Space data center cooling: why Google’s TPUs in orbit run 15 minutes

太空数据中心散热：为何谷歌的 TPU 在轨道上只能运行 15 分钟

A space data center sounds like it solves cooling for free: deep space sits at about −270 °C. Google’s own engineers say the opposite. In the post announcing Project Suncatcher’s first launch, Google calls cooling orbital data centers “a crucial research challenge”, and according to Ars Technica its first test satellite can run its TPUs for only about 15 minutes before it has to stop and let the radiators catch up. 太空数据中心听起来似乎能免费解决散热问题：深空的温度约为 -270 °C。但谷歌的工程师们却持相反观点。在宣布“追日者计划”（Project Suncatcher）首次发射的博文中，谷歌将轨道数据中心的散热称为“一项关键的研究挑战”。据 Ars Technica 报道，其首颗测试卫星在运行 TPU 时，仅能持续约 15 分钟，之后就必须停下来等待散热器降温。

If you care where AI compute gets built, and who pays for the power, this is the physics behind the headline. TL;DR On October 1 Google launches “MVP”, a fridge-sized satellite built with Planet that carries four TPUs, on SpaceX’s Transporter-18 rideshare. It is a test, meant to run for a few months. 如果你关心 AI 计算设施建在哪里，以及谁在为电力买单，那么这就是标题背后的物理学逻辑。简而言之：10 月 1 日，谷歌通过 SpaceX 的 Transporter-18 拼车任务发射了“MVP”卫星。这是一颗与 Planet 公司合作制造、冰箱大小的卫星，搭载了四颗 TPU。这只是一次计划运行几个月的测试。

Space is cold but it is a vacuum, so heat cannot be carried away by air or water. It can only leave as infrared light from a radiator, and a perfect radiator at room temperature sheds about 460 watts per square metre. The satellite gets about one kilowatt of solar power, roughly a hair dryer. Per Ars Technica, citing The New York Times, the cooling system runs in “brief spurts of about 15 minutes”. Then the TPUs shut down. 太空虽然寒冷，但它是真空环境，因此热量无法通过空气或水带走。热量只能以红外光的形式从散热器散发出去，而一个室温下的完美散热器每平方米大约只能散发 460 瓦的热量。这颗卫星获得的太阳能功率约为 1 千瓦，大约相当于一个吹风机的功率。据 Ars Technica 援引《纽约时报》的报道，其冷却系统以“约 15 分钟的短暂爆发”方式运行，随后 TPU 必须关机。

Why bother: in the right orbit a panel sees near-constant sunlight and, per Google, makes up to eight times more power than on Earth, with no permitting board in the way. Also in this episode: Oracle owes investors for a New Mexico data center even without electricity, UK iPhones with two levels of encryption, and F-Droid 2.0. 为什么要费这个劲？在合适的轨道上，太阳能板几乎能持续接收阳光。据谷歌称，其产生的电力比在地球上多出八倍，且没有审批委员会的阻碍。本期内容还包括：甲骨文公司即便在没有电力的情况下仍需为新墨西哥州的数据中心向投资者负责、英国 iPhone 的双重加密，以及 F-Droid 2.0。

Why is cooling a space data center so hard? On Earth, cooling is contact: air or water touches a hot chip and carries the heat away. In orbit there is nothing to touch. Google’s post says it in two sentences: Radiation is slow. The Stefan–Boltzmann law gives the power a surface can glow away: P = εσT⁴. For a perfect radiator (ε = 1) at about room temperature, 300 K, that is roughly 460 watts per square metre per side. 为什么太空数据中心散热如此困难？在地球上，散热依靠接触：空气或水接触发热的芯片并将热量带走。但在轨道上，没有任何介质可以接触。谷歌的博文用两句话概括了这一点：辐射散热很慢。斯特藩-玻尔兹曼定律（Stefan–Boltzmann law）给出了表面辐射散热的功率：P = εσT⁴。对于一个室温（300 K）下的完美散热器（ε = 1），每平方米每侧大约只能散发 460 瓦的热量。

The −270 °C of the cosmic background barely helps: the limit is what the radiator can emit, which depends on its own temperature. The International Space Station is the working example. Its external active thermal control system is “designed to provide 35 kW of heat rejection per loop for a total capability of 70 kW”, per NASA. It uses two rotating radiator wings with three panels each, and each panel is 23.3 by 3.4 metres, about the length of a tennis court. Six tennis-court-long panels, 70 kilowatts. -270 °C 的宇宙背景温度几乎没有帮助：散热极限取决于散热器自身能发射多少热量，而这又取决于它自身的温度。国际空间站就是一个现成的例子。据 NASA 称，其外部主动热控系统“设计为每个回路提供 35 千瓦的散热能力，总能力为 70 千瓦”。它使用了两个旋转散热翼，每个翼有三块面板，每块面板长 23.3 米、宽 3.4 米，大约相当于一个网球场的长度。六块网球场大小的面板，仅能处理 70 千瓦的热量。

