太空数据中心（第二部分）

Data Centers in Space Part II Edison Yu We hosted a very well attended expert call with Planet Lab’s Chief Space OfficerJames Mason before the holidays to discuss space data centers (SDCs) givenPlanet’s role in Google Project Suncatcher. Our discussion validated many pointsraised in our original report and more importantly highlighted some interesting andimportant nuances. Moreover, we attempt to quantify the potential cost of SDC Research Analyst Engineering challenges appear solvable Benjamin Black, CFAResearch Analyst We examined 4 main technical areas with the expert. The general view is that theseissues appear solvable over time. For Google's Project Suncatcher specifically(prototypes in 2027), Planet will use the new Owl satellite bus, increase the powerwith more solar panels (and radiators), and replace the camera payload with a Roshan Ranjit, CFAResearch Analyst nOrbit- The ideal position for SDCs to maximize sunlight exposure is aduskdawn sun-synchronous orbit(SSO) at roughly 600-800 kilometers aboveEarth. Naturally, operators will gravitate to put satellites in this area,potentially leading to overcrowding. The expert does not see this as aconstraint considering space debris and traffic management are alreadyinnately factored into existing constellation plans. In fact, certain layers areactually less congested than feared such as 600-700 km (as opposed to Research Associate nRadiation- There is ongoing development to make the electronics moreresilient to space radiation. Planet Labs has essentially been working on thisproblem in some shape or form since 2010 along with componentminiaturization and believes it is manageable. At the system level, there willinitially be redundancy built in by overprovisioning for a certain failure rate nLatency- Forspace-to-groundconnectivity, the signal travels through thevacuum of space which can be faster than terrestrial/fiber if the satellite isoverhead. As such, running inference workloads make sense especially forspace assets capturing a lot of orbital data whereas training may not besuitable (e.g., requires massive upload of datasets and frequent parametersynchronization). Forinter-satelliteconnections, this will involve usingmultiple optical laser terminals per satellite to create a mesh network androute data as quickly as possible. The expert indicated that optical laserterminals are a supply chain bottleneck due to the lack of scale/cost. Inother words, the core technology exists but there is yet to be a supplier thatcan achieve the volume/price point threshold desired. Therefore, Planet Strategic rationale Aside from pure cost, there are strategic reasons to deploy SDCs and the expertthinks governments and hyperscalers may “pay up” to at least have some capacity nSpeed- Constructing a new data center on the ground is lengthy requiringlandacquisition,permitting,and increasingly more scrutiny fromregulators. In relation, the data center may require construction of a parallelpower plant which has its own bottlenecks. Building a satellite and nRisk mitigation- While operating in space clearly presents risks, thesedangers are inherently different than those on the ground such as naturaldisasters, physical sabotage, and regulatory change. There may also besome “high ground” advantage where only certain entities have capabilityin orbit and those who don’t are at a disadvantage for geopolitical reasons.For example, there are Chinese companies working to send up SDCs Looking farther beyond, we think deploying large amounts of compute into orbit isimportant to cultivating efforts on the Moon and eventually Mars. SDCs orbiting theMoon can enable edge processing for AI autonomy, enabling real-time decision-making in habitats during communication delays. For instance, they might handle Preliminary economic analysis Assuming technical challenges can be addressed by engineering prowess andGPU/TPU costs are roughly equal between space and terrestrial, rocket launchappears to be the largest economic hurdle. For our preliminary analysis, we use anopen-source model created by Andrew McCalip, the head of R&D at Varda Space(private).Our initial finding is that the near-term cost of deploying a 1 GW space Launch: the current price of a reuseable Falcon 9 launch to low-Earth orbit (LEO) is~$70m or $4,000/kg. We estimate the gross margin to be as high as 60%, implying With Starship, we assume the initial cost will be $105m or $700/kg, then declinesignificantly with reusability to $30m or $200/kg and then ultimately reach $10m orbelow $70/kg with full-reuse and operational scale. Satellite: we estimate the cost of a Starlink modified SDC V3 satellite to be $2.1mor >$40k per kW based on the following parameters: nPower system generating 50 kW including solar arrays and thermalmanagement = $500knGPU payload of 50 Nvidia H100s @ $20k each = $1mnConnectivity using optical laser terminals = $200knPropulsion using argon hall thrusters = $150knLabor & other = $150k