Pith. sign in

REVIEW 3 major objections 5 minor 31 references

This paper argues that space-based data centers can host AI inference but not frontier-scale training, because satellite mesh networks cannot synchronize gradients across thousands of nodes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 04:34 UTC pith:J4XZNWBV

load-bearing objection Useful first-order quantitative case against orbital training, but the headline 100x/10-year claim rests on a bytes-per-parameter slip; direction survives, magnitude doesn't. the 3 major comments →

arxiv 2607.14172 v1 pith:J4XZNWBV submitted 2026-07-15 cs.DC cs.AI

The Cost and Network Limits of Space-Based AI Compute

Kees van Berkel This is my paper
classification cs.DC cs.AI
keywords orbital AILEO data centersLLM trainingbisection bandwidthroofline modelall-reducelaser inter-satellite linksnetwork topology
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper evaluates whether large AI data centers in low-Earth orbit could beat terrestrial facilities on cost. Using network models rather than detailed simulation, it argues that single-satellite inference is plausible, but training frontier-scale LLMs across 8,000 satellites would be one hundred to several thousand times slower and comparably more expensive than on the ground. The reason is that the torus-shaped laser inter-satellite networks proposed for such constellations have roughly three orders of magnitude less bisection bandwidth than terrestrial Clos networks, and the gradient all-reduce that keeps replicas in sync must cross that bottleneck. If the analysis holds, orbital training is not a near-term alternative to terrestrial data centers even at a ten-year horizon.

Core claim

The paper's central claim is that the network, not launch cost, power, or cooling, is the binding constraint on orbital AI. For inference, a workload can be confined to a single satellite and the terrestrial/satellite rooflines coincide, so cost advantages in solar power and passive cooling can matter. For training, the data-parallel all-reduce across all satellites crosses the network bisection, and the 2D/3D torus topologies proposed for orbital constellations deliver only a small fraction of a Clos network's bisection bandwidth; combined with bisection intensity ~500k FLOP/byte, the roofline model puts orbital training two to three orders of magnitude below terrestrial performance, implyi

What carries the argument

The argument rests on three linked quantities: bisection bandwidth (how many bytes per second can cross the cut between two halves of the network), bisection intensity (FLOPs performed per byte that crosses the cut, here ~500k for a 1T-parameter model with 1M-token batches), and the roofline model that combines them. For a Clos network, bisection bandwidth scales with the number of racks; for a 2D/3D torus it scales as N^{1/2} or N^{2/3}, which for 8,000 nodes gives 28.8 TB/s versus 2.25 or 10 TB/s. The all-reduce operation used to average gradients in data-parallel training must cross this bisection, and the paper's key move is to apply the roofline comparison at the network level.

Load-bearing premise

The 100x cost penalty assumes that training a frontier model across 8,000 satellites requires a single global all-reduce of gradients at every step, and that all of that traffic crosses the network bisection; if parallelism were clustered so only local subsets synchronize, or if collective algorithms avoided the bisection, the penalty would shrink.

