Orbital data centers: open problems and a research plan for the network軌道資料中心：待研究問題與網路研究計畫

13.1What is known, what is open已知的事與待研究的事

An orbital data center combines compute satellites through short optical links into one cluster. A dedicated ground segment handles data, while a separate communications constellation supplies user access.軌道資料中心以短距光學鏈路將運算衛星整合為單一叢集，資料經專用地面段進出，使用者則透過另一個通訊星座存取。

A research question specifies a test that could reject its hypothesis and a baseline it must exceed. A link-selection question needs terminal counts and blocking rules, which remain unspecified in the Suncatcher paper. Pénot and Balakrishnan already solve line-of-sight fabric feasibility with integer optimization ^{[102, Eq. 7]}, while van Berkel supplies torus and Clos bisection bounds ^{[101, Tab. 3]}. The six directions extend this evidence with explicit methods, metrics, baselines and stopping criteria, using data and tools available in October 2026. Table 13.1 maps each source-defined gap to a direction.

本章的研究問題，同時列出能檢驗假設的量測或模擬，以及必須超越的基準。鏈路選擇需要終端數及遮擋規則，Suncatcher 對這些項目的說明仍待補充。Pénot 與 Balakrishnan 已以整數最佳化求解視線織網可行性^{[102, Eq. 7]}，van Berkel 則提供環面與 Clos 對分界限^{[101, Tab. 3]}。六個方向據此延伸，使用 2026 年 10 月可取得的資料與工具，明列方法、指標、基準及停止條件，表 13.1 再將各文獻缺口對應至研究方向。

Table 13.1. Open problems recorded in the fact base for this chapter, each anchored to the document that states or exposes it. The letters refer to the research directions of Sections 13.2–13.7.

表 13.1。各項待研究問題及陳述或揭露缺口的來源文獻，字母對應第 13.2–13.7 節的研究方向。

Public milestones constrain the plan. The first listed opportunity for Tbps link measurements is therefore the planned 2027 test, which makes Direction F the initial priority.公開里程碑限制了研究時程。列出的 Tbps 鏈路實測機會因此始於預定 2027 年測試，使方向 F 成為起點。

13.2Research direction A — Collectives on a predictable, time-varying fabric研究方向 A — 可預測、時變織網上的集體通訊

A · The gap and the document that exposes itA · 缺口與揭露它的文獻

The reviewed literature still needs a collective simulation for a Tbps formation. Van Berkel assumes fixed-rate lasers and a static graph ^{[101, Tab. 3, §IV-E]}. Suncatcher's HCW geometry changes neighbor distances periodically ^{[36, p. 4, Fig. 3]}, while its spatial-multiplexing rule makes aggregate bandwidth depend inversely on distance ^{[36, p. 3, Fig. 1]}. Together, these assumptions imply predictable rate variation over half-orbit cycles. Direction A asks whether a collective schedule can use that variation.

已檢閱文獻仍需補上 Tbps 編隊的集體通訊模擬。van Berkel 採固定雷射速率與靜態圖^{[101, Tab. 3, §IV-E]}，Suncatcher 的 HCW 幾何則使鄰居距離週期變動^{[36, p. 4, Fig. 3]}，空間多工法則又讓總頻寬與距離成反比^{[36, p. 3, Fig. 1]}。這些假設共同產生可預測的半軌道週期速率變化，方向 A 因而檢驗集體通訊排程如何利用它。

A · The falsifiable questionA · 可檢驗的問題

Can advance knowledge of \(d_{ij}(t)\) keep exposed communication within ten percent of its best orbital phase? Compare with static rings and trees provisioned at the worst-case distance. Close the direction if a static schedule stays within ten percent of the orbit-aware result at every phase. If improvement requires repointing, charge the measured setup range ^{[33, Tab. 1]} and compare setup time with step time before simulation.

