Starcloud Targets $100 Billion Space AI Network to Undercut Terrestrial Compute Costs

Sequoia Capital

The company plans an 88,000-satellite constellation to bypass land and battery constraints on Earth, arguing that plunging launch costs will soon make orbit the cheapest location for AI inference.

AI-assisted summary of Starcloud's Philip Johnston: Why the Cheapest Compute Will Be in Space — original source. Check the original for full context.

Source published: 2026-05-06

Detailed notes

## Background Published on YouTube by Sequoia Capital on May 06, 2026, this video features Starcloud co-founder and CEO Philip Johnston at AI Ascent 2026 making the economic and technical case for orbital AI data centers. The page description outlines how space-based compute can become cheaper than terrestrial alternatives, previews an ambitious 88,000-satellite constellation plan, and explains how the company successfully trained the first model in space.

## TL;DR Philip Johnston presents Starcloud's breakthrough in successfully running and training AI models in space using Nvidia H100 GPUs, defying previous industry skepticism regarding radiation and thermal management. He details the superior economics of space-based solar energy, which yields eight times more power per square meter than on Earth and eliminates land and battery costs, reaching a financial break-even once launch costs fall below $500/kg via vehicles like SpaceX's Starship. To capitalize on this, Starcloud is planning an SEC-filed 88,000-satellite constellation to deliver 20 gigawatts of globally accessible, low-latency inference capacity. During the Q&A session, Johnston addresses technical solutions for heat dissipation using the Stefan-Boltzmann law, mitigations for Kessler syndrome via low-altitude orbits, and extensive radiation testing protocols.

## Atmosphere & style The presentation is energetic, forward-looking, and intellectually stimulating, kicking off with a lighthearted joke about 'abusing GPUs' before transitioning into a fast-paced, high-stakes pitch. Johnston exudes grounded confidence and engineering pragmatism, deftly balancing wild, sci-fi-esque visions of Dyson-sphere-level infrastructure with precise 'napkin math' formulas and rigorous testing data. The audience engagement feels sharp and highly technical, creating a collaborative, visionary atmosphere where profound skepticism about the harsh realities of space is systematically met with practical, optimistic solutions.

## Key takeaways - [01:01](https://www.youtube.com/watch?v=94b6i5jI1nE&t=61s) Star Cloud successfully proved that terrestrial, data-center-grade GPUs can operate in space by deploying Star Cloud 1, which ran nanoGPT, Gemini, and high-powered inference on SAR data. - [01:01](https://www.youtube.com/watch?v=94b6i5jI1nE&t=61s) The primary technical skepticism surrounding space-based GPUs centered on thermal dissipation in a vacuum and radiation-induced bit flips. - [01:43](https://www.youtube.com/watch?v=94b6i5jI1nE&t=103s) Building solar projects on Earth involves three primary cost drivers that do not apply or are vastly minimized in space: permitted land costs, battery storage for night hours, and solar cell manufacturing. - [01:43](https://www.youtube.com/watch?v=94b6i5jI1nE&t=103s) A single square meter of solar panel in space generates eight times the energy of the same panel on Earth because the sun's radiation is entirely unfiltered by the atmosphere. - [02:49](https://www.youtube.com/watch?v=94b6i5jI1nE&t=169s) Space-based data centers achieve economic break-even with Earth when satellite launch costs drop below $500 per kilogram. - [02:49](https://www.youtube.com/watch?v=94b6i5jI1nE&t=169s) SpaceX's Starship is expected to easily beat the space compute break-even threshold with its target launch costs of just $10 to $20 per kilogram. - [03:18](https://www.youtube.com/watch?v=94b6i5jI1nE&t=198s) Star Cloud has filed with the SEC for a massive constellation of 88,000 satellites designed to deliver 20 gigawatts of global inference capacity. - [03:18](https://www.youtube.com/watch?v=94b6i5jI1nE&t=198s) The planned 88,000-satellite constellation will operate in a dawn-dusk sun-synchronous orbit to maintain continuous, 24/7 exposure to sunlight. - [03:18](https://www.youtube.com/watch?v=94b6i5jI1nE&t=198s) The space constellation will leverage optical links to deliver sub-50 millisecond latency to users anywhere across the globe. - [03:18](https://www.youtube.com/watch?v=94b6i5jI1nE&t=198s) At a projected capex of $100 billion, Star Cloud's orbital network is estimated to be cheaper than building an equivalent 20-gigawatt compute infrastructure on Earth. - [04:43](https://www.youtube.com/watch?v=94b6i5jI1nE&t=283s) According to the Stefan-Boltzmann equation, raising chip operating temperatures by just 10% in Kelvin (from 50°C to 80°C) can nearly halve the required radiator surface area. - [04:43](https://www.youtube.com/watch?v=94b6i5jI1nE&t=283s) Star Cloud is partnering with Nvidia on a space-grade chip called 'Ruben 1' that is architected to run hotter to minimize radiator mass. - [06:40](https://www.youtube.com/watch?v=94b6i5jI1nE&t=400s) Kessler syndrome risks are mitigated for Star Cloud's initial satellites by flying at a low altitude of ~400 km, where atmospheric drag ensures natural deorbiting within months. - [06:40](https://www.youtube.com/watch?v=94b6i5jI1nE&t=400s) Visual depictions of orbital congestion are heavily distorted because tracking maps render satellite dots to be roughly the geographic width of California. - [08:36](https://www.youtube.com/watch?v=94b6i5jI1nE&t=516s) To harden chips against space radiation, Star Cloud utilizes the Knoxville cyclotron for proton testing and Brookhaven National Lab to simulate five years of heavy ion exposure in 24 hours. - [09:23](https://www.youtube.com/watch?v=94b6i5jI1nE&t=563s) Star Cloud is focusing exclusively on AI inference rather than training, as inference is projected to eventually command 99% of the total AI compute market. - [09:23](https://www.youtube.com/watch?v=94b6i5jI1nE&t=563s) A hypothetical 5-gigawatt space training structure would span 4 km by 4 km and feature a massive 1 km by 4 km radiator, but such a project is at least 15 years away.

