LLMSpace: Carbon Footprint Modeling for Large Language Model Inference on LEO SatellitesLLMSpace：LEO 衛星上大型語言模型推論的碳足跡建模

arXiv 2605.05615Jiang, Ildefonso, Loveless, Chen (Indiana University) · paper論文 preprint

LLMSpace calculates inference-satellite carbon from component masses, launch intensity and hardware lifetime. Solar power makes operational carbon small within the model, while manufacture and launch dominate. Mass, link assumptions and radiation lifetime determine how readers should interpret the comparison.LLMSpace 從元件質量、發射碳強度及硬體壽命計算推論衛星的碳排放。模型中的太陽能使營運碳相對較小，製造與發射則占主要部分。比較結果取決於質量、鏈路及輻射壽命的假設。

≈ 9 min · 5-question quiz≈ 9 分鐘 · 5 題測驗Chapter 12 ↗第 12 章 ↗Jump to the quiz ↓跳到測驗 ↓

If you read nothing else

只讀這一段也可以

LLMSpace is the first carbon model for language-model inference on a LEO satellite, built bottom-up from components rather than measured ^{[97, §4–5]} preprint. The main caveat for this primer is that the network appears only as mass and carbon. Thus, the paper leaves request deadlines outside its evaluated scope.LLMSpace 是第一個針對 LEO 衛星上語言模型推論的碳模型，由元件由下而上建構，量測則仍待完成^{[97, §4–5]}預印本。對本教材而言主要的限制是：網路只以質量與碳的形式出現，所以論文將從軌道服務的請求期限列於評估範圍之外。

Measure or comparison參數或比較項目	Values, conditions and evidence數值、條件與證據
Evaluation conditions	The DGX-H100-class payload has eight GPUs and weighs about 130 kg. The 4 nm COTS variant draws 10 kW and has a two-year radiation-limited life. The 14 nm rad-hard variant draws 20 kW and lasts ten years. The model combines this with a 191 kg Starlink-V1 bus, leaving the integration unresolved. ^{[97, §4.1, §5]}
Power budget	Each laser ISL terminal is specified as "up to 200 Gbps at ~1 W optical power", about 28 kg and 150 kgCO2e. Ground links exceed 96/32 Gbit/s. Latency lies outside the model. Its orbit is a 90–100 min period with a 30–35 min eclipse and batteries sized at 80% depth of discharge ^{[97, §4.1]} .
Power budget	Embodied carbon is 16.57 tCO2e for COTS and 30.45 tCO2e for rad-hard. Falcon 9 launch intensity is 14.5 kgCO2e per kilogram. Solar power makes orbital operational carbon "effectively negligible" in the authors' model. ^{[97, Tab. 6, Fig. 2]}
評估條件	它的酬載是約 130 kg、八張 GPU 的 DGX-H100 等級系統，可以是 10 kW、壽命兩年（受輻射限制）的 4 nm COTS 版本，也可以是 20 kW、壽命十年的 14 nm 抗輻射版本，搭載在取為 191 kg 的 Starlink-V1 平台上，而論文將其保留為待釐清這個組合^{[97, §4.1, §5]}。
功率預算	它的鏈路是「在約 1 W 光功率下最高 200 Gbps」的雷射 ISL，每具終端約 28 kg、約 150 kgCO2e，地面鏈路則高於 96/32 Gbit/s，延遲位於模型範圍之外。
軌道週期	軌道週期 90–100 分鐘、日蝕 30–35 分鐘，電池以 80% 放電深度設計^{[97, §4.1]}。
碳排估算	結果是 COTS 衛星的隱含碳為 16.57 tCO2e、抗輻射衛星為 30.45 tCO2e，發射碳強度取自 Falcon 9 的每公斤 14.5 kgCO2e，而作者稱在軌營運碳「實際上可忽略」，因為電力來自太陽^{[97, Tab. 6, Fig. 2]}。

