SpaceX - Powering AI from Orbit, Space Solar, and the Musk Stack (Pt.3)

SpaceX - Powering AI from Orbit, Space Solar, and the Musk Stack (Pt.3)

The orbital AI data center as tight satellite clusters: redefining photonic demand and unlocking compute beyond Earth’s physical limits.

Summary

• The Space DC is not a diffuse mesh of satellites spread around Earth — it is a constellation of tight orbital clusters, each functioning as a single coherent data center with nodes flying 20–160 km apart.

• This tight-cluster geometry means 60–70% of all terminals by unit count are short-range, low-power, cheap EML-based links, radically favoring direct detection over expensive coherent modules.

  
  

• The Space DC is not purely additive demand for the photonic supply chain — it partly substitutes terrestrial transceiver volume on the growth margin, though three genuinely net-new categories emerge (PAT systems, free-space optics, inter-orbit links) that have no terrestrial equivalent.

  
  

• At 800G per lane (6.4T–12.8T links), InP becomes a hard design requirement on power-constrained satellites because silicon CMOS hits its physical ceiling with no performance margin — making MACOM's core competence non-substitutable.

  
  

• MACOM's content per terminal scales superlinearly with bandwidth, and freed InP wafer capacity from the terrestrial CPO transition arrives just as the Space DC demands it — making MACOM perhaps the clearest single-company beneficiary of the shift.

Architecture Tiers, Cluster Geometry, and Link Types

Part 2 established the building blocks: a three-tier constellation — VLEO for inference, LEO for training, MEO for storage — wired together by laser links into a single distributed supercomputer, powered by panels whose sun-facing side generates electricity while their shaded side radiates waste heat into the void. Starship makes the launch math tractable, HJT and perovskite chemistry collapse the solar cost stack, and vacuum's zero dispersion rewrites the photonics supply chain by letting most links run on $10 EMLs rather than $1,000 coherent modules. Those are the pieces. Part 3 is about how they physically arrange themselves in orbit — and the arrangement is what most observers get wrong.

The fundamental architectural insight that most analysts miss: the Space DC is not a diffuse mesh of satellites spread evenly around Earth. It is a constellation of tight orbital clusters — each cluster functioning as a coherent data center in space. Satellites within each cluster fly in close formation, typically within a 100 × 100 mile (160 × 160 km) box, because distributed compute workloads — whether training or inference — demand the same high-bandwidth, low-latency interconnect fabric that a terrestrial data center achieves by placing racks close together. You would not build a GPU training cluster with nodes scattered across a continent, and you would not build an orbital training cluster with satellites scattered across an orbital shell that is Earth.

This tight-cluster architecture means that intra-orbit inter-satellite distances within all three tiers are short — typically 20–160 km between adjacent satellites within the same cluster. There are no 2,000–5,000 km intra-orbit backbone links in this architecture. The longest links in the system are the vertical cross-tier links connecting clusters at different altitudes: VLEO to LEO (~500 km), LEO to MEO (~1,000 km), and VLEO to MEO (~1,500 km), with slant angles pushing actual path lengths to 500–2,000 km depending on geometry.

This geometry has profound implications for the optical supply chain. The vast majority of all inter-satellite terminals by unit count are short-range, low-power, direct-detection links — radically favoring cheap EML-based transmitters and integrated silicon photonic receivers. The expensive, amplified, coherent-detection links that dominate the unit economics of traditional space laser communication are confined to a minority of terminals: the vertical cross-tier links.

VLEO — 500 km Altitude, 800,000 Satellites

Edge compute and real-time inference. This is the lowest tier and the only one that communicates directly with users on the ground, while also linking upward to the rest of the orbital compute fabric. That dual role — serving terrestrial users and coordinating with higher tiers — makes its communication requirements the most complex.

Earth-facing links (primary priority)

VLEO satellites are the customer-facing interface of the Space DC. They receive user queries from ground terminals — phones, base stations, enterprise dishes — and return inference results. This is the revenue-generating layer of the entire architecture. These links are predominantly RF (Ka-band and V-band phased arrays, the same technology Starlink already uses for broadband service) because atmospheric turbulence, clouds, and rain severely degrade optical links to the ground. RF remains the primary user-facing technology, though SpaceX has been testing optical ground links that may supplement RF for bulk data movement. Each VLEO satellite carries 1–2 RF phased array panels for user links and potentially one optical downlink terminal for backbone ground connectivity — high-bandwidth links to major fixed ground stations deliberately sited in dry, clear-sky locations (deserts, high-altitude plateaus) where weather rarely disrupts optical transmission. These stations act as on-ramps between the orbital network and terrestrial fibre infrastructure.