What would settle it

Run a well-engineered all-reduce for a 1T-parameter model on a simulated 8,000-node 3D torus with 100 Gb/s laser links and measure achieved average bandwidth; if it approaches the bisection roofline, the 100x penalty holds, but if ring-style algorithms sustain much higher throughput by never saturating a single bisection cut, the penalty is overstated.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Orbital AI inference on individual satellites is potentially cost-effective, since the workload need not cross a satellite network.
  • Frontier-scale LLM training in LEO would suffer model flop utilization of a few percent or less, making training runs months-to-years longer and costs 100x+ higher than terrestrial.
  • Even a 100x improvement in laser link bandwidth (to 10 Tb/s) leaves orbital training behind terrestrial Clos networks, and higher hop latencies further suppress utilization.
  • The claim that cost-effective orbital training is achievable within 2-3 years is not credible; a 10-year horizon looks optimistic on current trends.
  • No obvious alternative to all-reduce exists, so the training bottleneck is structural, not a matter of tuning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If training were spatially reorganized so only local subsets of satellites synchronize at each step (hierarchical parallelism), the all-reduce traffic crossing a global bisection would shrink; the direction of the penalty would remain, but its magnitude could drop well below 100x.
  • A hybrid architecture that sends gradient aggregates to a ground station via high-rate downlinks, rather than routing them through the satellite mesh, might sidestep the torus bisection and deserves a quantitative comparison.
  • A direct simulation of ring/recursive-halving all-reduce on a 3D torus with realistic link scheduling would test whether actual collective throughput approaches the bisection roofline or sits much higher, refining the 100x estimate.
  • The same roofline framework could be applied to orbital inference-at-scale (many satellites serving inference concurrently) to see whether bursty traffic changes the conclusion that inference is network-simple.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This paper compares a 1 GW terrestrial AI data center with two notional 8,000-satellite orbital constellations (2D and 3D torus topologies) for frontier-scale LLM training and inference. It combines a physics-side analysis (launch cost, solar power, radiative cooling, radiation, RF leakage, reentry) with a network-side roofline analysis based on bisection bandwidth and bisection intensity, parameterized by a 1T-parameter model with 1M-token batches. The central conclusions are: (i) single-satellite inference is plausible; (ii) orbital training hits network limits that make it at least 100x more expensive than terrestrial training; and (iii) Musk's 2-3 year 'space AI compute' claim is not credible, and even a 10-year horizon looks optimistic.

Significance. The paper addresses a high-profile claim with a transparent, reproducible quantitative framework. Its strengths are the use of publicly sourced inputs, the explicit sensitivity to 10 Tb/s FSO links, and the honest acknowledgment of model crudeness. The distinction between inference (feasible on one satellite) and training (network-bound) is valuable. However, the central numerical claim depends on three assumptions—the bytes per parameter crossing the bisection, the global all-reduce over all 8,000 replicas each step, and the torus bisection formula—which are not all consistent with standard data-parallel training practice. The direction of the conclusion is probably right; the magnitude is currently uncertain by roughly one order of magnitude in the optimistic direction.

major comments (3)
  1. [Section IV-C, Eq. (3)] The bisection intensity uses c = 12 bytes/parameter, described as 'weights+gradients+related numbers.' In the data-parallel setting assumed by the paper, each rack/satellite holds a full model copy; the only inter-replica traffic is gradient all-reduce. Parameters and optimizer states are updated locally, so no weights cross the bisection. Standard BF16 gradients are 2 bytes/param and FP32 master gradients 4 bytes/param, so c = 12 overstates the bisection-crossing volume by a factor of 3–6. Because the orbital roofline in Fig. 8 is proportional to I_bisect, the achievable orbital FLOP/s is understated by the same factor: for the 3D torus at 100 Gb/s the roofline rises from ~5 EF/s (c = 12) to ~15–30 EF/s (c = 4 or 2), shrinking the gap to the terrestrial 2000 EF/s from ~400x to ~130–70x. Section V's '100+ times' claim therefore depends on an unjustified choice of c. Please either justify
  2. [Section IV-C/IV-E] The model assumes a global all-reduce across all 8,000 replicas on every training step. Equation (3) is derived under that assumption. If training were organized with hierarchical/grouped data parallelism, local SGD with periodic synchronization, or asynchronous updates, the bisection-crossing volume per useful FLOP could be reduced by a factor K, shifting the roofline up by K. Section V asserts that 'there are no obvious alternatives to all-reduce' without analysis. This matters because the magnitude of the orbital penalty, not its direction, is at issue. Please discuss such algorithmic mitigations and their convergence trade-offs, or restrict the claim to dense per-step all-reduce synchronization.
  3. [Section IV-E] The statement that common all-reduce algorithms on torus networks require the data volume to cross the bisection 'once per dimension' is not reflected in Eq. (3) or Fig. 8. If c = 12 is meant to include this multi-pass effect, it should be explicit and dimension-dependent (2D vs 3D torus); if it is not included, the plotted rooflines for torus training are too optimistic by a factor of 2–3. As written, the relationship between the bisection intensity, Table 3's bisection bandwidths, and the multi-pass nature of torus all-reduce is ambiguous and needs clarification.
minor comments (5)
  1. [Section IV-B, Table 3] The text gives bisection bandwidth formulas N^(1/2)*Bl and N^(2/3)*Bl, while the table uses 2N^(1/2)*Bl and 2N^(2/3)*Bl. Clarify the factor of 2 (two directions of the torus) in the text.
  2. [Section IV-A] 8,000 satellites at 100 kW yields 800 MW, not 1 GW; if the comparison is with 8,000 racks at 120 kW (960 MW), the comparison is not equal-power. Use N = 10,000 satellites or adjust the stated power.
  3. [Section V] The sentence arguing that network-wide latencies may make terrestrial MFU values of 40–70% unrealistic conflicts with the IMEC model cited in II-C, where latency is said to be negligible for training. Either qualify or remove this line.
  4. [Section IV-D, Fig. 7] The caption says the bisection bandwidth applies to intra-rack communication. Clarify 'intra-rack' vs 'inter-node' because the roofline in a single satellite is memory-bound, not network-bound.
  5. [References] Several citations are to web pages and Wikipedia (e.g., [23], [25]). For a journal version, replace with primary sources or archival versions where available.