1. Where this paper sits1. 這篇論文的位置

Chapter 12 treats SpaceMoE, CollabLLM-LEO and LLMSpace as the published orbital-inference core. LLMSpace focuses on lifecycle carbon. SpaceMoE and CollabLLM-LEO study distributed inference latency. LLMSpace compares lifecycle carbon for an inference satellite with the same compute on a terrestrial grid. This is the environmental half of the argument for orbital data centers that Chapter 11 introduces through Suncatcher's power-cost comparison.第 12 章稱 SpaceMoE、CollabLLM-LEO 與 LLMSpace 為在軌推論文獻已發表的核心，而 LLMSpace 問的問題與另外兩篇各異。SpaceMoE 與 CollabLLM-LEO 問的是推論分散到多顆衛星時要花多久，LLMSpace 問的則是一顆推論衛星在生命週期中要付出多少碳，並比較其碳排放與地面電網執行相同運算的差距。這是軌道資料中心論證中環境面的那一半，而第 11 章透過 Suncatcher 的電力成本比較引入了經濟面。它也是這組論文中假設酬載最大的一篇，H100 等級的系統只飛過一次，就是 Starcloud-1，屬於公司宣布的展示、公開遙測資料仍待補充。它假設的 ISL 也最強，是最高 200 Gbps 的終端。

Measure or comparison參數或比較項目	Values, conditions and evidence數值、條件與證據
Evaluation conditions	The study assumes the largest payload in this set, an H100-class system, and the highest ISL rate, up to 200 Gbps. Starcloud-1 has flown an H100 in a company-announced demonstration. Public telemetry remains pending. Thus, it marks the far end of the hardware range that Chapter 12 lays out from SpaceMoE's 7.28 GFLOPS board through CollabLLM-LEO's Orin modules.
評估條件	因此它標出了第 12 章所鋪陳硬體範圍的最遠端，從 SpaceMoE 的 7.28 GFLOPS 板子，經 CollabLLM-LEO 的 Orin 模組，一路到這裡。

2. The problem, measured2. 問題，用量測說話

Environmental-benefit claims require explicit carbon accounting. Solar electricity addresses operational emissions, while satellite manufacture and launch add embodied emissions. The embodied carbon covers manufacturing and launch, while operational carbon covers electricity use. The authors model both from component inputs. ^{[97, §4]} preprint問題在於，關於在軌運算環境效益的說法需要明確的碳盤查，而最直覺的一句話論證，「陽光使用階段的排放為零」，漏掉了建造與發射衛星的碳。因此作者把隱含碳（embodied carbon），也就是製造與發射硬體所產生的排放，與營運碳，也就是硬體所用電力的排放，區分開來，兩者都由元件建模^{[97, §4]}預印本。

Measure or comparison參數或比較項目	Values, conditions and evidence數值、條件與證據
Evaluation conditions	The baseline bus is Starlink-V1 at 191 kg per satellite, citing Aili et al. 2025.
DGX H100 payload	The main compute platform is an NVIDIA DGX H100 with eight GPUs at about 130 kg
Alternative payloads	Alternative payloads include a roughly 2 kg Jetson, a 30 kg one-to-two-GPU system, and 130 kg DGX A100/B100 systems. ^{[97, §4.1, §5]}
Power budget	The COTS variant uses 4 nm hardware at 10 kW. Total ionizing dose, displacement damage and single-event effects limit its assumed life to two years. A rad-hard 14 nm system at 20 kW is given ten years, so the rad-hard design buys life with power and mass.
評估條件	基準平台是每顆 191 kg 的 Starlink-V1^[105]，主要運算平台是約 130 kg、八張 GPU 的 NVIDIA DGX H100，另外也考慮約 2 kg 的 Jetson、約 30 kg 的一到兩張 GPU 系統，以及約 130 kg 的 DGX A100/B100^{[97, §4.1, §5]}。
評估條件	兩個版本的差異在於處理輻射的方式：10 kW 的 COTS 4 nm 系統被賦予兩年壽命，受總游離劑量、位移損傷與單事件效應限制。
評估條件	20 kW 的抗輻射 14 nm 系統則有十年，所以抗輻射設計是以電力與質量換取壽命。