Vertical links to higher orbits (second priority)

A VLEO inference satellite performing a user query does not operate in isolation. It may need to fetch updated model weights from the LEO training tier at 1,000 km, pull reference data from the MEO storage tier at 2,000 km, or offload sub-tasks to more powerful compute nodes above. For large models — especially those using Mixture of Experts (MoE) architectures — the VLEO satellite might serve as a routing and assembly layer, receiving partial inference results from higher-tier satellites and composing the final response for the user. These vertical links span 500–1,500 km depending on the target tier and slant angle, making them the longest links that VLEO satellites maintain. They require meaningfully more transmit power than intra-cluster links, a larger-aperture telescope, and potentially a small EDFA (Erbium-Doped Fibre Amplifier - boosts optical signal strength without converting to electrical) or semiconductor optical amplifier to close the link at 400G. Each VLEO satellite likely needs 1–2 terminals dedicated to vertical links. These are the most expensive terminals on a VLEO satellite and the ones where the PAT (Pointing, Acquisition, and Tracking - the system that aims and locks the laser beam onto the target satellite) challenge is most acute, because the target satellites at higher orbits move at different orbital velocities.

Intra-orbit links within the VLEO cluster (third priority).

In a distributed MoE inference architecture, a single user query may activate different expert sub-networks residing on different VLEO satellites within the same cluster. The inference result must be assembled from partial computations across multiple co-orbital satellites. This requires high-bandwidth, low-latency intra-cluster communication — but the distances are short (20–80 km between adjacent satellites), relative velocities between co-orbital satellites are near zero, and the PAT problem is benign because the target is almost stationary in the satellite's reference frame. A few milliwatts of transmit power closes a 100G link trivially. No amplification needed, no complex optics. A basic EML or directly modulated laser with a small-aperture telescope handles this. Each VLEO satellite needs at least 2 intra-cluster terminals, and these are the cheapest terminals in the entire constellation.

The net result is that a VLEO satellite carries 5–7 communication terminals total: 1–2 earth-facing RF panels (primary revenue link), 1–2 vertical optical links to higher tiers (highest-cost optical terminals), and 2–3 ultra-cheap intra-cluster optical links (highest-volume terminal type). The mix of technologies and cost points across these terminals makes the VLEO supply chain analysis particularly complex.

LEO — 1,000 km Altitude, 300,000 Satellites

Regional compute and model training. This is where the heavy compute resides — training clusters that must synchronize gradients across hundreds or thousands of nodes. The key architectural point: a LEO training cluster is a tightly packed formation of satellites functioning as a single distributed GPU cluster. Intra-cluster spacing is 30–100 km. At these distances, the inter-satellite communication physics are identical to VLEO intra-cluster links: short-range, low-power, direct-detection, EML-based. But the bandwidth demands per link are far higher because gradient synchronization in distributed training requires all-to-all communication patterns. Each satellite needs 4–6 intra-cluster terminals running at 400G–1.6T aggregate per link to sustain training throughput. The cost optimization here is ferocious — every dollar of photonic content per terminal multiplies across millions of terminals.

LEO clusters also maintain vertical links: downward to the VLEO edge tier (500 km path) and upward to the MEO storage tier (1,000 km path). These vertical links move training data, model checkpoints, and weight updates between tiers. Each LEO satellite carries 1–2 vertical terminals. These are the more expensive terminals per unit but represent a minority of total terminal count.

MEO — 2,000 km Altitude, 118,000 Satellites

Cloud compute and data storage. These are the heavyweight nodes at 20,000 TOPS per satellite, serving as the persistent storage and high-compute backbone. MEO clusters are also tightly packed — 40–160 km intra-cluster spacing — with intra-cluster links using the same short-range EML-based architecture as other tiers. Each MEO satellite carries 4–6 intra-cluster terminals.