Circularity Check

0 steps flagged

No significant circularity: the orbital-training roofline conclusion follows from explicit stated assumptions and parameter values, with no fitted-parameter or self-citation reduction.

full rationale

I walked the claimed derivation chain. Section IV-C defines bisection intensity and Equation (3) gives I_bisect = (k/c) * B_tokens, with k≈6 FLOP/parameter/token and c≈12 bytes/parameter stated as typical values, not fitted to the orbital conclusion. Section IV-E then combines these intensities with bisection bandwidths from Table 3, which are themselves derived from the assumed Clos, 2D-torus, and 3D-torus topologies, to produce the rooflines. The Section V conclusion—that cost-effective frontier-LLM training in space within 2-3 years is not credible and even 10 years appears optimistic—is an arithmetic consequence of those explicit assumptions (8000 satellites, 100 Gb/s laser inter-satellite links, torus topologies, and gradient all-reduce crossing the bisection), not a result that is true by construction. The paper also includes a sensitivity case with 10 Tb/s free-space-optical links and still finds orbital networks inferior to terrestrial Clos networks, which shows the qualitative conclusion is not reverse-engineered from a single fitted constant. There are no load-bearing self-citations: the author does not cite prior work of his own, and the external references (Starcloud, Google's 81-satellite study, Starlink, the IMEC model) are used as stated inputs or opposing claims rather than as authority for the paper's own result. The skeptical concern about c=12 bytes/parameter versus the 2-4 bytes/parameter typically associated with gradients is a modeling-parameter critique, not a circularity: the parameter is an input assumption, and changing it changes the magnitude but not the directional conclusion under the paper's assumed all-reduce pattern. Therefore no circular step is established, and the honest finding is no significant circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 8 axioms · 3 invented entities

The paper's central 100x+ training penalty is arithmetic under a small set of point estimates and standard network formulas. The orbital constellation geometries are invented for the model and have no independent evidence, but the author tests a 10Tb/s FSO sensitivity that preserves the qualitative conclusion. No hidden fitted parameters are used; the 'star' at 2000 EFLOP/s is a sanity check, not a fit.