## Detailed walkthrough ### [00:01](https://www.youtube.com/watch?v=94b6i5jI1nE&t=1s) Introduction and the Star Cloud 1 Milestone

Philip Johnston, the co-founder and CEO of Star Cloud, opens his presentation with a lighthearted confession: much like the previous presenters, his company has "also been abusing GPUs in ways they were not designed for." Star Cloud's mission is to build data centers in space, primarily to harness the abundant solar energy available there. Johnston promises to spend five minutes explaining why space-based data centers will soon make far more sense than terrestrial ones, followed by five minutes of audience questions.

To kick things off, Johnston shares a dramatic deployment video of Star Cloud 1, a satellite equipped with five Nvidia GPUs, most notably the state-of-the-art Nvidia H100 chip. He jokes that getting such a clear deployment video is rare, as satellites are often launched "into the shadow." The success of Star Cloud 1 was a massive milestone because the industry widely believed that terrestrial, data-center-grade GPUs could not survive in space due to extreme thermal dissipation challenges and radiation-induced bit flips. Star Cloud proved the skeptics wrong by successfully training Andrej Karpathy’s nanoGPT model in space, running a version of Google's Gemini, and performing high-powered inference on Synthetic Aperture Radar (SAR) satellite data.

- **Key Hardware:** Star Cloud 1 carried five Nvidia GPUs, including the flagship Nvidia H100. - **Key Achievements:** First to train an AI model in space (nanoGPT), first to run a version of Gemini in orbit, and first to perform high-powered inference on SAR data. - **Primary Technical Hurdles Overcome:** Thermal dissipation of power-dense chips and radiation tolerance against bit flips.

[01:43](https://www.youtube.com/watch?v=94b6i5jI1nE&t=103s) The Economic Case: Space vs. Earth Solar Projects

Johnston presents a compelling economic comparison between building a solar-powered data center on Earth versus building one in space. On Earth, solar is the cheapest form of energy, but developers face three massive cost drivers: the cost of permitted land (which is often the single largest expense in North America), the cost of massive battery storage and backup power to handle the roughly 20 hours of the day without peak sunlight, and the cost of the solar cells themselves.

In space, the financial equation changes dramatically. Permitted land costs drop to zero, and because a satellite can sit in 24/7 continuous sunlight, there is no need for expensive battery storage or backup power. Furthermore, because solar radiation is unfiltered by Earth's atmosphere, one square meter of solar panel in space produces eight times the energy of the same panel on Earth, meaning space data centers require eight times fewer solar cells. The only major additional cost is launch. Johnston explains that the economic break-even point occurs when launch costs fall below $500 per kilogram—a 10x reduction from current average costs. This target is well within reach of upcoming launch vehicles; for instance, SpaceX's Starship is designed to bring launch costs down to a mere $10 to $20 per kilogram.

- **Earth Solar Cost Drivers:** Permitted land (highest cost in North America), battery storage/backup (due to only ~4 hours of peak daily sun), and solar cells. - **Space Solar Advantages:** Zero land costs, zero battery storage requirements (24/7 sun), and 8x higher energy yield per square meter of solar panel. - **Economic Break-Even:** Occurs at a launch cost of ~$500/kg. - **Starship Target Cost:** Designed to achieve launch costs of $10 to $20/kg.

[03:18](https://www.youtube.com/watch?v=94b6i5jI1nE&t=198s) The 88,000-Satellite Constellation

Star Cloud is already moving past individual test satellites. Johnston shares a concept video of a massive constellation the company is actively planning, having recently filed with the SEC for a constellation of 88,000 satellites. Each satellite will support roughly 200 kilowatts of power, enabling a staggering 20 gigawatts of new compute capacity. Johnston notes that this massive network is designed primarily for inference workloads, such as generating 3D video, powering back-office business processing agents, or running code-generation agents.

The satellites will operate in a dawn-dusk sun-synchronous orbit, ensuring they are perpetually exposed to the sun for uninterrupted 24/7 power. Users on Earth will connect to the constellation via optical links, achieving sub-50 millisecond latency to anywhere on the globe. Johnston frames this as "the start of the largest infrastructure project ever," noting that the $100 billion capex spend required for this constellation is actually far lower than what a comparable 20-gigawatt infrastructure project would cost terrestrially. He boldly concludes that this project represents the dawn of a "Kardashev type two" or Dyson sphere-type civilization, and potentially even Kardashev type three.

- **Constellation Scale:** 88,000 satellites filed with the SEC, targeting 20 gigawatts of total compute capacity. - **Power per Satellite:** Approximately 200 kilowatts. - **Orbit Type:** Dawn-dusk sun-synchronous orbit for continuous 24/7 solar exposure. - **Performance:** Sub-50 millisecond latency globally via optical links. - **Projected Capex:** $100 billion, which Johnston notes is cheaper than building the equivalent capacity on Earth.