預知 \(d_{ij}(t)\) 的排程，可將暴露通訊維持在最佳軌道相位的百分之十範圍內嗎？比較對象是按最差距離配置的靜態環與樹。若靜態排程在每個相位都位於軌道感知結果的百分之十內，即達到停止條件。若改善依賴重新指向，則需計入量得的建立時間範圍^{[33, Tab. 1]}，先比較建立與步驟時間，再進行模擬。

A · Method with today's data and toolsA · 用今日資料與工具的方法

Extrapolation below 0.32 km is a declared assumption. Compare a ring, lattice-based two-dimensional reduce-scatter and all-gather, and a tree, using Chapter 10's compute-time model.

A · Metrics and baselinesA · 指標與基準

Report step time across a full orbit as its mean, peak-to-trough ratio and exposed fraction \(T_{\text{exposed}}/T_{\text{step}}\). Count repointings per orbit as a schedule cost. Compare a conservative ring provisioned at 200 m with an optimistic ring provisioned at 100 m. The conservative ring leaves closest-approach capacity idle, while the optimistic ring stalls when rates fall below its allocation. Require gains over both static baselines at every orbital phase.

以整圈軌道的步驟時間平均、峰谷比及暴露比例 \(T_{\text{exposed}}/T_{\text{step}}\) 作為指標，並以每圈重新指向次數計入成本。比較基準為按 200 m 配置的保守環，以及按 100 m 配置的積極環。保守環在最近距離留下閒置容量，積極環則在速率低於分配份額時停滯，因此需在每一個軌道相位都勝過這兩個靜態基準，控制基準選擇的影響。

A · What would falsify itA · 停止研究的條件

The rate-versus-range measurement is the first decision gate, using the proposed 100 m–2 km test interval. If the tested link stays effectively fixed across distance, use a static-fabric model. That trade needs steps well below a minute to matter. Confocal multiplexing estimates also exclude beam quality, truncation and jitter ^{[36, p. 8]}.速率對距離的量測是第一個決策關卡。若鏈路在各距離維持近似固定速率，則改採靜態織網模型。

A · Why nowA · 為什麼是現在

Public formation geometry and the published rate rule support a software study before flight calibration. The planned two-satellite test can supply the rate-versus-range curve needed to compare predictions with measured behavior.

公開編隊幾何與已發表速率法則，可支持飛行校準前的軟體研究。預定雙星測試可提供速率對距離曲線，讓模型預測與量測行為直接比較。

13.3Research direction B — The cluster fabric with real constraints研究方向 B — 含真實限制的叢集織網

B · The gap and the document that exposes itB · 缺口與揭露它的文獻

The two fabric studies cover complementary constraints. Two-layer designs are usually infeasible, while three or more layers succeed across their sweep. Suncatcher specifies an “8-neighborhood” only for its central satellite ^{[36, p. 4, Fig. 2]}. Computing its full-cluster bisection therefore requires terminal and switching details beyond the published specification (Part D item 21).兩篇織網研究涵蓋互補的約束。

B · The falsifiable questionB · 可檢驗的問題

Which lattice, torus or dedicated-switch Clos delivers the most bisection per launched kilogram under visibility, range-dependent rate and setup constraints? Determine the scale where the lattice reaches the training intensity requirement.在視線、距離相依速率與建立時間約束下，點陣、環面及專用交換 Clos，哪一種提供最多的每發射公斤對分頻寬？研究需找出點陣達到訓練對分強度需求的規模。

B · Method with today's data and toolsB · 用今日資料與工具的方法

Reproduce Pénot and Balakrishnan's four constraints: minimum spacing, maximum envelope, an unobstructed Sun vector and stable clear links. Use mixed-integer optimization and share position traces with Direction A. Evaluate max-flow across balanced cuts, then take the worst cut and minimum over orbital phase. Mass uses the Joule paper's 575 kg v2-mini analogue ^{[36, p. 10, Table 1]}, with terminal mass left as a parameter.重現 Pénot 與 Balakrishnan 的四項約束：最小間距、最大包絡、清晰太陽向量及穩定清晰鏈路。採混合整數最佳化，並與方向 A 共用位置軌跡。

B · Metrics and baselinesB · 指標與基準

Report minimum orbital bisection per launched mass, compute fraction r, network diameter, mean hops and repointings needed to preserve connectivity. Use van Berkel's torus and Clos formulas at matching N and the plain lattice baseline. Follow Table 3's planar-torus coefficient, whose factor of two differs from the text (fact base Part C item 41). The lattice preserves compute positions by avoiding dedicated switch satellites.