MEO's vertical links connect downward to LEO training clusters (1,000 km) and to VLEO edge nodes (1,500 km). These are the longest links in the entire constellation. At 500–2,000 km path length with slant angles, the link budget tightens significantly: an EDFA must boost the transmit signal, a larger-aperture telescope (10–15 cm versus 5 cm for intra-cluster) is needed, and a high-sensitivity InGaAs APD (Indium Gallium Arsenide Avalanche Photodiode) on the receive end is required — a detector made from a semiconductor material sensitive to the infrared wavelengths used in optical comms, which internally multiplies weak incoming photons into a stronger electrical signal through an avalanche effect. This sensitivity is essential because over these longer distances, the laser signal arrives far weaker than over short intra-cluster hops. Each MEO satellite carries 1–2 vertical terminals, and these are by far the most expensive terminals per unit in the entire system — potentially $5,000–15,000 each versus $50–200 for an intra-cluster terminal.

Summary of Link Types and Distances

• Intra-cluster links across all three tiers span 20–160 km. Short-range, low-power, EML + direct detection. Highest volume, lowest cost per terminal. Approximately 60–70% of all terminals by unit count.

  
  

• Vertical cross-tier links span 500–2,000 km depending on tier pair and slant angle. Higher power, potentially amplified, larger aperture. Most expensive per unit. Approximately 15–25% of all terminals by count.

  
  

• Earth-facing links (VLEO only) use RF phased arrays plus selective optical ground links. An entirely different technology stack. Approximately 10–15% of all terminals by count.

SUPPLY CHAIN FRAMEWORK

The Volume Shift Framework

The most common analytical error in sizing the Space DC optical supply chain is treating it as purely additive — a new market layered on top of existing terrestrial demand. This understates the substitution effect. But the opposite error — treating Space DCs as a wholesale replacement of terrestrial infrastructure — overstates it. The reality is more nuanced.

Existing terrestrial data centers represent trillions in sunk capital: buildings, power infrastructure, cooling systems, fibre connectivity. They will continue operating through their useful lifespans, and certain workloads — latency-sensitive serving, regulated data with sovereignty requirements, anything needing frequent physical maintenance — will stay grounded indefinitely. Nobody is decommissioning a functioning hyperscale facility because orbital compute exists.

Where the substitution effect bites is on the growth margin. Hyperscaler IT load is projected to surge roughly 6x by 2035, requiring over 100 GW of new capacity. That new capacity faces increasingly severe terrestrial constraints: grid connection queues, permitting battles, water scarcity for cooling, and land availability near population centres. Space DCs offer a relief valve — abundant solar power, no grid queues, no permitting, no water — and as Starship drives launch costs down to projected levels, they become the path of least resistance for a significant share of new AI compute builds. GPUs that would have been racked in new facilities in Texas or Iowa will instead sit on satellites. Optical connections that would have linked those GPUs through fiber-optic transceivers and patch cables in new builds will instead link them through free-space laser terminals across 20–160 km of vacuum.

For the core photonic components — the lasers, modulators, detectors, drivers, and TIAs that form the active elements of any optical link — the implications depend on what share of new capacity migrates to orbit. If Space DCs capture even 30–50% of incremental builds over the next decade, the component volumes involved are substantial — potentially millions of optical terminals displacing millions of terrestrial transceivers that would otherwise have been deployed. The form factor changes. The packaging changes. The qualification requirements change. But the silicon and III-V die count for that migrated share does not change dramatically.

This means that for companies like Lumentum (EMLs) and Tower Semiconductor (SiPh PICs), the Space DC is not purely a windfall of net new volume — it is partly a migration of future demand from one domain to another. Total photonic demand still grows with overall market expansion, and existing Earth DCs will keep buying upgrades, so the picture is not zero-sum. But the investment thesis for these companies rests less on volume expansion and more on three secondary effects: pricing premiums from space qualification (perhaps 2–3x terrestrial pricing per component in the early years, compressing to 1.3–1.5x at scale), changes in the mix of component types required (more direct-detection, fewer coherent modules, different driver specifications), and the genuinely new demand categories that exist in space but not on Earth.

Three Genuinely New Demand Categories

The previous section argued that much of the Space DC's optical component demand is substitutive — it displaces terrestrial transceiver volume rather than adding to it. But three elements of the Space DC have no equivalent in terrestrial data centers. These represent genuinely net new demand for the photonic supply chain — volume that would not exist in any scenario where compute stayed on the ground.

Pointing, acquisition, and tracking (PAT) systems.