[04:43](https://www.youtube.com/watch?v=94b6i5jI1nE&t=283s) Q&A: Thermal Dissipation and the Stefan-Boltzmann Equation

During the Q&A session, an audience member asks for the "napkin math" on how Star Cloud handles heat dissipation in the vacuum of space, where there is no air to conduct heat away. Johnston explains that while space has a very low ambient temperature of 3 Kelvin, dissipating heat in a vacuum requires a large surface area to emit energy as infrared radiation.

To illustrate the math, Johnston explains that solar panels in space generate about 200 watts of power per square meter. A radiator operating at 50°C can dissipate about 800 watts per square meter. This means a satellite needs a radiator surface area equal to about one-quarter of its solar panel surface area (for example, a 400-square-meter solar array requires a 100-square-meter radiator). However, the Stefan-Boltzmann equation dictates that thermal dissipation is proportional to the fourth power of absolute temperature ($T^4$). By raising the operating temperature of the chips from 50°C to 80°C—just a 10% increase in Kelvin—they can nearly halve the required radiator surface area. To achieve this, Star Cloud is collaborating with Nvidia on a new space-grade chip called "Ruben 1," designed to run hotter without increasing failure rates, thereby significantly reducing radiator mass.

- **Solar Generation vs. Radiator Dissipation:** Solar panels generate ~200 W/m², while a 50°C radiator dissipates ~800 W/m². - **Stefan-Boltzmann Relationship:** Dissipation scales with the fourth power of temperature ($T^4$). - **Nvidia Collaboration:** Working on the "Ruben 1" space chip to operate at higher temperatures (e.g., 80°C) to reduce radiator mass. Jensen Huang highlighted this partnership during his GTC keynote.

[06:40](https://www.youtube.com/watch?v=94b6i5jI1nE&t=400s) Q&A: Kessler Syndrome and Space Congestion

Another audience member raises the inevitable concern of Kessler syndrome—the threat of cascading satellite collisions creating an impassable field of space debris. Johnston emphasizes that Star Cloud takes space safety incredibly seriously, stating that "everybody needs to be a responsible user of space." He explains that they mitigate this risk in a few ways. For their initial satellites, they fly at a relatively low altitude of around 400 kilometers. At this height, atmospheric drag causes satellites to naturally deorbit and burn up within a few months. If a collision were to occur, the debris would quickly lose altitude and decay, making a runaway Kessler event highly unlikely.

For higher orbits, Johnston points out that these regions are currently extremely unpopulated because they begin to edge into the high-radiation Van Allen belts. He points to SpaceX as a highly successful case study, noting they operate roughly 10,000 satellites in low Earth orbit without a single collision, thanks to highly sophisticated, automated collision avoidance systems. Johnston also debunks the visual perception of space congestion, explaining that on orbital tracking maps, each satellite dot is rendered "about the width of California" to be visible, which wildly distorts how crowded space actually is. In reality, space is vast enough to easily accommodate terawatts of compute in dawn-dusk orbits without major collision risks.

- **Initial Orbit Altitude:** ~400 km, ensuring natural deorbiting within a few months. - **SpaceX Case Study:** ~10,000 active satellites with zero collisions due to advanced collision avoidance. - **Orbital Capacity:** The dawn-dusk sun-synchronous orbit can easily support terawatts of compute capacity.

[08:36](https://www.youtube.com/watch?v=94b6i5jI1nE&t=516s) Q&A: Radiation Testing, Inference vs. Training, and the 5 GW Concept

When asked about how they handle radiation and bit flips, Johnston explains that they rely on extensive ground testing. They have conducted four rounds of testing using a high-velocity proton particle accelerator at the cyclotron in Knoxville, Tennessee. For heavy ion testing, they use the Brookhaven National Lab, where they can expose chips to five years' worth of space radiation dosage in a single 24-hour period. This rigorous telemetry directly informs their physical shielding designs and software-level error correction.

An audience member asks if Star Cloud does contiguous compute or focuses on something else. Johnston clarifies that for the foreseeable future, their focus is almost exclusively on inference. He notes that inference will soon make up 99% of the total AI compute market anyway, and training large models in space is incredibly difficult because it would require docking massive structures together. Prompted by the audience's curiosity, Johnston shares a concept video of what a massive 5-gigawatt, 4-kilometer-by-4-kilometer training structure would look like. It would require launching 40-megawatt modules via Starship to connect to a central spine, backed by a massive 1-kilometer-by-4-kilometer radiator. Johnston admits such a structure is at least 15 years away. He closes the session with an audience poll on when space-based compute will become cheaper than terrestrial compute, revealing a mix of optimism and skepticism, with many voting for the 5-to-10-year timeframe.

- **Radiation Testing Facilities:** Cyclotron in Knoxville (protons) and Brookhaven National Lab (heavy ions, simulating 5 years of radiation in 24 hours). - **Workload Focus:** Exclusively inference, which Johnston projects will represent 99% of the AI market. - **5 GW Training Concept:** A hypothetical 4 km x 4 km structure composed of 40 MW Starship-delivered modules, featuring a 1 km x 4 km radiator.