報告每發射質量的軌道最小對分頻寬、運算比例 r、網路直徑、平均跳數，以及維持連通所需的重新指向次數。基準採相同 N 的 van Berkel 環面與 Clos 公式，以及保留運算位置的單純點陣。平面環面係數依 Table 3，與正文省略的兩倍係數差異並列記錄（C 部第 41 項）。

B · What would falsify itB · 停止研究的條件

Close the comparison if lattice capacity meets every formation size admitted by the envelope, or if a published fabric resolves the assumed design choices. Greater measured spatial-multiplexing gain would move the crossover beyond feasible planar packing. The calculation's scale and array assumptions remain explicit in the table.

若點陣容量涵蓋包絡內所有可容納規模，或公開織網已釐清假定的設計選項，即達到比較的停止條件。更高的實測空間多工增益，會將交叉點移至平面堆積可行範圍之外，表格保留計算尺度及陣列假設。

B · Why nowB · 為什麼是現在

Published packing constraints, bisection formulas and the shared geometry support a solver-based comparison before the planned flight test. A measured rate curve can later calibrate it.

公開堆積約束、對分公式與共用幾何，可支持預定飛行測試前的求解器比較，再以實測速率曲線校準。

13.4Research direction C — Failure arithmetic in orbit研究方向 C — 軌道上的故障算術

C · The gap and the document that exposes itC · 缺口與揭露它的文獻

The paper says the training effect “requires further study” (p. 5). SpaceX's filing assumes five-year satellite life (DCD, 2026 company).SpaceX 申請假設五年壽命（DCD，2026 company）。

C · The falsifiable questionC · 可檢驗的問題

For synchronous formation training, do detection and recovery exceed ten percent of step time or bisection? Can job interruption remain within the terrestrial reference? A neighboring replica transfer is sub-second under the assumed aperture rate, so silent-corruption detection may dominate the network work. Compare verification traffic with checkpoint movement.

同步編隊訓練的偵測與恢復，會占用超過百分之十的步驟時間或對分頻寬嗎？工作中斷能維持在地面參考範圍內嗎？假定孔徑速率下的鄰居複本傳輸低於一秒，因此靜默損毀偵測可能主導網路工作，需比較驗證與檢查點流量。

C · Method with today's data and toolsC · 用今日資料與工具的方法

Model per-chip radiation events as Poisson processes using the published per-rad rates. Add scheduled thermal pauses and satellite replacement, then inject events into the shared simulator. Price cross-replica weight-hash checks after each all-reduce. Include step recomputation after hash disagreement and collective rerouting around powered-down members. Staggering is a design variable: 81 satellites pausing every 15 minutes create membership changes about every 11 s, while aligned pauses stall the cluster.

以公開每 rad 速率建立各晶片的卜瓦松事件模型，加入排定的熱暫停與衛星汰換，再注入共用模擬器。成本包含每次 all-reduce 後的跨複本權重雜湊比對、結果有差異時的步驟重算，以及繞過斷電成員的集體通訊改道。錯開方式是設計變數，81 顆衛星各每 15 分鐘暫停，均勻錯開約每 11 s 發生成員變動，對齊暫停則使整個叢集停滯。

C · Metrics and baselinesC · 指標與基準

Define verified goodput as the fraction of a full orbit spent on steps that pass verification. Report this fraction, verification and recovery’s bisection-bandwidth share, and monthly job interruptions. Compare with HPN statistics and SpaceMoE's independent survival model. Thermal pauses and reacquisition correlate across time and neighbors, giving an explicit test of that independence assumption.

以整圈軌道中投入通過驗證步驟的時間比例，定義經驗證的有效產出，並報告此比例、驗證與恢復占用的對分頻寬，以及每月工作中斷數，並比較 HPN 統計與 SpaceMoE 的獨立存活模型。熱暫停及重新擷取在時間與鄰居間具有相關性，可用於檢驗獨立假設。

C · What would falsify itC · 停止研究的條件

Beam-test rates need orbital validation. The peer-reviewed flight study recorded zero single-event computational errors over six months on commercial boards ^[85]. Much rarer MVP corruption than the predicted rate would shift attention toward thermal scheduling. Close the network direction if verification and checkpoints use less than one percent of bisection and interruptions remain below HPN. The remaining work would concern system reliability.