3. The design3. 設計

抗輻射版本功率加倍時兩者都加倍，這就是抗輻射衛星的隱含碳幾乎是兩倍的原因（見計算表）。

Measure or comparison參數或比較項目	Values, conditions and evidence數值、條件與證據
Model configuration	Each subsystem adds manufacturing carbon and mass-dependent launch carbon. Falcon 9 supplies 3.3 × 10⁵ kgCO2e for 22,800 kg, giving 14.5 kgCO2e per kilogram. ^{[97, Tab. 6]}
Orbital period	The subsystems are the solar array, the battery, the cooling panel, the computing hardware, and the network and satellite bus. The orbit has a 90–100 min period, 55–65 min sunlight and 30–35 min eclipse. It repeats 15–16 times daily. Batteries cover eclipse at 80% depth of discharge. ^{[97, §4.1]}
Model configuration	That orbit model is where the physics of Chapter 8 enters. A 10 kW payload needs 5.0–5.8 kWh through eclipse. At 80% depth of discharge, battery capacity is about 6.3–7.3 kWh. The rad-hard variant doubles power and both energy requirements. Thus, the rad-hard satellite carries almost twice the embodied carbon (calc table).
模型配置	模型把每個子系統的製造碳與發射碳相加，發射碳是質量乘以取自 Falcon 9 的每公斤 14.5 kgCO2e，即每 22,800 kg 入軌 3.3 × 10⁵ kgCO2e^{[97, Tab. 6]}。
星座配置	子系統包括太陽能陣列、電池、散熱板、運算硬體，以及網路與衛星平台，它們的大小由軌道決定：週期 90–100 分鐘，日照 55–65 分鐘、日蝕 30–35 分鐘，每天 15–16 圈，電池容量要能以 80% 放電深度撐過日蝕^{[97, §4.1]}。
模型配置	第 8 章的物理正是從這個軌道模型進入：10 kW 的酬載跨越一次日蝕需要 5.0–5.8 kWh，在 80% 放電深度下需要約 6.3–7.3 kWh 的電池。

The network enters in the same currency. Latency, topology and traffic models lie outside the link description. This is the gap a networking reader notices first.網路也以同樣的單位進入模型。CubeSat 的 RF 地面鏈路則為 kbit/s 到數百 Mbit/s^{[97, §4.1]}。這些鏈路將延遲、拓樸與流量列於模型範圍之外，這是網路讀者最先注意到的缺口。採用多接面陣列與抗輻射電池的「rad-opt」設計，則比抗輻射版本減少 7% 的隱含碳^{[97, §6.1]}。

Measure or comparison參數或比較項目	Values, conditions and evidence數值、條件與證據
Power budget	The Starlink V2-attributed laser terminal is specified as "up to 200 Gbps at ~1 W optical power". It weighs about 28 kg and adds about 150 kgCO2e of manufacturing carbon. Ground links exceed 96/32 Gbit/s down/up, with EIRP up to 1.75 kW. CubeSat RF links range from kbit/s to hundreds of Mbit/s. ^{[97, §4.1]}
Model configuration	Launching a 28 kg terminal adds about 0.41 tCO2e, nearly three times its manufacturing carbon. Terminal mass therefore dominates its modeled carbon.
Energy	Solar electricity makes modeled orbital operational carbon "effectively negligible". Terrestrial comparisons use 20 gCO2e/kWh for clean power and 380 gCO2e/kWh for dirty power. Rad-hard variants fall between these cases. A "rad-opt" design uses a multi-junction array and rad-hard battery, reducing embodied carbon by 7% relative to rad-hard. ^[97] ^{[97, §6.1]}
功率預算	雷射 ISL 終端取為「在約 1 W 光功率下最高 200 Gbps」，約 28 kg、製造碳約 150 kgCO2e，出處歸於 Starlink V2。
容量比較	高容量地面鏈路為下行／上行超過 96/32 Gbit/s，EIRP 最高 1.75 kW。
模型配置	以論文的碳強度發射一具 28 kg 的終端約需 0.41 tCO2e，幾乎是其製造碳的三倍，所以在這個模型中終端的碳主要由質量決定。
評估條件	營運碳接著由能源來源計算，在軌道上是太陽能陣列，因此「實際上可忽略」，並與乾淨電網 20 gCO2e/kWh 與高碳電網 380 gCO2e/kWh 的地面電網比較，抗輻射版本落在兩者之間^[97]。