free parameters (6)
  • Solar conversion efficiency = 30%
    Assumed in III-B to get 400 W/m² from 1300 W/m² insolation; controls per-satellite power and cooling.
  • Launch cost to LEO = $200/kg
    Expected Starship cost used in III-A; affects Capex comparison but not the network conclusion.
  • Radiator temperature = 370 K
    Taken from Musk's X post in III-C; sets required radiator area via Stefan-Boltzmann.
  • Global batch size for 1T model = ~1M tokens
    Used in IV-C to set bisection intensity at 500k FLOP/Byte.
  • Bytes per parameter = 12 bytes
    Assumed weights+gradients+related numbers in IV-C; affects I_bisect.
  • Per-link laser bandwidth baseline = 100 Gb/s
    Starlink V2 ISL value used in Table 3; the paper tests 10 Tb/s FSO as a sensitivity.
axioms (8)
  • standard math Stefan-Boltzmann radiative cooling law
    Used in III-C to estimate radiator area.
  • standard math Bisection bandwidth formulas for Clos and torus networks
    Used in IV-B to compute Table 3.
  • domain assumption Roofline model bounds achievable FLOP/s as min(compute peak, bisection bandwidth x bisection intensity)
    Core modeling framework in IV-D and IV-E.
  • domain assumption A 120 kW terrestrial rack and a 100 kW V3 satellite are compute-equivalent nodes
    Explicitly stated in Section IV introduction; enables the 8000-node comparison.
  • domain assumption Frontier LLM training requires a global all-reduce of gradients across all nodes
    Stated in IV-C; determines that bisection traffic scales with full model replicas.
  • domain assumption All-reduce traffic crosses the network bisection
    Used in IV-C and IV-E to apply bisection bandwidth to training rooflines.
  • domain assumption Sun-synchronous LEO provides 24/7 insolation at 1300 W/m²
    Used in III-B to justify lower orbital OpEx.
  • ad hoc to paper 8000 satellites can be maintained in tight 2D or 3D torus formation with 100 Gb/s laser ISLs
    Proposed in IV-A as the study's orbital network; no such constellation exists and the paper calls it speculative.
invented entities (3)
  • 2D planar 89x90 satellite constellation with torus network no independent evidence
    purpose: Model orbital training network with four laser links per satellite
    Hypothetical architecture assumed for the study; no prototype, mission data, or external design reference.
  • 3D cubic 20x20x20 satellite constellation with torus network no independent evidence
    purpose: Model orbital training network with six laser links per satellite
    Hypothetical architecture assumed for the study; the paper acknowledges orbital dynamics and shadowing challenges.
  • Bi-planar Clos satellite constellation no independent evidence
    purpose: Discussed as a potential way to recover Clos-like bisection in orbit
    Proposed only in Section V as a direction; no design, link budget, or feasibility analysis.

pith-pipeline@v1.3.0-alltime-deepseek · 11039 in / 13463 out tokens · 141972 ms · 2026-08-02T04:34:03.312482+00:00 · methodology

0 comments
read the original abstract

This paper evaluates whether large-scale AI data centers deployed in low-Earth orbit (LEO) could become a cost-effective alternative to terrestrial facilities. The analysis compares orbital and ground-based systems across launch cost, power generation, cooling, radiation exposure, and atmospheric reentry, as well as compute-network performance. A key distinction is the shift from terrestrial Clos networks to space-based mesh networks using laser inter-satellite links. Using bisection bandwidth, bisection intensity, and roofline-style models, we show that while LEO-based inference may be feasible, training frontier-scale LLMs in orbit is unlikely to be competitive with terrestrial data centers.

Figures

Figures reproduced from arXiv: 2607.14172 by Kees van Berkel.