質子束速率需要在軌驗證，同儕審查飛行研究在商用板上觀測六個月，記錄零次單事件運算錯誤^[85]。若 MVP 損毀率遠低於預測值，研究重點將轉向熱排程。當驗證與檢查點合計低於對分頻寬百分之一、中斷低於 HPN，即達到網路方向的停止條件，其餘工作則屬系統可靠度。

C · Why nowC · 為什麼是現在

The June 2026 preprint and Joule paper publish the event rates. Telemetry from the prototype's one-year mission can test the model as releases become available.

2026 年 6 月預印本與 Joule 論文已提供事件速率，原型一年任務期間釋出的遙測，可用來檢驗模型。

13.5Research direction D — The front end: feeding and draining an orbital data center研究方向 D — 前端：為軌道資料中心進料與出料

D · The gap and the document that exposes itD · 缺口與揭露它的文獻

Suncatcher identifies its ground segment as an open design problem. The pilot uses radio, with optical downlink planned later and TBIRD cited as the reference. The paper also predicts higher latency to some ground locations from the dawn–dusk orbit ^{[36, pp. 2, 7]}. Extending these studies to TB–PB training-data ingest requires an uplink model, since TBIRD demonstrates space-to-ground transfer.Suncatcher 將地面段列為待研究設計，試驗先採電波，後續規劃光學下行，並引用 TBIRD 作參考，同時指出晨昏軌道會增加部分地面位置的延遲^{[36, pp. 2, 7]}。延伸至 TB–PB 訓練資料輸入時，需另建上行模型，因為 TBIRD 展示的是太空至地面傳輸。

D · The falsifiable questionD · 可檢驗的問題

Which station network meets the cluster's ingest and checkpoint deadlines under realistic cloud cover? Compare many small L2D2-style stations with fewer large telescopes.在真實雲量下，哪種站網能達成叢集的資料輸入與檢查點期限？

D · Method with today's data and toolsD · 用今日資料與工具的方法

Propagate the 650 km dawn–dusk orbit using Kepler relations and each station's elevation mask. Scale measured TBIRD pass volumes by contact duration and slant range. Public cloud records supply site-specific weather because L2D2's attenuation statements conflict (fact base Part C, item 48). Evaluate satellite–station matching using predicted link quality. Sweep ingest rate from radio capability to the TBIRD downlink rate, making uplink uncertainty an explicit sensitivity.

以克卜勒關係及各站仰角遮罩推演 650 km 晨昏軌道，依接觸時間與斜距縮放 TBIRD 實測傳輸量。L2D2 的衰減陳述互有衝突，因此氣象輸入採各站公開雲量紀錄。依預測鏈路品質評估衛星與地面站配對，再將資料輸入速率從電波能力掃描至 TBIRD 下行速率，明列上行能力的敏感度。

D · Metrics and baselinesD · 指標與基準

Report daily transfer volume in both directions, ingest completion time, checkpoint arrival delay and station count. Compare one high-latitude TBIRD-class station, the bent-pipe station-count rule and distributed L2D2-style sites.

報告雙向每日傳輸量、資料輸入完成時間、檢查點抵達延遲及地面站數，比較單一高緯度 TBIRD 級站、彎管站數法則，以及分散式 L2D2 站網。

D · What would falsify itD · 停止研究的條件

Two sufficient stations would narrow this to site selection. Launching datasets with storage, or training on orbital data, would shift the objective toward checkpoint egress. The calculation's clear-sky checkpoint target fits one pass.

若兩個站已足以達標，研究範圍便集中於選址。資料隨儲存設備發射，或訓練只採在軌資料時，目標則轉向檢查點回傳，計算中的晴空檢查點目標可容納於一次過頂。

D · Why nowD · 為什麼是現在

TBIRD supplies measured optical transfer, L2D2 supplies scheduling methods, and public orbit and weather records support site sizing. Complete that study before selecting sites for future flight tests.