## Notable quotes > [00:01](https://www.youtube.com/watch?v=94b6i5jI1nE&t=1s) also been abusing GPUs in ways they were not designed for. — Philip Johnston

> [00:01](https://www.youtube.com/watch?v=94b6i5jI1nE&t=1s) into the shadow. — Philip Johnston

> [03:18](https://www.youtube.com/watch?v=94b6i5jI1nE&t=198s) the start of the largest infrastructure project ever — Philip Johnston

> [03:18](https://www.youtube.com/watch?v=94b6i5jI1nE&t=198s) Kardashev type two — Philip Johnston

> [06:40](https://www.youtube.com/watch?v=94b6i5jI1nE&t=400s) everybody needs to be a responsible user of space. — Philip Johnston

> [06:40](https://www.youtube.com/watch?v=94b6i5jI1nE&t=400s) about the width of California — Philip Johnston

## Mentioned Philip Johnston, Starcloud, Sequoia Capital, Andrej Karpathy, nanoGPT, Google, Gemini, Nvidia, Nvidia H100, SpaceX, Starship, SEC, Ruben 1, Jensen Huang, GTC, Van Allen belts, Knoxville, Tennessee, Brookhaven National Lab

[Source](https://www.youtube.com/watch?v=94b6i5jI1nE)

中文简报

Starcloud瞄准千亿美元太空AI网络以压低地面算力成本该公司计划部署由8.8万颗卫星构成的星座，以规避地球上的土地和电池限制，并指出发射成本的骤降将很快使轨道空间成为AI推理成本最低的场所。 Philip Johnston介绍了Starcloud利用Nvidia H100 GPU在太空成功运行并训练AI模型的突破，打破了此前业内对辐射和热管理的质疑。他详细阐述了天基太阳能的经济优势：其每平方米发电量是地面的八倍，且免去了土地和电池成本，一旦借助SpaceX的Starship等运载工具使发射成本降至每公斤500美元以下即可实现财务收支平衡。为把握这一机遇，Starcloud正筹划部署已向SEC申报的8.8万颗卫星星座，以提供20吉瓦全球可访问的低延迟推理算力。在问答环节，Johnston还探讨了基于斯忒藩-玻尔兹曼定律（Stefan-Boltzmann law）的散热技术方案、通过低轨部署缓解凯斯勒现象（Kessler syndrome）的措施，以及详尽的辐射测试方案。 Star Cloud 成功发射搭载 Nvidia H100 等芯片的 Star Cloud 1 卫星，并在轨运行 nanoGPT 与 Gemini，验证了数据中心级 GPU 克服真空散热与辐射位翻转在太空运行的可行性。 (0:01) 太空太阳能单平米发电量是地面的 8 倍，且无需昂贵的储能电池与土地许可成本；当火箭发射成本降至 500 美元/公斤以下时，太空算力成本将优于地面，而 SpaceX 星舰目标成本仅为 10 至 20 美元/公斤。 (1:43) Star Cloud 已向 SEC 提交申请部署由 88,000 颗卫星组成的星座，运行于晨昏太阳同步轨道，预计提供 20 吉瓦（GW）全天候推理算力，并通过激光光链路实现全球低于 50 毫秒的延迟。 (3:18) 针对太空散热挑战，Star Cloud 依托斯特藩-玻尔兹曼定律，正与 Nvidia 合作开发可承受 80°C 高温运行的「Ruben 1」太空芯片，以大幅缩减卫星辐射散热板所需的表面积。 (4:43) 针对凯斯勒现象与太空垃圾风险，公司将初始卫星轨道设在约 400 公里高度，确保报废或碰撞碎片能在数月内受大气阻力自然坠毁烧毁。 (6:40)

Reading summary

Star Cloud co-founder and CEO Philip Johnston argues that space-based data centers will become economically and operationally superior to terrestrial solar facilities as launch costs fall. By leveraging continuous solar exposure, zero land acquisition costs, and higher solar irradiance, Star Cloud plans to deploy orbital inference clusters at lower capital expenditure than comparable Earth-based infrastructure. Space Solar Economics and Launch Cost Thresholds (1:43) Johnston compares the economics of terrestrial solar data centers against orbital deployments. On Earth, solar developers face three major cost drivers: permitted land acquisition (which Johnston cites as the single largest expense in North America), substantial battery storage and backup systems to handle the roughly 20 daily hours without peak sunlight, and solar cells. In contrast, orbital facilities operating in dawn-dusk sun-synchronous orbits receive continuous 24/7 sunlight, eliminating the need for battery storage and land permits entirely. Furthermore, solar panels in space yield eight times more energy per square meter than on Earth due to unfiltered solar radiation. According to Johnston, space-based compute reaches economic parity when launch costs decline below $500 per kilogram, approximately a tenfold reduction from current averages. Star Cloud expects upcoming heavy-lift vehicles such as SpaceX's Starship—designed to target launch costs of $10 to $20 per kilogram—to comfortably beat this break-even threshold, fundamentally altering infrastructure economics. Thermal Dissipation and Radiation Management (4:43) Deploying data-center silicon in a vacuum requires overcoming heat dissipation and radiation hazards. Because vacuum lacks air for convective cooling, satellites must shed heat purely as infrared radiation. In orbit, solar panels generate approximately 200 watts per square meter, while a radiator running at 50°C dissipates roughly 800 watts per square meter. Under the Stefan-Boltzmann law, thermal dissipation scales with the fourth power of absolute temperature. Star Cloud is collaborating with Nvidia on a space-grade chip termed "Ruben 1" designed to operate at 80°C rather than 50°C, a modest temperature rise that nearly halves the necessary radiator surface area. To manage radiation-induced bit flips, Star Cloud tested hardware using proton beams at a Knoxville cyclotron and heavy ions at Brookhaven National Laboratory, where components received five years of equivalent space radiation in 24 hours. The company validated this design with its Star Cloud 1 test satellite, deploying five Nvidia GPUs (including an H100) to train Andrej Karpathy's nanoGPT, run a version of Google's Gemini, and process synthetic aperture radar inference in orbit. Constellation Architecture and Orbital Congestion Risks (3:18) Star Cloud has filed with the SEC for a constellation of 88,000 satellites targeting 20 gigawatts of aggregate compute capacity, with each satellite supporting approximately 200 kilowatts. Johnston projects the deployment will require $100 billion in capital expenditure, asserting this remains cheaper than constructing equivalent 20-gigawatt capacity terrestrially. Orbital nodes will communicate with terrestrial users via optical links designed to deliver sub-50 millisecond latency worldwide. Addressing concerns regarding Kessler syndrome and orbital debris, Johnston notes initial satellites operate at low altitudes around 400 kilometers, where atmospheric drag forces satellites and fragments to naturally deorbit and burn up within months. Johnston dismisses visual tracking maps that represent satellites as California-sized markers, asserting that dawn-dusk orbits possess ample volume to accommodate terawatts of capacity without significant collision hazards, citing SpaceX's record of operating roughly 10,000 satellites without a collision using automated avoidance systems. Workload Specialization and Long-Term Scaling Limits (8:36) Star Cloud is designing its orbital constellation almost exclusively for inference workloads, such as 3D video generation, business processing automation, and code-generation agents. Johnston forecasts that inference will eventually account for 99% of the total AI compute market. He acknowledges that large-scale model training in space presents severe structural barriers, as high-bandwidth distributed training requires docking massive physical structures together. While Star Cloud developed a concept for a hypothetical 5-gigawatt training facility spanning 4 kilometers by 4 kilometers—assembled from 40-megawatt Starship-delivered modules connected to a 1-by-4-kilometer radiator—Johnston concedes that such contiguous training platforms remain at least 15 years away from technical and operational realization. While Johnston envisions orbital data centers scaling to multi-gigawatt constellations, the model depends heavily on launch costs dropping below $500 per kilogram, regulatory clearance for tens of thousands of satellites, and the thermal longevity of specialized hardware in orbit. Johnston's informal audience poll revealed persistent skepticism, with many attendees estimating that cost competitiveness remains 5 to 10 years away.