Figure 1
Figure 1. Figure 1: Projected power growth for frontier AI training, from [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: A breakdown of normalized training times across multiple GPU [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cost of launch to low earth orbit. Adapted from [ [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Unintended emission as observed by LOFAR, from [ [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Al2O3 aerosol accumulation assuming satellite reentry above the equator (red box), from [28]. F. Reentry LEO satellites are designed for a lifetime of about 5 years. End-of-life implies reentry into earth’ atmosphere. The best possible outcome of such reentry is total vaporization. The effects of such reentries has been simulated by the US NOAA [28], assuming a total reentry mass of 104 ton/year. A satelli… view at source ↗
Figure 7
Figure 7. Figure 7: The rooflines for terrestrial and orbital inference of a 1T-parameter [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Rooflines for terrestrial/orbital training of a 1T-parameter LLM, [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 6 linked inside Pith

  1. [1]

    Why we should train ai in space,

    E. Feilden, A. Oltean, and P. Johnston, “Why we should train ai in space,” Starcloud, Inc., Tech. Rep. White Paper v1.03, Sep. 2024, formerly Lumen Orbit. [Online]. Available: https: //starcloudinc.github.io/wp.pdf

    2024

  2. [2]

    xai joins spacex,

    SpaceX, “xai joins spacex,” https://www.spacex.com/updates# xai-joins-spacex, 2025, accessed: 2026-06-01

    2025

  3. [3]

    Orbital data center loa – narrative,

    C. Tenge-Rietberg, “Orbital data center loa – narrative,” Space Exploration Technologies Corp., Hawthorne, CA, Tech. Rep., January 2026, fCC Application for Launch and Operating Authority for the SpaceX Orbital Data Center System. [Online]. Available: https: //cdn.geekwire.com/wp-content/uploads/2026/01/SpaceX-Center.pdf

    2026

  4. [4]

    Sec filing by space exploration technologies corp

    Space Exploration Technologies Corp., “Sec filing by space exploration technologies corp.” https://www.sec.gov/Archives/edgar/data/1181412/ 000162828026036936/spaceexplorationtechnologi.htm, May 2026, ac- cessed: 2026-06-03

    arXiv 2026

  5. [5]

    Towards space- based computing infrastructure network: Development trends, network architecture, challenges analysis, and key technologies,

    L. Kuang, J. Sun, J. Zhang, H. Cui, and K. Liu, “Towards space- based computing infrastructure network: Development trends, network architecture, challenges analysis, and key technologies,” 2025. [Online]. Available: https://arxiv.org/abs/2503.06521

    arXiv 2025

  6. [6]

    Space computing: Architectures, challenges, and future directions,

    E. Xue, Z. Zhang, J. Xue, H. Wang, I. E. Carvajal-Roca, Z. He, H. Zhang, H. Wang, Z. Wan, and C. Li, “Space computing: Architectures, challenges, and future directions,”Intelligent Computing, vol. 4, 2025. [Online]. Available: https://spj.science.org/doi/10.34133/ icomputing.0134

    2025

  7. [7]

    Towards a future space-based, highly scalable ai infrastructure system design,

    B. Ag ¨uera y Arcas, T. Beals, M. Biggs, J. V . Bloom, T. Fischbacher, K. Gromov, U. K ¨oster, R. Pravahan, and J. Manyika, “Towards a future space-based, highly scalable ai infrastructure system design,” 2025. [Online]. Available: https://arxiv.org/abs/2511.19468

    Pith/arXiv arXiv 2025

  8. [8]

    Nvidia gb200 nvl72: Rack-scale blackwell platform for large-scale ai,

    NVIDIA Corporation, “Nvidia gb200 nvl72: Rack-scale blackwell platform for large-scale ai,” https://www.nvidia.com/en-us/data-center/ gb200-nvl72/, 2024, accessed: 2026-06-16

    2024

  9. [9]

    Jupiter rising: A decade of clos topologies and centralized control in google’s datacenter network,

    A. Singh, J. Ong, A. Agarwal, G. J. Anderson, A. Armistead, R. Bannon, S. Boving, G. Desai, B. Felderman, P. Germano, P. Kanagala, J. Provost, J. Simmons, E. Tanda, A. Vahdat, D. Wanderer, and U. H ¨olzle, “Jupiter rising: A decade of clos topologies and centralized control in google’s datacenter network,” inProceedings of the ACM SIGCOMM Conference, 2015...