Now scale. Google’s Ironwood TPU “scales up to 9,216 liquid cooled chips … spanning nearly 10 MW”. Ten megawatts divided by 70 kilowatts is about 140 space stations’ worth of radiators for one pod. 现在来算算规模。谷歌的 Ironwood TPU “可扩展至 9,216 颗液冷芯片……总功率接近 10 兆瓦”。10 兆瓦除以 70 千瓦，意味着一个计算舱需要相当于 140 个国际空间站规模的散热器。

Hotter radiators do much better, because the output grows with the fourth power of temperature. So treat “140 space stations” as a rough sense of scale, nothing you could build to. The direction is clear either way: in space the chips are easy to power and hard to cool. Which explains one word in Google’s post. Future satellites “will each carry dozens of TPU chips”. Only dozens. 温度更高的散热器效果会好得多，因为散热功率随温度的四次方增长。所以，“140 个空间站”只是一个粗略的规模概念，现实中根本无法建造。无论如何，方向很明确：在太空中，芯片很容易供电，但很难散热。这解释了谷歌博文中的一个词：未来的卫星“每颗将携带数十颗 TPU 芯片”。仅仅是数十颗。

Project Suncatcher: what Google is launching on October 1. The announcement came from Travis Beals, Senior Director for Paradigms of Intelligence. Google is “launching a prototype satellite to evaluate how Google Tensor Processing Units (TPUs) perform in space”, developed with Planet and flying on SpaceX’s Transporter-18 rideshare. The New York Times got the first inside look: a Falcon 9 from Vandenberg Space Force Base, a satellite named MVP. For once, an honest product name. “追日者计划”：谷歌 10 月 1 日发射了什么？该公告由智能范式高级总监 Travis Beals 发布。谷歌正在“发射一颗原型卫星，以评估谷歌张量处理单元（TPU）在太空中的表现”。该卫星与 Planet 公司合作开发，搭乘 SpaceX 的 Transporter-18 拼车任务发射。《纽约时报》获得了首次内部观察机会：一枚从范登堡太空军基地发射的猎鹰 9 号火箭，搭载了一颗名为“MVP”的卫星。难得有一个诚实的产品命名（MVP 在科技界常指“最小可行性产品”）。

Per Ars Technica, MVP is about the size of a refrigerator, carries four TPUs, and is built on a spacecraft Planet had already made for an early test. The original plan was two custom satellites in 2027; Google wanted to go sooner. The heat path runs from the chips through a thermal interface material into aluminium and copper heat pipes and out to a radiator, tested in a thermal vacuum chamber. 据 Ars Technica 报道，MVP 大小如冰箱，携带四颗 TPU，基于 Planet 公司此前为早期测试制造的航天器平台构建。最初的计划是在 2027 年发射两颗定制卫星，但谷歌希望提前行动。热量路径是从芯片通过导热界面材料进入铝铜热管，最后到达散热器，并在热真空室中进行了测试。

The radiation result goes back to Google’s November 2025 research paper: the first irregularities showed up in high-bandwidth memory at 2 krad(Si), about three times the shielded five-year dose, with no hard failures up to 15 krad. That paper also listed “thermal management” among the “significant engineering challenges” remaining. A year later it is still the headline problem. 辐射测试结果可追溯到谷歌 2025 年 11 月的研究论文：高带宽内存在 2 krad(Si) 剂量下首次出现异常，这大约是屏蔽后五年总剂量的三倍，但在 15 krad 之前没有出现硬故障。该论文还将“热管理”列为剩余的“重大工程挑战”之一。一年过去了，这仍然是首要难题。

The 15-minute limit: how Google’s first orbital data center actually runs. Here is the number Google’s own post leaves out. MVP’s solar panels produce “only about one kilowatt of power, The New York Times reports”, Ars writes. “That’s enough to power a microwave or a hair dryer.” And the cooling can’t keep up even with that: So they survived the radiation in a proton beam and the rocket on a shake table. The limit is the heat. Right now physics is winning, and Google’s first space data center works in shifts, with naps. That is fine for a test: a duty cycle is a measurement too. It is a problem only for the phrase “data center”. 15 分钟限制：谷歌首个轨道数据中心是如何运行的。这是谷歌博文中遗漏的数据。Ars 写道，据《纽约时报》报道，MVP 的太阳能电池板“仅能产生约 1 千瓦的电力”。“这只够驱动一台微波炉或吹风机。”即便如此，散热系统也跟不上：它们在质子束中经受住了辐射测试，在振动台上经受住了火箭发射测试。真正的限制是热量。目前物理定律占据了上风，谷歌的第一个太空数据中心只能轮班工作，并需要“打盹”休息。作为测试，这没问题：占空比本身也是一种测量数据。这只是对于“数据中心”这个词来说是个问题。