This is the forward-looking companion to the live Earth map. The cool dots and the named astronomy platforms are exactly what is in orbit right now — the same live Celestrak catalog, SGP4-propagated in your browser. The warm haze is what one filing would add: SpaceX's January 2026 FCC application for a constellation of up to one million orbital data-center satellites, here rendered as a representative realization consistent with the filing's stated envelope (500–2,000 km altitude; 30° and sun-synchronous inclinations; orbital shells up to 50 km thick). It is not a forecast that this will be built — it is a faithful picture of what has been asked for, drawn to the same scale as reality so the gap is legible. The deploy slider scales the projected layer continuously; the readouts on the right are recomputed from the deployed count using closed-form geometry, with every figure sourced below.
What is actually being proposed
The projected layer is one specific filing, not a generic "more satellites" scenario. Three things are routinely conflated in coverage, and only the middle one is what this map draws:
1 · The chips are made on Earth. SpaceX, Tesla and xAI announced TERAFAB, a chip-fabrication venture in Austin targeting roughly one terawatt of compute per year, with the stated intent to direct the large majority of that output "to space." That is a terrestrial factory.
2 · The data centers go in low Earth orbit. In January 2026 SpaceX filed with the U.S. Federal Communications Commission for a constellation of up to one million "orbital data center" satellites — compute hardware powered by continuous sunlight and cooled by radiating heat to space, linked by optical inter-satellite laser and relayed to the ground via Starlink. The filing's stated operating envelope is altitudes of 500–2,000 km, inclinations of 30° and sun-synchronous, in orbital shells up to 50 km thick. This is the layer rendered here.
3 · The Moon is a later, far more speculative factory-and-launcher. The lunar mass driver and an in-situ satellite factory are a separate, aspirational stage — no hardware, mid-2030s at the earliest. The satellites would still operate in Earth orbit; the Moon would only be a cheaper place to build and fling them.
SpaceX has not published per-shell satellite counts. The realization on this map distributes the one-million ceiling across eleven representative shells inside the filed envelope (weighted toward the sun-synchronous and low-inclination bands a continuously-solar-powered compute constellation would favor). The geometry, density, and visibility numbers below are exact for that realization; the true distribution, if it is ever built, will differ in detail but not in order of magnitude.
Projected impact · optical astronomy
The harm to ground-based astronomy does not require the full million to be a problem, but the full million makes it categorical. All figures below are at the one-million ceiling; the live panel on the map recomputes them for any deployed count.
~52,000
Above the horizon at any instant exact geometry
Summed over all shells, the count geometrically above a typical observer's local horizon at any moment — each visible within the spherical cap of half-angle arccos(R⊕/r).
cap fraction (1−R⊕/r)/2 per shell · today the entire active catalog above the horizon is a few hundred
~720,000
Sunlit at any instant · 72% exact geometry
Fraction outside Earth's cylindrical shadow. Most of the constellation is lit at all times; the question is only how much is also above your horizon and how dark your sky is.
sunlit fraction (1+√(1−(R⊕/r)²))/2 per shell
~18
Trails per LSST exposure, twilight scaling estimate
~52,000 above the horizon spread over the visible hemisphere is 2.5 sources/deg². A single 9.6-deg² Rubin/LSST field then contains ~24 satellites, ~18 of them sunlit at twilight — each a saturated streak.
cf. Hainaut & Williams 2020; Bassa+ 2022; Tyson+ 2020 (streak mitigation saturates well below this density)
~202,000
In shells ≥1,000 km — lit much of the night exact geometry
High shells stay in sunlight through local midnight at mid-latitudes for much of the year, so their contribution to sky brightness and trails is not confined to twilight. This is the regime Lawler+ 2022 flagged as the worst case.
sunlit-through-midnight altitude threshold; cf. Lawler, Boley & Rein 2022
Two structural points. First, the diffuse floor: even the satellites too faint or too numerous to extract as individual trails add an unresolved scattered-light haze to the whole sky, raising the background against which every faint source is measured (Kocifaj+ 2021). Second, mitigation does not scale to this regime — darkening coatings, orbit-aware scheduling and streak-masking were designed for tens of thousands of objects; at a density where a wide field holds ~20 trails per exposure, masking removes most of the image. A constellation like this does not degrade ground-based optical astronomy at the margin; for twilight survey science and for the high-altitude shells, much of the night, it forecloses it.
Projected impact · collisions & debris
The optical harm is recoverable in principle — deorbit the satellites and the sky clears. The collisional harm is not, because fragments outlive the constellation that made them.
5.3×10⁻⁶
Peak spatial density · sats/km³ @ 600 km exact geometry
The busiest 50-km shell is 600 km, which carries both the 30° and the sun-synchronous allocations — ~293,000 satellites in one physical volume. Spatial density is count / (4πr²·Δr).