Figure A. Left, the model combines a roughly 130 kg eight-GPU payload with a 191 kg Starlink-V1 bus. Each laser terminal adds about 28 kg. Right, the calculation spreads satellite embodied carbon over its assumed lifetime. The reference is annual grid carbon from a 10 kW ground system. The orbit lands between the clean and dirty grids, as the paper's Fig. 2a reports for the rad-hard variants. The COTS satellite's short radiation-limited life is what makes it costlier per year than the rad-hard one.圖 A。左：LLMSpace 所組合的質量，約 130 kg 的八 GPU 酬載搭載在 191 kg 的 Starlink-V1 平台上，每具雷射終端約 28 kg。右：對論文總量的粗略解讀，把每顆衛星的隱含碳攤到其所述壽命上，並與在地面以 10 kW 運轉一年的電網碳並列。軌道落在乾淨與高碳電網之間，與論文 Fig. 2a 對抗輻射版本的報告一致。而 COTS 衛星受輻射限制的短壽命，正是它每年成本比抗輻射版本更高的原因。Sources來源: ^{[97, §4.1, §5, Tab. 6, Fig. 2]}

Original figure note原始圖說出處

Drawn for this primer from Jiang et al., arXiv 2605.05615 v2, §4.1, §5, Table 6 and Fig. 2. The per-year values are computed in the table in Section 5.

4. Evaluation setup4. 實驗設定

The bottom-up analytical model is checked against a reference assembled from prior work. Measured lifecycle validation, traffic simulation, testbeds and flight hardware lie outside its scope. The "up to 200 Gbps" rate affects only terminal mass, power and manufacturing carbon. ^{[97, §5]} Latency, topology and failures lie outside the model. Thus, the number of terminals per satellite and their pointing remain outside the result calculations.

Period and eclipse fraction represent the orbit. They size batteries and arrays, with propagated constellation geometry and time-varying load outside the model. preprint

評估是一個由下而上的解析模型，驗證的對象是作者根據先前文獻建立的參考值，實測生命週期資料、流量模擬、測試平台與飛行硬體則位於評估範圍之外^{[97, §5]}預印本。ISL 速率寫為「最高 200 Gbps」，但它只透過終端的質量、功率與製造碳進入模型，延遲位於模型範圍之外、拓樸及故障模型皆位於範圍之外，所以每顆衛星的終端數與其指向皆位於結果的計算範圍之外。軌道以週期與日蝕比例表示，星座傳播動態則位於模型範圍之外，這固定了電池與陣列的大小，卻遺漏了任何隨時間變化的負載。

Measure or comparison參數或比較項目	Values, conditions and evidence數值、條件與證據
Model configuration	The hardware is the DGX-H100 class at 10 kW for COTS and 20 kW for rad-hard on a Starlink-V1 bus of 191 kg.
Carbon reference	The reference comparison is 18.3 tCO2e for COTS and 32.5 tCO2e for rad-hard
Carbon reduction	against which the model lands 9.4% and 6.3% lower ^{[97, Tab. 6]}.
評估條件	硬體是 DGX-H100 等級，COTS 為 10 kW、抗輻射為 20 kW，搭在 191 kg 的 Starlink-V1 平台上。
模型配置	參考比較值為 COTS 18.3 tCO2e、抗輻射 32.5 tCO2e，模型各自低了 9.4% 與 6.3%^{[97, Tab. 6]}。