精炼阅读摘要

Star Cloud联合创始人兼首席执行官Philip Johnston提出，太空数据中心能够利用无遮挡的持续太阳能并规避高昂的地面土地与储能成本，一旦火箭发射成本降至每千克500美元以下，轨道算力的经济效益将显著超越地面。该公司已通过搭载英伟达H100等芯片的Star Cloud 1卫星初步验证了商用GPU在轨运行的可行性，并规划推进由88,000颗卫星组成的20吉瓦太空计算星座。 Star Cloud 1与在轨计算验证 (0:01) 此前工业界普遍认为，由于缺乏空气对流导致的极端散热困难以及宇宙辐射引发的位翻转问题，地面数据中心级GPU无法在太空环境中正常运转。Star Cloud发射了搭载五颗英伟达GPU（包括旗舰芯片Nvidia H100）的Star Cloud 1测试卫星，并在轨成功训练了Andrej Karpathy的nanoGPT模型、运行了谷歌Gemini的一个版本，同时完成了合成孔径雷达（SAR）卫星数据的高算力推理任务。 Johnston指出，该卫星的成功打破了常规认知，证明未经极端航天加固的先进地面芯片在经过特定的热管理和软硬件抗辐射设计后，能够在太空低轨环境中稳定执行高密度AI计算任务，为后续大规模部署奠定了工程实验基础。太空光伏经济性与发射成本阈值 (1:43) 地面建设太阳能数据中心面临三项主要资本开支：合规土地购置许可（北美地区通常最高的单项成本）、应对每日长达约20小时无峰值光照所需的巨额储能电池与备用电力系统，以及光伏组件本身的成本。相比之下，太空环境无需任何土地许可支出，且卫星部署在恒定日照轨道中可实现24小时不间断发电，直接省去了昂贵的电池储能环节。此外，由于太空中太阳辐射不受大气层衰减与散射，单位面积太阳能电池板的发电量达到地面的八倍，意味着所需光伏板面积可缩减至八分之一。Johnston指出，这一模式的经济平衡点在于运载火箭发射成本降至每千克500美元（约为当前平均水平的十分之一）；而SpaceX星舰的设计目标成本为每千克10至20美元，一旦达成，太空数据中心将在综合拥有成本上对地面形成明显优势。 8.8万颗卫星星座与AI推理网络 (3:18) Star Cloud已向美国证券交易委员会（SEC）提交了包含88,000颗卫星的星座建设计划。按照方案设计，单颗卫星将支持约200千瓦的电力负荷，全网预计释放20吉瓦的新增算力。该星座将部署于晨昏太阳同步轨道（dawn-dusk sun-synchronous orbit），确保卫星持续暴露在阳光下维持全天候供电，并通过激光光链路与地面通信，提供全球范围内低于50毫秒的低延迟连接。 Johnston透露，该星座的资本支出预计为1,000亿美元，但他强调这一金额仍低于在地面建设同等20吉瓦算力设施的总成本。该网络当前几乎完全针对AI推理任务设计（包括3D视频生成、企业后台自动化Agent以及代码生成等），因为推理将占据AI计算需求的99%。对于集中式大模型训练，Johnston虽展示了远期由40兆瓦模块拼装而成的5吉瓦平台构想，但坦言此类需要大规模轨道对接的结构至少需要15年才可能实现。斯忒藩-玻尔兹曼散热与抗辐射工程 (4:43) 在太空真空环境中，散热完全依赖红外热辐射，由斯忒藩-玻尔兹曼定律决定，散热能力与绝对温度的四次方（T⁴）成正比。太空光伏板每平方米约产生200瓦电力，而50°C工况下的辐射散热板每平方米可散逸约800瓦，这意味着卫星需要相当于太阳能板四分之一面积的散热器。为减轻散热器质量，Star Cloud与英伟达联合开发名为“Ruben 1”的太空专用芯片，旨在允许芯片在80°C的高温下持续运行，通过约10%的开尔文绝对温度提升将所需散热面积缩减近半。在辐射防护与轨道安全方面，公司在田纳西大学回旋加速器及布鲁克海文国家实验室对芯片进行了四轮质子与重离子束流测试，单日测试可模拟轨道五年的重离子辐射通量，以此指导软硬件容错设计。针对凯斯勒综合征（空间碎片雪崩）的担忧，Johnston说明早期卫星部署在约400公里的低空，残骸受大气阻力会在数月内自然再入大气层烧毁；结合全自动轨道防撞机制，他认为晨昏轨道有充足空间容纳太瓦级算力部署。 Star Cloud勾勒出一条将AI高耗能算力迁移至外太空的激进路线，试图依托无限太阳能与新型热辐射架构替代昂贵的地面电网及电池配套。然而，这一愿景在商业上成立的前提，高度依赖于SpaceX星舰等下一代运载工具能否兑现亚500美元甚至数十美元每千克的极低发射成本，同时该系统还需经受超大规模卫星群在轨运维、辐射抗扰度衰减与极端热循环的长期严苛检验。