    2015

  10. [10]

    Servers account for 60% of the total cost of ownership of a one-gigawatt ai data center,

    A. Michael and B. Cottier, “Servers account for 60% of the total cost of ownership of a one-gigawatt ai data center,” May 2026, accessed 2026-06-14. [Online]. Available: https://epoch.ai/data-insights/ ai-datacenter-cost-breakdown

    2026

  11. [11]

    L. A. Barroso, J. Clidaras, and U. H ¨olzle,The Datacenter as a Computer: Designing Warehouse-Scale Machines, 4th ed. Cham: Springer, 2025. [Online]. Available: https://link.springer.com/book/10. 1007/978-3-031-99489-0

    2025

  12. [12]

    Wlb-llm: workload-balanced 4d parallelism for large language model training,

    Z. Wang, A. Cai, X. Xie, Z. Pan, Y . Guan, W. Chu, J. Wang, S. Li, J. Huang, C. Cai, Y . Hao, and Y . Ding, “Wlb-llm: workload-balanced 4d parallelism for large language model training,” inProceedings of the 19th USENIX Conference on Operating Systems Design and Implementation, ser. OSDI ’25. USA: USENIX Association, 2025

    2025

  13. [14]

    Palm: Scaling language modeling with pathways,

    A. Chowdhery, S. Narang, J. Devlinet al., “Palm: Scaling language modeling with pathways,”arXiv preprint arXiv:2204.02311, 2022

    Pith/arXiv arXiv 2022

  14. [15]

    Reducing acti- vation recomputation in large transformer models,

    J. Hoffmann, S. Borgeaud, A. Menschet al., “Reducing acti- vation recomputation in large transformer models,”arXiv preprint arXiv:2205.05198, 2022

    Pith/arXiv arXiv 2022

  15. [16]

    Scaling infrastructure to support multi-trillion parameter llm training,

    M. Isaev, N. McDonald, and R. Vuduc, “Scaling infrastructure to support multi-trillion parameter llm training,”OpenReview, 2023. [Online]. Available: https://openreview.net/forum?id=rqn2v1Ltgn0

    2023

  16. [17]

    Megascale: Scaling large lan- guage model training to more than 10,000 gpus,

    Z. Jiang, H. Lin, Y . Zhonget al., “Megascale: Scaling large lan- guage model training to more than 10,000 gpus,”arXiv preprint arXiv:2402.15627, 2024

    Pith/arXiv arXiv 2024

  17. [18]

    The llama 3 herd of models,

    Llama Team, “The llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  18. [19]

    Scaling intelligence: The exponential growth of ai’s power needs,

    EPRI Staff, “Scaling intelligence: The exponential growth of ai’s power needs,” Electric Power Research Institute (EPRI), Tech. Rep., December 2025, white paper, accessed 2026-06-03. [Online]. Available: https://restservice.epri.com/publicattachment/94532

    2025

  19. [20]

    The next big shifts in ai workloads and hyperscaler strategies,

    C. Arora, M. Sorel, P. Sachdeva, A. Bhan, J. He, N. Shaw, R. Garg, and S. Ravishankar, “The next big shifts in ai workloads and hyperscaler strategies,” https://www.mckinsey.com/ industries/technology-media-and-telecommunications/our-insights/ the-next-big-shifts-in-ai-workloads-and-hyperscaler-strategies, Dec. 2025, mcKinsey & Company, accessed 2026-06-03

    2025

  20. [21]

    Evaluating Cross- Architecture Performance Modeling of Distributed ML Workloads Using StableHLO ,

    J. Svedas, N. Laubeuf, R. Harvey, A. Singh, C. Man, A. Nada, T. Krishna, J. Myers, and D. Bhattacharjee, “ Evaluating Cross- Architecture Performance Modeling of Distributed ML Workloads Using StableHLO ,” in2026 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS). Los Alamitos, CA, USA: IEEE Computer Society, Apr. 2026, ...

    arXiv 2026

  21. [22]