5. What was measured5. 量到了什麼

Claim宣稱	Setup and caveat設定與附註
Embodied carbon of 16.57 tCO2e for the COTS satellite and 30.45 tCO2e for the rad-hard one ^{[97, Tab. 6]}.COTS 衛星的隱含碳為 16.57 tCO2e，抗輻射衛星為 30.45 tCO2e^{[97, Tab. 6]}。	Modelled, unmeasured. Within 9.4% and 6.3% of a reference built from prior work. The safe quotation per fact base Part C, item 40.依據為元件模型，實測則仍待補充。與根據先前文獻建立的參考值相差 9.4% 與 6.3% 以內。依事實庫 Part C 第 40 項為可安全引用的數字。
Launch intensity of 14.5 kgCO2e per kilogram ^{[97, Tab. 6]}.發射碳強度為每公斤 14.5 kgCO2e^{[97, Tab. 6]}。	From Falcon 9, 3.3 × 10⁵ kgCO2e per 22,800 kg. A different vehicle or a reusable upper stage would change it.取自 Falcon 9，每 22,800 kg 為 3.3 × 10⁵ kgCO2e。相異的運載器或可重複使用的上節都會改變它。
Operational carbon in orbit is "effectively negligible". Rad-hard variants fall between grids at 20 and 380 gCO2e/kWh ^[97].在軌營運碳「實際上可忽略」。抗輻射版本落在 20 與 380 gCO2e/kWh 的電網之間^[97]。	Solar power treated as carbon-free once the array is built. The comparison is of carbon only, with latency and throughput outside the comparison.陣列建成後太陽能被視為零碳。比較的只有碳，延遲及吞吐量位於比較範圍之外。
A "rad-opt" design cuts embodied carbon by 7% against rad-hard ^{[97, §6.1]}.「rad-opt」設計比抗輻射版本減少 7% 的隱含碳^{[97, §6.1]}。	Multi-junction array and rad-hard battery. Depends on the rad-hard overheads that the paper states inconsistently.多接面陣列與抗輻射電池。取決於論文前後說法有差異的抗輻射額外負擔。

Question問題	Formula公式	Result結果
What energy must the battery supply through one eclipse?電池在一次日蝕中必須提供多少能量？	\( E = P\,t_{\text{ecl}},\quad C = E/0.8 \)	10 kW: 5.0–5.8 kWh → 6.3–7.3 kWh · 20 kW: 10.0–11.7 kWh → 12.5–14.6 kWhcomputed. 30–35 min eclipse and 80% DoD from §4.1. Payload power only計算值。30–35 分鐘日蝕與 80% DoD 取自 §4.1。只計酬載功率
How much heavier in carbon is the rad-hard satellite?抗輻射衛星的碳重多少？	\( 30.45 / 16.57 \)	1.84×computed from Table 6由 Table 6 計算
Embodied carbon per year of life?每年壽命的隱含碳？	\( 16.57/2,\qquad 30.45/10 \)	≈ 8.3 · ≈ 3.0 tCO2e/yrcomputed. Lifetimes from §4.1計算值。壽命取自 §4.1
Grid carbon of running 10 kW for a year on the ground?在地面以 10 kW 運轉一年的電網碳？	\( 10\ \text{kW} \times 8{,}760\ \text{h} \times \{20, 380\}\ \text{g/kWh} \)	≈ 1.75 · ≈ 33.3 tCO2e/yrcomputed. Grid intensities from Fig. 2a. Excludes ground hardware's embodied carbon and cooling overhead計算值。電網碳強度取自 Fig. 2a。計算範圍排除地面硬體的隱含碳與冷卻負擔
Launch carbon of one laser terminal?一具雷射終端的發射碳？	\( 28\ \text{kg} \times 14.5\ \text{kgCO}_2\text{e/kg} \)	≈ 0.41 tCO2ecomputed. ≈ 2.7× the ≈ 150 kgCO2e to manufacture it ^{[97, §4.1]}計算值。約為其製造碳約 150 kgCO2e 的 2.7 倍^{[97, §4.1]}
What share of the bus mass is the payload?酬載佔平台質量多少？	\( 130 / 191 \)	68%computed. The two masses remain unresolved in the paper計算值。論文將其保留為待釐清這兩個質量