600 km sun-synchronous band
~58×
Denser than today's busiest shell scaling estimate
Today's peak — the ~550 km Starlink band — sits near 1.7×10⁻⁷ sats/km³ for the operational population. The filed constellation is roughly thirty times denser at its peak.
comparator: ~5,000 active payloads in a 50-km shell at 550 km; cf. ESA Space Environment Report
~1,000×
Pairwise conjunction-rate index scaling estimate
Catastrophic-collision rate scales with the square of spatial density (Kessler & Cour-Palais 1978). A shell ~58× denser carries of order 3×10³× the intrinsic conjunction rate before any avoidance manoeuvres.
rate ∝ ρ²·vrel·σ·V — order-of-magnitude
runaway
Kessler-cascade regime scaling estimate
Above a critical density, collision fragments are produced faster than drag removes them and the debris population grows on its own. These densities sit well into that regime for the higher, slow-decaying shells.
cf. Liou & Johnson 2006 (instability of the LEO debris population)
Two amplifiers specific to this filing. The high shells (1,000–2,000 km) sit above where atmospheric drag clears debris on a useful timescale — a fragment cloud there persists for centuries, not years, so a single cascade is effectively permanent. And the operational reality is already visible at 1% of this scale: today's ~10,700 Starlink satellites perform on the order of 145,000 automated collision-avoidance manoeuvres every six months — more than 300,000 across 2025. The maneuver burden, the conjunction-screening load, and the fragment risk all scale super-linearly from there.
What this projection is — and isn't
Read this before quoting the numbers
It is a filing, not a forecast. One million is the ceiling SpaceX asked the FCC to authorize, not a committed build. Treat the full-deploy view as "what has been requested," and use the slider to see any partial scenario. The harm is already categorical at a few hundred thousand.
The shell breakdown is RAVEN's representative realization, not SpaceX's design — the company has not released per-shell counts. The total, the altitude band, and the inclinations match the filed envelope; the allocation across shells is ours and is labeled as such.
Above-horizon and sunlit counts are exact closed-form geometry for the deployed realization. Density and conjunction figures are order-of-magnitude scaling estimates against today's busiest shell, tagged accordingly on every card. They are meant to convey scale, not to substitute for a full conjunction-rate simulation.
Brightness is unmodeled. Whether these satellites are naked-eye depends on their size and surface — unknown for the data-center design. The "above horizon & sunlit" count is therefore an upper bound on what could be visible, not a prediction of apparent magnitude.
Live baseline · the real platforms inside the swarm
The cool points and named markers are the current Celestrak catalog, propagated by SGP4 exactly as in the live Earth map — including the astronomical observatories below. They are here to set the scale: the working space telescopes humanity actually operates, shown inside the projected haze. Click a marker to fly the camera to its current position.
Lagrange-Point Astronomical Platforms
Catalog only — SGP4 propagation isn't valid at the Sun–Earth Lagrange points (1.5 million kilometres from Earth, dominated by Solar rather than Earth gravity), so these aren't rendered in the live map. Their state vectors come from JPL Horizons / mission DSN tracking.
How the Positions Are Computed
Live TLE pipeline
On page load, the browser fetches the current Celestrak GP catalog (group active) via the public HTTPS endpoint. Celestrak refreshes TLEs every two hours, so each visit gets data within that window. The TLE block is parsed by satellite.js into per-satellite SGP4 state structs (satrec). On every animation frame, each satrec is propagated forward by the elapsed simulated time and the resulting TEME-frame position is written into a single packed Float32Array driving a THREE.Points draw call.
SGP4
SGP4 is the analytic propagator that NORAD designed for TLEs, accounting for J2-J4 zonal harmonics, atmospheric drag, and lunisolar perturbations in a closed-form expansion. It's accurate to roughly a kilometre over a day for LEO targets, falling off for high-eccentricity or very-high-altitude orbits. We use it for every satellite in the catalog including GEO (where it's still fine) but exclude the L1/L2 platforms whose orbits are dominated by Solar gravity (SGP4 wasn't designed for that regime).
Earth rotation
Earth's body-fixed frame is rotated about its spin axis (= scene +Z) by the Greenwich Mean Sidereal Time of the current simulated instant — the standard GMST = 280.46061837 + 360.98564736629·(JD − 2451545.0) formula. SGP4 returns positions in TEME (an inertial-ish frame that differs from GCRF by sub-arc-second angles, negligible at rendering scale), so satellites and Earth surface line up correctly.
Classification
After parsing TLEs, each satellite is classified by its mean motion (rev/day): >11.25 = LEO, 1.5–11.25 = MEO/HEO, <1.5 = GEO. Note the bins are by mean motion, not orbit family: everything above ~26,000 km altitude falls in the GEO bin, including graveyard and high science orbits, while Molniya-class HEO (mm ≈ 2) files under MEO. Starlink is broken out separately (name match) because at ~7,000 entries it dominates the shell. Layer toggles let you isolate any class.