Source quotes

[0:01] Philip Johnston: also been abusing GPUs in ways they were not designed for. [0:01] Philip Johnston: into the shadow. [3:18] Philip Johnston: the start of the largest infrastructure project ever [3:18] Philip Johnston: Kardashev type two [6:40] Philip Johnston: everybody needs to be a responsible user of space. [6:40] Philip Johnston: about the width of California

Original transcript

[0:01] Thanks so much for having me. [0:03] Um my name is Philip Johnston and I'm [0:05] the co-founder and CEO of Star Cloud and [0:08] just like the previous company we have [0:09] also been abusing GPUs in ways they were [0:11] not designed for. [0:13] >> [laughter] [0:14] >> Um so yeah, we're building data centers [0:16] in space um mainly for the energy that [0:18] we can draw and I will spend the next 5 [0:21] minutes um explaining why it will soon [0:23] make much more sense to build data [0:25] centers in space than it does to build [0:26] them on Earth and then I'll take 5 [0:28] minutes for questions. So please um [0:30] start thinking of some questions. Before [0:32] I do that though, I want to show a quick [0:34] video which is actually the deployment [0:36] of Star Cloud 1 um [0:38] and this was it had five in video GPUs [0:40] on it but the most significant was the [0:42] in video H100 chip and uh I'll just [0:44] quickly play the video first. [0:50] Star Cloud 1 separation confirmed. [0:54] So [0:56] we don't you don't normally get as great [0:57] a deployment video by the way. Half the [0:59] time they'd like deploy it into the [1:00] shadow. [laughter] [1:01] Um so the reason this was so significant [1:04] is until this point many people thought [1:05] you actually couldn't run state of the [1:07] art terrestrial data center grade GPUs [1:09] in space for two main reasons. One is [1:11] the thermal dissipation. So they they're [1:13] very power dense. They produce a lot of [1:15] heat and the second is the radiation [1:17] tolerance. So people thought that um you [1:19] would have bit flips at too high rate [1:21] and so by um with this chip we were the [1:24] first to train a model in space. We [1:25] actually trained nano GPT from Andrej [1:27] Karpathy um and then we also uh [1:31] were the first to run a version of [1:32] Gemini, the first to do high powered [1:33] inference on uh SAR data. So other [1:36] satellite data. [1:37] Um and so it was a very significant step [1:39] in proving that we can actually run the [1:41] state of the art terrestrially. [1:43] Um but the um [1:45] yeah, I think maybe to make the case for [1:47] why it will soon make more sense in [1:49] terms of energy cost, I'd like to [1:51] quickly draw a comparison with uh [1:55] with a solar project on Earth since [1:56] solar is the cheapest form of energy [1:58] that we have on Earth. So if you want to [1:59] build a solar project to power a new [2:01] data center, [2:02] you have three main costs. So the first [2:04] is the cost of permitted land and in [2:06] fact in North America that's actually [2:08] the largest cost or can be for most new [2:10] solar projects. [2:12] The second is the cost of battery [2:13] storage and backup power because we're [2:15] only you know, we only have peak power [2:16] for about 4 hours of the day. So we need [2:18] to charge those batteries to use at [2:19] night. And then the last is the cost of [2:23] the solar cells themselves. [2:25] So how does that compare to building a [2:27] similarly sized solar project in space? [2:30] Well, in space number one we don't need [2:31] to pay for permitted land. So your [2:32] biggest cost is gone. You don't need to [2:35] pay for battery storage and backup power [2:36] because we're 24/7 in the sun. So your [2:39] second biggest cost is gone. And then [2:40] you need eight times less solar cells [2:42] cuz 1 square meter of solar panel in [2:44] space produces eight times the energy of [2:47] 1 square meter of solar panel on Earth. [2:49] So the only additional cost or the main [2:50] additional cost we have in space is the [2:52] launch cost. [2:54] >> [snorts] [2:54] >> And so you can clearly see there's a [2:55] break even point where the launch cost [2:57] comes below the cost of permitted land, [3:00] batteries and solar and we see that [3:02] break even cost to be around $500 a [3:04] kilo. So about a 10x reduction from [3:06] where we are today. But that's well [3:08] within range of the launch vehicles that [3:10] are coming online. So for comparison, [3:12] Starship is designed to produce launch [3:15] cost of around 10 to $20 a kilo. [3:17] Um [3:18] And so I think I'll just finish by [3:20] playing you a uh one final concept video [3:23] and this shows uh a constellation that [3:26] we're building now. So we've just filed [3:27] with the SEC for a constellation of [3:29] 88,000 satellites. Each one's about 200 [3:32] kilowatts. It will enable us to deploy [3:34] on the order of 20 gigawatts of new [3:36] compute capacity. Really just scratching [3:38] the surface um with this new [3:40] constellation. [3:41] Um and it will enable it's basically for [3:43] all inference workloads. And so this [3:46] could be and yeah, maybe I'll I'll start [3:48] the video and you can [3:49] get a sense of it. [3:51] So in this case it's to generate a 3D [3:53] video but it could also be for back [3:56] office business processing agents, code [3:57] generation agents. [3:59] Um they will come up via optical link to [4:01] this constellation in this dawn dusk sun [4:03] synchronous orbit. Means it's always in [4:05] the sun. 24/7 power. Sub 50 millisecond [4:08] latency to anywhere on Earth. [4:10] Um all optically linked. [4:12] >> [snorts] [4:13] >> And this really is the start of the [4:14] largest infrastructure project ever. I [4:17] mean we're talking about just for this [4:18] constellation of 88,000 we're talking [4:20] about a hundred billion dollars of capex [4:21] spend which is actually much lower than [4:23] it would cost to do to do terrestrially. [4:26] Um and uh not only is it the start of [4:28] the largest infrastructure project, it's [4:29] also in my opinion the start of a [4:31] Kardashev type two uh Dyson sphere type [4:34] civilization and potentially Kardashev [4:36] type three. [4:37] Um I will finish there and we have about [4:39] 4 minutes for questions. So any [4:41] questions? Yeah, we'll start at the [4:42] front. [4:43] Yeah. [4:45] Uh the intuition on the availability [4:47] solar is obvious. Um can you just give [4:49] us the napkin math on the the radiator [4:51] equation again? So like dissipating heat [4:53] for anyone who's thought about it, it [4:54] always feels hard. And then and also [4:56] please say something about the [4:57] availability of dawn dusk. That orbit is [4:59] finite, right? Yes. Yeah. Yeah. [5:01] Um it's a great question. So because [5:03] space is a vacuum, it's actually much [5:05] harder you know, space is only 3° [5:07] Kelvin. So very low ambient temperature. [5:10] But because it's a vacuum um as you [5:12] rightly point out, it's actually quite [5:14] difficult to dissipate that heat and [5:15] what it requires is a large surface area [5:18] so that you can emit that in infrared. [5:19] So everything that's warm is glowing in [5:21] in infrared all the time. If you had an [5:23] infrared camera on my face, you'd see [5:24] that I'm glowing. Um and so the rough [5:27] math on these surface area is you would [5:30] your solar panels generate around 200 [5:32] watts per square meter and the radiator [5:34] if you keep it around 50° C [5:37] uh will dissipate around 800 watts per [5:39] square meter. So what that means is if [5:41] you've got a um [5:43] you need about a quarter again the [5:44] surface area in radiator than you have [5:47] on um on solar panel. So if you had a [5:49] 400 square meter solar panel, you'd need [5:51] an additional 100 square meters of [5:53] radiator to dissipate that heat. Um [5:56] there's a very nice equation called the [5:58] Stefan-Boltzmann equation which [5:59] basically says that the rate that the [6:01] the thermal dissipation is proportional [6:04] to the to the fourth power of the [6:06] temperature. So if you can jack up that [6:08] temperature instead of being 50° to 80°, [6:10] which is like a 10% increase in Kelvin, [6:13] then you you can actually half the [6:14] surface area of your radiator or close [6:16] to half the surface area of radiator. [6:18] And so that's what we're working on with [6:19] Nvidia now. If anybody was at GTC, [6:21] you'll have seen Jensen walk out to this [6:23] um to the deployment video of Star Cloud [6:25] 1 and then he spent uh 5 minutes talking [6:27] about the um the new space Ruben 1 chip [6:30] that we're working on and it's designed [6:32] to run at a hotter temperature without [6:34] having a a higher failure rate. And the [6:36] reason you want it to run at a hotter [6:37] temperature is so that you can lower the [6:39] mass on the radiator. [6:40] Great question then. Yeah. Yeah. All [6:42] right. Kessler syndrome, the favorite [6:44] thing, right? Now we're going to get [6:45] more satellites there yet more in uh [6:48] Kessler orbits cuz everyone will want [6:50] What happens then? Yeah, it's a great [6:52] question and it's also related to the [6:53] question um asked about the space in [6:56] space. [6:57] Um [6:58] I think so it's something we take [7:00] incredibly seriously. You know, [7:01] everybody needs to be a responsible user [7:03] of space. We do and and everybody else. [7:05] We're you know, keen keen to make sure [7:07] that space is usable forever. Um [7:10] For the first few satellites, so you can [7:12] solve it in a few ways. If you fly at a [7:14] relatively low altitude, the chance of a [7:16] Kessler type effect is is extremely low. [7:18] So um our first satellite we're flying [7:20] around 400 km altitude. That means that [7:23] it will naturally deorbit within a few [7:24] months. And so if you were to have a [7:26] collision at that altitude, by the time [7:28] it gets around to the next orbit, you're [7:30] already a a few hundred meters below [7:31] where you had the collision and the [7:33] chance yeah, chance of um Kessler is [7:35] very very low. Um [7:37] Yeah, as you fly higher, it's actually [7:39] extremely