In a terrestrial data center, the optical connection between two GPUs is a fiber plugged into a port. Alignment is mechanical and static — push in the connector, it clicks. In space, the optical connection is a laser beam that must be continuously steered with sub-microradian precision onto a moving target. Every single inter-satellite link needs a PAT subsystem: a fast-steering mirror, a quad-cell detector, a control ASIC, and associated power electronics. None of this exists in the terrestrial architecture. It is pure net new demand, multiplied across every terminal on every satellite. This is where STMicroelectronics captures genuinely incremental content.

Free-space optics hardware

Every terminal needs a telescope (3–5 precision optical elements), anti-reflection coatings, bandpass filters, and a beam-forming assembly. In a terrestrial data center, the equivalent is a $0.50 fiber ferrule. In space, it is a $50–500 precision optical assembly depending on the terminal type. This is where Coherent captures genuinely incremental content — their optics and coating capabilities serve a demand category that simply does not exist on Earth.

Inter-orbit communication links

In a terrestrial data center, all compute nodes are co-located — same building, same campus, at most same metro area connected by dark fiber. The "tiers" of the architecture (edge, training, storage) sit next to each other on the same LAN. In the Space DC, these tiers are at different orbital altitudes. VLEO inference nodes at 500 km must communicate with LEO training clusters at 1,000 km and MEO storage at 2,000 km. This inter-tier traffic has no terrestrial equivalent. It generates genuinely incremental terminal demand and creates additional relay capacity requirements on Starlink satellites that serve as the inter-orbit routing backbone.

One Major Category Disappears: Fiber

In the terrestrial architecture, Corning and Prysmian supply hundreds of millions of meters of single-mode fiber, plus connectors, patch panels, and cable assemblies, for every large data center. In space, the propagation medium is vacuum — free and infinite. Corning's data center fiber business faces a structural headwind if a meaningful fraction of hyperscale compute migrates to orbit.

One Category Shrinks Significantly: Coherent DSP Complexity

On Earth, chromatic dispersion in fiber forces data center interconnects to use increasingly complex coherent detection with digital signal processing — Marvell and Broadcom sell DSP ASICs at $50–200 per transceiver for this purpose. In vacuum, dispersion is zero. Direct detection works at any distance. The coherent DSP market loses volume as compute migrates to space. This is a secondary headwind for Marvell's and Broadcom's optical DSP businesses, partially offset by potential entry into the space terminal market through other product lines.

Why 6.4T–12.8T Per Link Changes Everything

If the Space DC is to match terrestrial training cluster performance — and it must, because this is the core value proposition — inter-satellite links within a training cluster must deliver equivalent bandwidth. A VLEO or LEO satellite in a training cluster running at 20,000 TOPS needs the same interconnect bandwidth as a terrestrial GPU node at equivalent compute density. Whether the optical link traverses 2 meters of fiber in a rack or 80 km of vacuum between satellites is irrelevant to the compute layer above — the bandwidth requirement is identical.

This means the intra-cluster inter-satellite link roadmap tracks the terrestrial data center interconnect roadmap: 1.6T in early deployment (2028–2029), 6.4T in the main buildout phase (2030–2032), and 12.8T in the mature phase (2033+). Each generation roughly quadruples per-link bandwidth through a combination of more lanes per terminal (4 → 8 → 16) and higher per-lane data rates (100G → 200G → 200G+).

At 100G per lane (50 Gbaud PAM4), the key analog components — laser drivers and transimpedance amplifiers (TIAs) — must switch fast enough and maintain sufficient linearity to cleanly resolve four amplitude levels at high speed (hence the PAM4 name). Traditionally, these circuits require specialty materials like InP or SiGe BiCMOS because their transistors are inherently faster and cleaner at analog tasks. But advanced silicon CMOS at 5nm or 3nm nodes is now fast enough to handle 100G per lane. That matters because it opens the door to silicon photonics (SiPho) PICs with integrated silicon drivers — a single chip combining optical and electrical functions in cheap, scalable CMOS fabrication rather than expensive specialty processes.

For short intra-cluster links (tens of kilometers between satellites in the same training cluster), this silicon-native approach likely suffices. MACOM's InP technology offers superior analog performance — faster switching, better linearity, lower noise — but if silicon CMOS meets the spec at 100G per lane, InP becomes a premium option rather than a hard requirement. The investment implication is pointed: the highest-volume segment of the Space DC interconnect market may not need MACOM's core differentiator.