Projected layer — Keplerian propagation on the GPU
The filed constellation cannot use SGP4: there are no TLEs for satellites that do not exist. Instead each projected satellite is given circular two-body elements — semi-major axis a = R⊕ + altitude, inclination i, right ascension of the ascending node Ω, and an initial argument of latitude u₀ — drawn to populate the shells of the realization. Position is the closed-form circular orbit x = r(cosΩ·cos u − sinΩ·sin u·cos i), y = r(sinΩ·cos u + cosΩ·sin u·cos i), z = r·sin u·sin i, with u = u₀ + n·t and mean motion n = √(μ/a³). This expression is evaluated per vertex on the GPU in a custom shader from a single time uniform, so the full million points propagate every frame without touching the CPU. Sun-synchronous inclinations are set from the standard nodal-precession condition cos i = −(a/12352 km)³·⁵; J2 nodal drift itself is omitted (negligible over the visualization timescale). The projected points live in the same inertial scene as the SGP4 baseline, so Earth rotates correctly beneath both.
Impact readouts
Above-horizon and sunlit counts are summed analytically over the shells from the deployed count, using the spherical-cap fraction (1−R⊕/r)/2 and the cylindrical-shadow sunlit fraction (1+√(1−(R⊕/r)²))/2. Trail density per Rubin/LSST field follows from the above-horizon areal density on the visible hemisphere times the 9.6-deg² field. Spatial density is count/(4πr²·Δr) with Δr = 50 km, computed on physical shells so that co-altitude allocations at different inclinations are summed rather than compared; the conjunction-rate index applies the Kessler ρ²-scaling against today's busiest operational shell. Exact-geometry quantities and order-of-magnitude scaling quantities are tagged separately throughout.
References
- Hoots, F. R. & Roehrich, R. L. 1980 — Spacetrack Report #3: Models for propagation of NORAD element sets (canonical SGP4 reference)
- Vallado, D. A. et al. 2006 — Revisiting Spacetrack Report #3, AIAA 2006-6753 (modern reformulation)
- Celestrak GP catalog — live TLE data, refreshed every 2 hours
- satellite.js — battle-tested SGP4 implementation
- Petit, G. & Luzum, B. (eds.) 2010 — IERS Conventions (2010), Tech. Note 36 — GMST & reference-frame definitions
- Weisskopf, M. C. et al. 2002 — Chandra X-ray Observatory overview, SPIE 4012
- Gehrels, N. et al. 2004 — Swift gamma-ray burst mission, ApJ 611
- Atwood, W. B. et al. 2009 — Fermi LAT instrument, ApJ 697
- Gendreau, K. C. et al. 2016 — NICER on ISS, SPIE 9905
- Ricker, G. R. et al. 2015 — TESS mission overview, JATIS 1
- Harrison, F. A. et al. 2013 — NuSTAR mission, ApJ 770
- Weisskopf, M. C. et al. 2022 — IXPE mission, JATIS 8
- Yuan, W. et al. 2022 — Einstein Probe mission, Sci. Bull. 67
- Projected constellation — SpaceX files plans for million-satellite orbital data center constellation, SpaceNews 2026 (FCC application; 500–2,000 km, 30° + sun-synchronous, ≤50 km shells)
- SpaceX files for million-satellite orbital AI data center megaconstellation, Data Center Dynamics 2026
- Musk's TERAFAB chip venture; ~1 TW/yr, majority directed to space, Fortune 2026
- Walker, C. et al. 2020 — SATCON-1 Workshop Report (NSF NOIRLab / AAS) — constellation impact on optical astronomy
- Hainaut, O. R. & Williams, A. P. 2020 — Impact of satellite constellations on astronomical observations, A&A 636, A121
- Lawler, S. M., Boley, A. C. & Rein, H. 2022 — Visibility predictions for megaconstellations, AJ 163, 21 (high-altitude shells lit much of the night)
- Bassa, C. G. et al. 2022 — Analytical simulations of satellite trails in astronomical images, A&A 657, A75
- Tyson, J. A. et al. 2020 — Mitigation of LEO satellite trails for Rubin/LSST, AJ 160, 226
- Kocifaj, M. et al. 2021 — Proliferation of space objects and the diffuse night-sky brightness, MNRAS 504, L40
- Kessler, D. J. & Cour-Palais, B. G. 1978 — Collision frequency of artificial satellites: the creation of a debris belt, JGR 83, 2637 (the Kessler-cascade paper)
- Liou, J.-C. & Johnson, N. L. 2006 — Risks in space from orbiting debris, Science 311, 340 (LEO debris-population instability)
- ESA Space Environment Report — operational spatial-density and conjunction statistics
Entry IX in the RAVEN Satellite Location series · projection companion to Entry III · v1 · 2026-05-29
R.A.V.E.N.Research Labs