unpopulated those high orbits [7:41] because then you start to edge into the [7:43] the Van Allen radiation belt. Um but I [7:46] mean we actually have a pretty [7:48] good case study for this and that is um [7:50] SpaceX is now operating around 10,000 [7:52] satellites without ever having a single [7:54] collision um in in low Earth orbit. And [7:57] the the way that you do that is by [7:59] having ex-pretty sophisticated collision [8:01] avoidance. Um [8:04] Uh the the the the other reason I think [8:05] people think this is more of an issue [8:06] than it [8:07] than it actually is and and the reason [8:09] that space is so much larger than it [8:11] looks is when you see a map of all of [8:13] those satellites, each dot on those maps [8:16] is about the width of California and [8:18] you're representing something that might [8:19] be this wide by something the width of [8:20] California. And so people can often [8:22] think that space is very congested. It's [8:24] actually [8:25] it you you we can easily fit on the [8:27] order of uh terawatts of compute in just [8:29] this dawn dusk sun synchronous orbit [8:30] without having [8:32] um [8:33] you know, huge problems with collision [8:34] avoidance. [8:36] Any other questions? Yeah. [8:39] Uh [8:39] is radiation like bit flipping, is that [8:43] something you actually have to think [8:43] about or consider? How does that impact [8:45] stuff? [8:46] Yes, it is something I have to think [8:47] about. Um [8:48] so the way that we're solving it is just [8:50] an enormous amount of ground testing. So [8:52] we've done uh four rounds of testing at [8:54] the cyclotron down in Knoxville. Um it's [8:56] a high velocity proton particle [8:58] accelerator. [8:59] Um and we take all of that telemetry and [9:01] then that informs our choice on [9:02] shielding. And then for heavy ions we [9:04] have to go to the Brookhaven National [9:06] Lab and we basically run all the chips [9:08] through um [9:09] through the the space environment. So if [9:11] over 24 hour period, you can put it [9:13] through 5 years worth of radiation dose [9:15] um and then we take all that data and we [9:17] then use that to inform shielding but [9:19] also software development choices for [9:20] it. [9:21] Yeah. [9:22] Yeah. Do you do contiguous compute or [9:24] are you doing something else? [9:26] Uh this is almost exclusively I mean [9:29] actually for the foreseeable future will [9:32] just be for inference. Um and the reason [9:34] the reason it's for inference is number [9:36] one, inference is going to be like 99% [9:38] of the compute market very soon anyway. [9:39] So even if we you know, we wouldn't want [9:41] to write be running a large training set [9:43] um [9:44] well, running large training sets will [9:45] be a very small percentage of the total [9:47] in 5 to 10 year time of AI workloads. [9:50] But secondly, it's very hard. We would [9:52] need a [9:54] yeah, we We need to dock together a [9:56] large 5 gigawatt um kind of structure. [9:59] The I actually have a video of that [10:00] here, but I won't waste everyone's time [10:01] with it. [10:02] >> [laughter] [10:03] >> Unless [snorts] somebody wants to see a [10:04] video of a 5 gigawatt data center in [10:06] space. Do you guys want to see that? All [10:08] right. [10:08] >> [laughter] [10:10] >> We we made this video cuz we didn't want [10:11] people to be like, "Oh, you could never [10:13] train a model in space." So, this is [10:14] what a 5 gigawatt 4 km by 4 km structure [10:18] in space would look like. [10:20] So, this would be a Starship launch [10:22] vehicle with a 40 megawatt is what you [10:25] can fit per Starship launch vehicle. [10:27] Which will connect to a central spine. [10:30] Which is connected to this enormous [10:33] solar panel. On the back there, we have [10:34] a 1 km by 4 km radiator. [10:39] Yeah, that that that's how you would [10:40] train a large model, but as I say, it'll [10:42] probably be at least 15 years before we [10:44] get to anything [10:45] like that. [10:48] One more I think we're Okay, 40 seconds. [10:50] Or do we [10:51] Okay. Oh, sorry. [10:53] Um by when do you think the majority of [10:55] the data center will be in the space? [10:57] Oh, that's a great question. And I [10:58] actually wanted to ask all of you. Let's [10:59] run a poll. So, the [11:01] the question I want to ask is when do [11:03] you think it will be cheaper to run [11:05] compute in space for anybody? It could [11:06] be for Space X or for us. And the four [11:08] answers will be [11:09] the next 5 years it will be cheaper, in [11:12] 5 to 10 years it will be cheaper, [11:14] sometime after 10 years, or never. Okay, [11:17] so who thinks within 5 years it will be [11:20] cheaper to run compute in space than [11:22] terrestrially? [11:24] Interesting. Who thinks 5 to 10 years? [11:29] Who thinks beyond 10 years? [11:32] And who thinks never? [11:34] Brave. [11:35] Okay, that's that's an interesting for [11:37] me to see. [11:38] Um I don't Yeah, I think that's we are [11:41] out of time, so I will leave it there. [11:42] Thank you very much for your time. [11:46] >> [applause]