6. Limits and open questions6. 限制與待解的問題

Carbon with request timing outside its scope. The links appear as mass and carbon only, with latency, topology and failure outside the model, so request deadlines lie outside its evaluation. The ground comparison uses grid carbon per kilowatt-hour. An equivalent serving design with matched tokens, users and latency requires separate evaluation.聚焦於碳，請求時序則位於範圍之外。鏈路只以質量與碳出現，延遲、拓樸及故障模型皆位於範圍之外，所以請求期限位於論文評估範圍之外。它與地面電網的比較是每千瓦時碳的比較，若要與地面端服務方案比較，須另行確保使用者、token 交付量及延遲條件相同。
Energy modeled as an average orbit, with orbital dynamics outside its scope. The period and eclipse fraction size the battery, but load is constant and thermal behavior lies outside the model. The model excludes measured duty cycling. Xing et al. use a 5–10 W board, while the Suncatcher prototype reportedly operates for 15–20 minutes, a company statement. Chapter 13 asks how joint energy and orbital dynamics affect operation.能源以平均軌道建模，軌道動態則位於模型範圍之外。週期與日蝕比例決定了電池大小，但負載是常數、熱行為位於模型範圍之外，而 Xing 等人在 5–10 W 板子上量到的工作週期，以及據報 Suncatcher 原型以 15–20 分鐘運作的工作週期公司皆位於範圍之外，所以能源與軌道動態的聯合建模仍待完成，而這正是第 13 章提出的聯合問題。
Unreconciled mass and overheads. A 130 kg payload on a 191 kg bus leaves 61 kg. That remainder must cover 10–20 kW arrays and batteries plus 28 kg terminals. Chip-area overhead is 15× in §5, 2× in the §4.1 RAD750 example, and 10×/20× in §6. Power overhead is 2× versus 1.5×/3× (fact base Part C, item 40). Only the embodied totals are safe to quote.仍待釐清的質量與額外負擔。130 kg 的酬載放在 191 kg 的平台上，只剩 61 kg 給其他所有東西，包括 10–20 kW 所需的陣列與電池，以及每具 28 kg 的終端。抗輻射晶片面積額外負擔在 §5 為 15 倍、在 §4.1 的 RAD750 例子為 2 倍、在 §6 為 10/20 倍，功率額外負擔則為 2 倍對 1.5/3 倍（事實庫 Part C，第 40 項）。只有隱含碳總量可以安全引用。
Lifetime assumptions drive the answer. The assumed two-year COTS life raises annual carbon above the rad-hard case. Suncatcher tests find HBM irregularities only from about 2 krad. One chip shows zero hard TID failures up to 10 krad(Si), or 15 krad(Si) in arXiv v2. The five-year requirement is 750 rad. COTS life beyond two years could halve annual carbon or improve it further. ^{[108, §2.3]} ^[36]壽命假設主導答案。COTS 衛星受輻射限制的兩年壽命讓它每年的成本比抗輻射版本高，然而 Suncatcher 的質子測試發現 HBM 要到約 2 krad 才出現異常、單顆晶片直到 10 krad(Si) 觀測到的硬性 TID 失效為零（arXiv v2 預印本寫 15 krad(Si)^{[108, §2.3]}），而五年的需求是 750 rad^[36]，這意味著 COTS 加速器可能活超過兩年，屆時每年的碳會減半甚至更少。
Inputs awaiting primary-source evidence. The "up to 200 Gbps at ~1 W" terminal is attributed to Starlink V2, while SpaceX keeps laser rate proprietary. The H100 has one company-announced flight demonstration, with public telemetry pending. The 191 kg bus mass comes from a secondary citation. Thus, the network and payload inputs are assumptions rather than measurements.仍待一手來源的輸入。「在約 1 W 下最高 200 Gbps」的終端被歸於 Starlink V2，但 SpaceX 從仍待公布其雷射速率。酬載所代表的 H100 只飛過一次，屬於公司宣布的展示、公開遙測仍待補充。191 kg 的平台質量則來自二手引用，所以網路與酬載的輸入都是模型假設，量測則須另行確認。

