for data.

The second advantage is security. A laser beam is roughly 20 meters wide after a thousand kilometers of travel — narrower than a city block — and aimed at a specific receiver. To intercept it, an adversary would need a telescope physically positioned within that narrow cone, which in practice means flying their own spacecraft into the beam path. Radio signals, by contrast, spread across thousands of kilometers and can be picked up by any antenna in the footprint. Optical links are, for practical purposes, both unjammable and undetectable from the ground.

The third advantage is regulatory. Radio spectrum is a finite, licensed, and increasingly congested resource — every Ka-band frequency has to be coordinated with every other operator using it, and the available bands are filling up fast as more constellations launch. Light is unregulated. A satellite can fire as many laser links as it has terminals for, at any wavelength its hardware supports, without filing a single spectrum application.

Finally, there is latency, which matters surprisingly much for distributed AI. Light travels about 47% faster in vacuum than through fiber. A 160-kilometer hop between two satellites takes around 0.5 milliseconds; the same distance through fiber takes about 0.8 milliseconds. The 0.3-millisecond saving sounds trivial, but when training a large model requires thousands of synchronizing exchanges per second across the constellation, those microseconds compound into meaningful throughput. In a sense, space is not just a place to put compute — it is a faster medium for moving data between compute nodes than anything available on Earth.

This is marketing material for NewDeal Invest. The investment style is aimed at investors with a high risk profile and a long-term investment horizon, and past performance is no guarantee of future returns. The stocks that NewDeal Invest invests in are more volatile than the broader stock market, and therefore fluctuate more both up and down. NewDeal Invest does not provide investment advice, and all investment decisions are made by you yourself. No information from NewDeal Invest should be considered a recommendation. Short-term investments, with a time horizon of 5 years or less, are discouraged both in NewDeal Invest, PMINDI, and in the stocks mentioned. See relevant risk factors at https://newdealinvest.dk/ and https://newdealinvest.dk/pmindi/