Note
Go to the end to download the full example code.
Solar-powered AI clusters: links, formation, radiation and cost#
Explore arXiv:2511.19468v2 (Agüera y Arcas et al., Google, June 2026) with version-pinned, manually curated evidence. The paper reports a short-path 800 Gbps one-way optical demonstration, not an operational orbital AI cluster. The optical curves and HCW formation below are engineering approximations; the power, thermal and cost comparison uses explicit synthetic assumptions.
Run python analysis/space_ai_cluster/run.py to export nine CSV tables and
input provenance. This example uses packaged data and makes no network calls.
See the modeling guide
for equations, evidence boundaries, direction conventions and exclusions.
import matplotlib.pyplot as plt
import numpy as np
from orbitalpv.orbit import circular_orbit_period_s
from orbitalpv.space_ai import FormationSpec, hcw_formation, load_space_ai_evidence
from orbitalpv.space_ai_analysis import space_ai_trade_study
evidence = load_space_ai_evidence()
tables = space_ai_trade_study(evidence)
colors = ["#157F91", "#CA8730", "#785AA0", "#C5524B"]
plt.rcParams.update({"axes.titlesize": 11, "axes.labelsize": 9, "font.size": 9})
fig, axes = plt.subplots(2, 3, figsize=(15, 9), layout="constrained")
fig.suptitle(
"Solar-powered AI in orbit | Evidence and engineering screens", fontsize=17, weight="bold"
)

Text(0.5, 0.99537, 'Solar-powered AI in orbit | Evidence and engineering screens')
links = tables["optical_links"]
ax = axes[0, 0]
ax.loglog(
links.distance_km, links.photon_ceiling_gbps / 1000, color=colors[0], label="Photon ceiling"
)
ax.loglog(
links.distance_km,
links.screened_rate_gbps / 1000,
color=colors[1],
label="With electronics cap",
)
ax.axhline(9.6, color="#555555", linestyle=":", label="24 x 400 Gbps convention")
ax.set(
xlabel="Separation [km]",
ylabel="One-direction ceiling [Tbps]",
title="5 W, 10 cm optics | far field only",
)
ax.legend(fontsize=8)
<matplotlib.legend.Legend object at 0x790bc7d38190>
spec = FormationSpec(**evidence["formation"]["inputs"])
period = circular_orbit_period_s(spec.altitude_m)
ax = axes[0, 1]
for fraction, color in zip((0, 0.125), colors[:2], strict=True):
state = hcw_formation([period * fraction], spec)[0]
ax.scatter(state[:, 1], state[:, 0], s=12, color=color, label=f"t = {fraction:g} orbit")
ax.set(xlabel="Along-track y [m]", ylabel="Radial x [m]", title="81 satellites | ideal HCW, no J2")
ax.set_aspect("equal")
ax.legend(fontsize=8)
<matplotlib.legend.Legend object at 0x790bc7d3a0d0>
compute = tables["compute_screen"]
ax = axes[0, 2]
bars = ax.barh(compute.scenario, compute.sustainable_compute_kw, color=colors)
ax.bar_label(bars, fmt="%.1f", padding=3)
ax.axvline(20, color="#555555", linestyle=":")
ax.invert_yaxis()
ax.set(
xlabel="Continuous compute electrical power [kW]",
xlim=(0, 23),
title="Analyst scenarios | 20 kW target",
)
[Text(0.5, 432.77744444444437, 'Continuous compute electrical power [kW]'), (0.0, 23.0), Text(0.5, 1.0, 'Analyst scenarios | 20 kW target')]
events = tables["radiation_events"].set_index("event_kind")
names = ["tpu_sdc_nominal", "hbm_uecc", "tpu_sefi", "host_cpu_sefi", "host_ram_sefi"]
ax = axes[1, 0]
ax.barh(
["TPU SDC", "HBM UECC", "TPU SEFI", "Host CPU SEFI", "Host RAM SEFI"],
events.loc[names, "events_per_device_year"],
color=colors[2],
)
ax.set_xscale("log")
ax.invert_yaxis()
ax.set(
xlabel="Expected events per device-year [log scale]", title="Conditional on 150 rad(Si)/year"
)
[Text(0.5, -1.3885555555555538, 'Expected events per device-year [log scale]'), Text(0.5, 1.0, 'Conditional on 150 rad(Si)/year')]
ax = axes[1, 1]
for (learning, group), color in zip(
tables["learning_curve"].groupby("learning_rate"), colors, strict=False
):
ax.loglog(
group.cumulative_mass_t,
group.price_usd_kg,
label=f"{learning:.0%} learning rate",
color=color,
)
ax.axhline(200, color="#555555", linestyle=":", label="Scenario: $200/kg")
ax.set(
xlabel="Cumulative launched mass [t]",
ylabel="Conditional launch price [USD/kg]",
title="Learning sensitivity | no date forecast",
)
ax.legend(fontsize=8)
<matplotlib.legend.Legend object at 0x790bc7d1d090>
costs = tables["launched_power"].query("launch_price_usd_kg == 200")
ax = axes[1, 2]
x = np.arange(len(costs))
ax.bar(x - 0.18, costs.calculated_usd_kw_year, width=0.36, label="Calculated", color=colors[0])
ax.bar(
x + 0.18, costs.paper_reported_usd_kw_year, width=0.36, label="Paper reported", color=colors[1]
)
ax.set_xticks(x, ["Starlink v2", "Starlink v1", "OneWeb", "Iridium"], rotation=15)
ax.set(ylabel="Launch-only USD / nameplate kW-year", title="Table 1 proxies | $200/kg scenario")
ax.legend(fontsize=8)
for ax in axes.flat:
ax.grid(alpha=0.15)
ax.spines[["top", "right"]].set_visible(False)
plt.show()
print("Near-field geometry only; no qualified multiplexed throughput:")
print(tables["spatial_geometry"].to_string(index=False))
print("Launch-only/electricity cost parity, excluding all other costs:")
print(compute[["scenario", "launch_only_energy_cost_parity_usd_kg"]].to_string(index=False))
# The paper's 9.6 Tbps direction label is ambiguous; the model explicitly assumes
# 400 Gbps per channel per direction and never doubles the reported benchmark.
# The bench demonstration, analytical capacity and synthetic useful power are
# separate evidence classes. Compute kW is not FLOPS, tokens/s or training speed.
Near-field geometry only; no qualified multiplexed throughput:
beam_count subaperture_diameter_m ideal_distance_m pointing_allowance_urad
1 0.100 5067.084925 0.986761
4 0.050 1266.771231 1.973521
16 0.025 316.692808 3.947043
Launch-only/electricity cost parity, excluding all other costs:
scenario launch_only_energy_cost_parity_usd_kg
High beta 146.253913
Low beta eclipse 32.500142
Half radiator area 100.220103
25 percent PV loss 118.058284
Total running time of the script: (0 minutes 0.398 seconds)