TBIRD 提供實測光學傳輸，L2D2 提供排程方法，公開軌道與氣象紀錄則支持站網規模研究，應在後續飛行測試選址前完成評估。

13.6Research direction E — Serving across ground and orbit研究方向 E — 跨地面與軌道的服務

E · The gap and the document that exposes itE · 缺口與揭露它的文獻

Deployment comparison still requires a common ground baseline, joint time-and-energy accounting and scheduled reconfiguration ^{[13, §3–4]}.部署比較仍需共同地面基準、時間與能源聯合計算，以及排程式重新配置^{[13, §3–4]}。

E · The falsifiable questionE · 可檢驗的問題

Which request mixes, user locations and link rates favor an orbital cluster for first-token time and energy per token? Include relay transit, reconfiguration and KV cache locality. Report the region where orbit wins, or establish that ground serving wins across the tested user space.

哪些請求組合、使用者位置及鏈路速率，使軌道叢集在首 token 時間與每 token 能源上占優？模型需包含中繼、重新配置及KV 快取局部性，結果應界定軌道勝出的參數區域，或證實地面在測試的使用者範圍內全面占優。

E · Method with today's data and toolsE · 用今日資料與工具的方法

Declare constellation hops and node delay as parameters, then add relay geometry. Compare grid-powered ground service with the mostly sunlit dawn–dusk cluster. Cluster-local KV state stays at one serving site while access satellites change. Serving directly on moving communications satellites instead requires cache migration or rebuilding.星座跳數與節點延遲明列為參數，再加入中繼幾何。叢集本地 KV 狀態留在固定服務位置，存取衛星更換時只改變路徑。

E · Metrics and baselinesE · 指標與基準

Report first-token and output-token time, energy and cost per token, and the share of sessions requiring cache movement. Compare ground serving, SpaceMoE-style serving on communications satellites, and split prefill and decoding placed by data and path locality.

報告首 token 與輸出 token 時間、每 token 能源及成本，以及需要搬移快取的會話比例。基準包含地面服務、SpaceMoE 式通訊衛星服務，以及依資料和路徑局部性配置的 prefill 與解碼分割方案。

E · What would falsify itE · 停止研究的條件

Close the orbital-benefit hypothesis if ground serving wins throughout the tested space after matched access overhead. Remote users remain a candidate regime: Bose et al. measured Mozambique CDN latency of 138.7 ms versus 7.2 ms ^{[89, Tab. 1]}. An orbital path must save more terrestrial detour than its relay adds.

在存取開銷相同的條件下，若地面服務於測試範圍全面占優，即達到停止條件。遠端使用者仍是候選情境，Bose 等人量得 Mozambique CDN 延遲為 138.7 ms 對 7.2 ms^{[89, Tab. 1]}。軌道路徑需節省超過中繼成本的地面繞行，才能取得優勢。

E · Why nowE · 為什麼是現在

Published latency measurements and inference methods support immediate ground-baseline collection. Directions A, C and D later supply calibrated orbital inputs.

公開延遲量測與推論方法，可立即支持地面基準資料蒐集，方向 A、C、D 再提供校準後的軌道輸入。

13.7Research direction F — Emulation and the 2027 two-satellite test研究方向 F — 模擬與 2027 年雙星測試

F · The gap and the document that exposes itF · 缺口與揭露它的文獻

Orbital-cluster emulation still needs a calibrated link model. Cionca et al. find measurement gaps across StarryNet, OpenSN and Celestial, including Celestial's ISL representation ^[81]. Measured beam quality and jitter would then constrain the model.軌道叢集模擬仍需校準鏈路模型。Cionca 等人指出 StarryNet、OpenSN 與 Celestial 和量測間的落差，其中包含 Celestial 的 ISL 表示^[81]。

F · The falsifiable questionF · 可檢驗的問題

Can acquisition time, delivered rate versus range and thermal-cycle pointing jitter calibrate availability predictions within a declared error bound? Test whether short range improves on the long-range setup baseline. Register best-case and degraded-model predictions before launch so that observations discriminate between models.