At 800G per lane (400 Gbaud PAM4), the physics change qualitatively. The laser driver must produce voltage swings of 2+ volts across a 25-ohm load — large enough to create four cleanly separated amplitude levels — while doing so at analog bandwidths exceeding 80 GHz. That combination of high voltage swing and extreme speed is punishing: you're asking a transistor to push significant current very quickly into a resistive load, and any nonlinearity smears the four PAM4 levels together, making them unreadable at the receiver. On the receive side, the TIA faces a mirror challenge — amplifying a faint photocurrent at equivalent bandwidth while keeping the four signal levels evenly spaced rather than compressing or distorting them.

Silicon CMOS transistors at these frequencies are operating near the ceiling of their transit frequency (the maximum speed at which a transistor can usefully amplify). They can technically function, but with poor power efficiency and almost no performance margin. InP HBTs (heterojunction bipolar transistors — a faster transistor architecture built on indium phosphide) offer intrinsic transit frequencies above 400 GHz, giving them 3–5x the speed headroom over silicon at these data rates. That translates directly into either half the power consumption or twice the signal quality at the same bandwidth. On a power-constrained satellite where every watt competes with compute for the solar power budget, InP's efficiency advantage is not a nice-to-have — it is a design requirement.

At 6.4T per link with 8 lanes of 800G, each terminal needs 8 MACOM InP driver ICs and 8 MACOM InP TIAs — 16 analog ICs per terminal at perhaps $8–20 per IC, yielding MACOM content of $130–320 per terminal. At 12.8T per link with 16 lanes of 800G, that doubles to 32 analog ICs per terminal and $260–640 of MACOM content.

This is where the earlier point about silicon sufficiency at 1.6T becomes important context. In the first deployment phase (1.6T links at 100G per lane), MACOM's InP is a premium option — silicon CMOS likely wins on cost and integration, and MACOM's content per terminal is modest. But the investment thesis isn't about the first generation. It's about what happens when the network scales to 6.4T and 12.8T, where 800G per lane makes InP a hard requirement on power-constrained satellites.

Compare this to MACOM's content in a current-generation 800G terrestrial transceiver: roughly $15–40 per module. The Space DC terminal at 6.4T carries 5–10x more MACOM content, and at 12.8T potentially 10–20x. Even if total terminal count is roughly volume-neutral versus terrestrial deployment — because orbital compute shifts capacity rather than adding to it — MACOM's dollar content per terminal increases enormously with each bandwidth generation. MACOM is the one company for which the volume-shift framework does not diminish the thesis. It strengthens it, because content per link scales superlinearly with bandwidth as physics pushes the design toward InP.

This parallels MACOM's trajectory in terrestrial data centers. As the industry moved from 100G to 400G to 800G to 1.6T transceivers, MACOM's revenue per module increased at every generation because higher speeds demand more sophisticated analog. The same dynamic plays out in space — arguably more forcefully, because satellite power constraints make InP's efficiency advantage a harder requirement than it ever was on Earth, where designers could afford to brute-force solutions in inefficient silicon.

The terrestrial supply picture actually reinforces this. Today, InP wafer capacity is a genuine bottleneck — and it is precisely this scarcity that is pushing the terrestrial industry toward SiPho pluggables and co-packaged optics (CPO) as alternatives. Crucially, this shift does not eliminate MACOM's terrestrial relevance — it reshuffles the product mix. CPO architectures absorb the laser driver function into silicon, reducing MACOM's driver content per link. But they still require high-performance TIAs on the receive side, and because CPO removes the pluggable connector bottleneck, it enables more total bandwidth per switch ASIC — meaning more lanes, more TIAs, and likely higher total MACOM content per switch package than the pluggable architecture it replaces. The terrestrial story is fewer drivers but more TIAs at higher aggregate volume — a mix shift, not a revenue decline.

Meanwhile, as terrestrial demand migrates away from discrete InP driver ICs, wafer capacity gradually frees up — capacity that can be redirected toward space terminals just as the Space DC buildout scales to 6.4T and 12.8T. MACOM ends up in an unusual position: its terrestrial business evolves with sustained or growing TIA content per switch, while its space business scales into the freed driver capacity with higher content per terminal and a captive customer base that cannot substitute silicon regardless of price.

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/