    Performance modeling and workload analysis of dis- tributed large language model training and inference,

    J. Kundu, W. Guo, A. BanaGozar, U. De Alwis, S. Sengupta, P. Gupta, and A. Mallik, “Performance modeling and workload analysis of dis- tributed large language model training and inference,” in2024 IEEE International Symposium on Workload Characterization (IISWC), 2024, pp. 57–67

    2024

  22. [23]

    Starlink — Wikipedia, the free encyclopedia,

    Wikipedia contributors, “Starlink — Wikipedia, the free encyclopedia,” 2026, [Online; accessed 14-July-2026]. [Online]. Available: https: //en.wikipedia.org/wiki/Starlink

    2026

  23. [24]

    Laser intersatellite links in a starlink constellation: A classification and analysis,

    A. U. Chaudhry and H. Yanikomeroglu, “Laser intersatellite links in a starlink constellation: A classification and analysis,”IEEE Vehicular Technology Magazine, vol. 16, no. 2, pp. 48–56, 2021

    2021

  24. [25]

    Cost to leo,

    Marspedia Contributors, “Cost to leo,” https://marspedia.org/File:Cost To LEO.png, 2026, image file, accessed 2026-06-03

    2026

  25. [26]

    Spacecraft chassis and component configuration,

    J. A. Dunford, G. P. Simard, S. T. Reineman, M. W. Read, and B. Crabtree, “Spacecraft chassis and component configuration,” US Patent US11 834 205B1, Dec. 5, 2023, granted patent, accessed 2026-06-03. [Online]. Available: https://patents.google.com/patent/US11834205B1/en

    2023

  26. [27]

    Unintended electromagnetic radiation from starlink satellites detected with lofar between 110 and 188 mhz,

    Di Vruno, F., Winkel, B., Bassa, C. G., J ´ozsa, G. I. G., Brentjens, M. A., Jessner, A., and Garrington, S., “Unintended electromagnetic radiation from starlink satellites detected with lofar between 110 and 188 mhz,”A&A, vol. 676, p. A75, 2023. [Online]. Available: https://doi.org/10.1051/0004-6361/202346374

    doi:10.1051/0004-6361/202346374 2023

  27. [28]

    Investigating the potential atmospheric accumulation and radiative impact of the coming increase in satellite reentry frequency,

    C. M. Maloney, R. W. Portmann, M. N. Ross, and K. H. Rosenlof, “Investigating the potential atmospheric accumulation and radiative impact of the coming increase in satellite reentry frequency,”Journal of Geophysical Research: Atmospheres, vol. 130, no. 6, p. e2024JD042442, 2025, e2024JD042442 2024JD042442. [Online]. Available: https: //agupubs.onlinelibra...

    doi:10.1029/2024jd042442 2025

  28. [29]

    Optimization of collective reduction operations for massively parallel architectures,

    R. Rabenseifner, “Optimization of collective reduction operations for massively parallel architectures,” inInternational Conference on Com- putational Science, 2004, pp. 1–9

    2004

  29. [30]

    Collective algorithms for multiported torus networks,

    P. Sack and W. Gropp, “Collective algorithms for multiported torus networks,”ACM Transactions on Parallel Computing, vol. 1, no. 2, pp. 12:1–12:26, 2015

    2015

  30. [31]

    Efficiently scaling transformer inference,

    R. Pope, D. Sholtoet al., “Efficiently scaling transformer inference,” arXiv preprint arXiv:2211.05102, 2022

    Pith/arXiv arXiv 2022

  31. [32]

    Overview of free-space laser communica- tions technology state-of-the-art: Challenges, solutions, and standards,

    D. Boroson and H. Hemmati, “Overview of free-space laser communica- tions technology state-of-the-art: Challenges, solutions, and standards,” IEEE Journal of Selected Topics in Quantum Electronics, vol. 32, no. 1: Advances in Free Space Laser Communications, pp. 1–16, 2026

    2026