擷取時間、送達速率對距離，以及熱循環下的指向抖動，能將可用度預測校準至明列誤差內嗎？測試需比較短距與長距建立基準，並在發射前登錄最佳情況及劣化模型的預測，讓觀察可區分模型。

F · Method with today's data and toolsF · 用今日資料與工具的方法

Build a link-level trace emulator sharing geometry and rate models with Directions A and B. Represent setup time, range-dependent rate and jitter-driven outages, then replay telemetry to measure prediction error. First prepare measurements of rate at every visited separation, command-to-first-frame acquisition time, jitter against panel or radiator temperature, and thermal-cycle availability. This shorter-range geometry motivates a test of faster acquisition.建立與方向 A、B 共用幾何及速率模型的鏈路層軌跡模擬平台，表示建立時間、距離相依速率及抖動造成的中斷，再以遙測重播量測預測誤差。先準備各實際間距的速率、指令至第一幀的擷取時間、抖動對面板或散熱器溫度，以及熱循環可用度。

F · Metrics and baselinesF · 指標與基準

Report held-out availability and rate errors, plus the acquisition-time distribution. Baselines are modern long-range terminals ^{[33, Tab. 1]}, demonstrated TBIRD pointing and Starlink's daily acquisition frequency. Their operating conditions define each comparison.

報告保留資料的可用度及速率誤差，以及擷取時間分布。基準為現代長距終端^{[33, Tab. 1]}、TBIRD 實證指向及 Starlink 每日擷取頻率，各項操作條件界定比較範圍。

F · What would falsify itF · 停止研究的條件

A fixed separation permits point calibration, leaving the range curve open. Slow acquisition at short range narrows Direction A to fixed fabrics. If flight telemetry remains proprietary, use bench measurements and retain the emulator's status as a testable hypothesis.

固定間距可做單點校準，距離曲線則仍待檢驗。短距擷取若仍耗時，方向 A 便集中於固定織網。若飛行遙測維持保密，則採實驗室量測，並保留模擬平台作為可檢驗假設的定位。

F · Why nowF · 為什麼是現在

The early-2027 schedule sets the deadline for the measurement list and registered predictions. Prepare them before the flight test configuration is fixed.

2027 年初時程設定了量測清單與預先登錄預測的期限，應在飛行測試配置確定前備妥。

13.8Priorities, dependencies and a 24-month plan優先順序、相依關係與 24 個月計畫

The planned pair can calibrate a short optical link's range, acquisition and thermal-pointing behavior. Cluster fabrics, collectives, ground stations and user paths require additional models. Start F before launch, then use its calibration in B and A. Run C in parallel using beam tests and prototype telemetry. D needs C's checkpoint sizes, while E needs A's step model, C's reliability and D's ground segment. Ground-baseline collection for E can begin immediately, and Table 13.2 assigns deliverables and stopping criteria across the plan.

預定雙星可校準短距光鏈路的距離、擷取及熱指向行為，叢集織網、集體通訊、地面站與使用者路徑則需要額外模型。F 在發射前啟動，再將校準提供 B 與 A，C 以質子束測試及原型遙測平行進行。D 需要 C 的檢查點大小，E 則需要 A 的步驟模型、C 的可靠度及 D 的地面段。E 的地面基準資料可立即蒐集，表 13.2 再安排各期交付物及停止條件。

Table 13.2. The research plan assigns deliverables, required data and stopping criteria around the planned optical test. The stated decision criteria make each deliverable testable.

表 13.2。研究計畫圍繞預定光學測試，安排交付物、所需資料及停止條件，讓各項產出具備可檢驗的決策標準。

Calibration depends on telemetry release because the Suncatcher link parameters come from Google's paper. Proprietary flight data would leave A, B and F as models awaiting direct validation. These conditions could change the assumed cluster altitude before the plan ends.Suncatcher 鏈路參數來自 Google 論文，因此校準依賴遙測公開。飛行資料若維持保密，A、B 與 F 便保留為等待直接驗證的模型。這些條件可能在計畫結束前改變假定叢集高度。

Papers behind this chapter本章所依據的論文