7. How it connects7. 與其他內容的關聯

Section 12.4: the chapter's treatment of LLMSpace alongside SkyMemory, Constella and the cached-model framework.第 12.4 節：該章對 LLMSpace 的討論，與 SkyMemory、Constella 及快取模型框架並列。
Section 12.2 and Section 12.3: SpaceMoE and CollabLLM-LEO, the two papers that model time with carbon outside their scope.第 12.2 節與第 12.3 節：SpaceMoE 與 CollabLLM-LEO，兩篇模擬時間碳則位於模型範圍之外的論文。
Section 12.5: LLMSpace's 200 Gbps terminal is the top rung of the ISL ladder whose sensitivity Panel 12.1 explores.第 12.5 節：LLMSpace 的 200 Gbps 終端是面板 12.1 探討其敏感度的 ISL 階梯的最上層。
Section 12.6 compares ground and orbital serving sites. Energy sources and latency jointly determine the comparison.第 12.6 節：地面與軌道作為兩個服務地點，地面電網上的能源與日照軌道上的能源必須與延遲一起權衡。
Section 10.6: a single DGX-class satellite is the single-satellite regime, where inference is self-contained and the ISL matters least.第 10.6 節：單顆 DGX 等級衛星屬於單衛星區間，推論自成一體，ISL 的影響相對較小。
Section 13.6: research direction E combines the time model of the first two papers with the energy accounting of this one.第 13.6 節：研究方向 E 把前兩篇的時間模型與本篇的能源核算結合起來。
SpaceMoE [78] and CollabLLM-LEO [96] are the sibling digests. Suncatcher [36] supplies the radiation results and the power-cost argument this paper's lifetimes and carbon totals should be read against. Xing et al. [85] is the only flight measurement of the thermal and battery behavior the model abstracts away.SpaceMoE [78]與 CollabLLM-LEO [96]是姊妹導讀。Suncatcher [36]提供了應該拿來對照本文壽命與碳總量的輻射結果與電力成本論證。Xing 等人 [85]是對本模型所抽象掉之熱與電池行為唯一的飛行量測。

Key numbers關鍵數字

Number	數字	Meaning (setup)
≈ 130 kg · 8 GPUs	DGX-H100-class payload ^{[97, §4.1]}	DGX-H100 等級酬載^{[97, §4.1]}
10 kW / 2 yr · 20 kW / 10 yr	COTS 4 nm vs rad-hard 14 nm power and life ^{[97, §4.1, §5]}	COTS 4 nm 與抗輻射 14 nm 的功率與壽命^{[97, §4.1, §5]}
191 kg	Starlink-V1 bus mass, unreconciled with the payload ^{[97, §4.1]}	Starlink-V1 平台質量，與酬載仍待釐清^{[97, §4.1]}
≤ 200 Gbps · ≈ 1 W · ≈ 28 kg · ≈ 150 kgCO2e	Laser ISL terminal rate, optical power, mass, manufacturing carbon ^{[97, §4.1]}	雷射 ISL 終端速率、光功率、質量、製造碳^{[97, §4.1]}
> 96/32 Gbit/s	High-capacity ground link, down/up ^{[97, §4.1]}	高容量地面鏈路，下行／上行^{[97, §4.1]}
90–100 min · 30–35 min · 80%	Orbit period, eclipse, battery depth of discharge ^{[97, §4.1]}	軌道週期、日蝕、電池放電深度^{[97, §4.1]}
16.57 / 30.45 tCO2e	Embodied carbon, COTS / rad-hard satellite ^{[97, Tab. 6]}	隱含碳，COTS／抗輻射衛星^{[97, Tab. 6]}
14.5 kgCO2e/kg	Launch intensity from Falcon 9 ^{[97, Tab. 6]}	取自 Falcon 9 的發射碳強度^{[97, Tab. 6]}