“在奔向太空之前，我们应当先把地面所有可行方案都探索清楚。”阿曼达・彼得森・科里奥在WTF峰会上说道。该峰会汇聚了科技与金融行业的资深女性从业者。就在她发表上述观点之际，SpaceX的轨道数据中心规划，是支撑其约1.8万亿美元市值的重要逻辑之一。
彼得森・科里奥还提到，当下美国社会对地面数据中心的反对声渐起，很大一部分责任在于一批资历尚浅的新企业，这类企业理应接受更高监管标准。她表示，在这一轮AI算力热潮之前，深耕数据中心建设多年的企业“做法都比较规范”，相关项目很少引发争议。
把计算硬件送入轨道的发射成本“极其高昂，最高可达地面建设成本的100倍”，彼得森・科里奥称。除成本之外，运营商还需要解决在轨设备的维护与升级难题。在地面，数据中心工作人员可以随时更换服务器、重新布线调试系统。
尽管对太空算力持谨慎态度，谷歌研究团队本月官宣，将和卫星影像企业PlanetLabs合作，借助卫星探索大规模太空数据中心集群的可行性。在名为“Suncatcher（捕光计划）”的项目框架下，谷歌已经向太空发射了一台小型太空服务器原型机，搭载了和博通联合定制的AI张量处理单元（TPU）。
去年秋天，参与捕光计划的谷歌研究员连同公司首席研究员詹姆斯・曼伊卡共同发表论文：如果发射成本降至每公斤200美元及以下，太空单位功耗对应的算力成本，就有望和地面水平基本相当。谷歌研究员认为，这一门槛有望在2035年前后达成，不过时间点取决于SpaceX星舰的研发进度以及发射频次提升速度。
但其他企业态度更为乐观。SpaceX首席执行官埃隆・马斯克称，公司最快明年就能建成太空数据中心；亚马逊创始人、SpaceX竞争对手蓝色起源的掌舵人杰夫・贝佐斯则预判，十年内就会出现功耗达到1吉瓦的太空数据中心。另有报道称，谷歌与SpaceX已经探讨过合作建设在轨数据中心的方案。
彼得森・科里奥并未全盘否定太空算力构想，她称这个方向“非常有意思”；谷歌之所以持续布局这类高风险前沿项目，是因为企业习惯“提前思考10年、20年、30年之后的事”。但这项技术落地“绝非近在眼前”。OpenAI首席执行官萨姆・奥特曼的表态更为直接，他认为在当前条件下，把数据中心搬去太空的想法“很不切实际”，并质疑马斯克给出的激进时间表。
真正瓶颈：审批许可，而非电力
针对马斯克等人提出的观点——美国电力供给不足以支撑AI算力扩张，彼得森・科里奥提出了不同看法。
“电力够不够？当然足够。”她表示。真正的瓶颈，是项目并网所需的审批流程、漫长的评估研究。字母表旗下X实验室推出了Tapestry项目，正和美国东海岸最大电网运营商PJM合作，利用AI加速相关评估流程。
在得克萨斯州，州电力监管机构已经放缓部分数据中心项目审批，加强监管，避免新建算力设施加重电网负担、推高电价。
行业其他高管也指出，电网本身的利用率远未达到理想水平。数据中心土地与电力服务商Cloverleaf Infrastructure联合创始人布莱恩・贾诺斯近期接受采访时提到，电网平均约45%的容量长期闲置，这是为了应对用电峰值。一天当中绝大多数时段，电网都有富余容量可以承接数据中心用电需求。
值得一提的是，在电能利用效率（PUE）指标上，谷歌相比云业务同行具备优势，这项指标衡量数据中心算力承载效率，代表从同等电力里压榨出更多算力的能力。
对于美国国内兴起的反数据中心浪潮，彼得森・科里奥将原因归于“大量新入局玩家，它们在拿地、建设算力项目时，缺少行业积淀与成熟经验”，但她没有点名具体企业。
近几年大批新手开发商涌入AI数据中心赛道，但不止新玩家，甲骨文这类知名企业的项目同样遭遇过当地居民抵制。
更高行业标准
无论如何，她支持行业出台监管规则、设立更高标准。她没有提出具体条款，但举例说明谷歌正在提高项目属地社区投入，同时设置预案，规避项目烂尾带来的风险。
其中包括承诺在15至20年内持续承担电网升级费用，避免项目停摆后，当地社区“被迫买单”。这类前置承诺承担电网基建成本的合作模式，目前已经在美国九个以上州落地，并且“扩张速度极快”。
配套标准还包括：谷歌足额缴纳电费与财产税；缴纳保证金，若项目废弃，用于土地复原；取消和地方政府签署的保密协议。
她此番发言的背景是，投资者正在重新审视一批新锐AI云厂商的估值与生存能力：这类企业大多获得英伟达重仓支持，靠大举借贷、在缺乏长期运营记录的情况下疯狂上马数据中心。
彼得森・科里奥口中需要收紧标准的企业，大概率就包含这类厂商。与此同时，谷歌也透露，自身算力需求将越来越多地依赖第三方算力资源，但没有披露具体合作方。SpaceX就是其中之一。谷歌需要外部算力，是因为自建数据中心的速度跟不上需求。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
“Before moving into space, we should explore all possible options on the ground.” Amanda Peterson Corio said at the WTF summit. The Summit brought together senior female practitioners in science, technology and finance. As she made that point, SpaceX ' s orbital data centre planning was one of the important logic underpinning its market value of approximately $1.8 trillion.
Peterson Corio also referred to the current rise in opposition of American society to ground data centres and the fact that a large part of the responsibility lay with a group of new enterprises with low qualifications that should be subject to higher regulatory standards. She stated that before this round of AI-calculations, deep-tilled data centres had been built for many years with “standardized practices” and related projects had rarely been controversial.
The launch cost of placing calculated hardware in orbit was “extremely high, up to 100 times the cost of ground construction”, according to Peterson Corio. In addition to costs, operators need to address the maintenance and upgrading of on-orbit equipment. On the ground, data centre staff are ready to replace the server and rewiring the commissioning system.
Despite a cautious approach to space computing, the Google Research Team announced this month that it would work with the satellite image company PlanetLabs to explore the feasibility of a large space data centre cluster using satellites. Within the framework of the project entitled “Suncatcher” Google has launched a prototype of a small space server into space, carrying a joint custom-made version of the AI Block Processing Unit (TPU).
Last fall, Google researchers involved in the Lighting Project, together with the company's chief researcher, James Manyka, published a paper: if the cost of launching falls to $200 per kilogram and below, the corresponding calculator costs of unit effort in space are expected to be roughly equal to the ground level. According to Google researchers, this threshold is expected to be reached around 2035, although the point of time will depend on progress in the development of the SpaceX ship and the speed of frequency increase.
But others are more optimistic. The CEO of SpaceX, Elon Mask, said that the company would have a space data centre as soon as next year; while the competitor of Amazon and SpaceX, Jeff Bezos, predicted that a space data centre with up to 1 kiva would be available in 10 years. It has also been reported that Google and SpaceX have explored options for cooperation in building on-orbit data centres.
Peterson Corio did not completely reject the idea of space computing, which she described as “very interesting”; Google continued to organize such high-risk front-line projects because business was accustomed to “reflecting 10, 20, 30 years later”. But this technology is “never close”. CEO Open AI Sam Ottman's statement is more direct, considering the idea of moving data centres into space “unrealistic” under the current conditions, and questioning the radical timetable given by Mask.
Real bottlenecks: licensing, not electricity
In response to the argument put forward by Musk and others that the United States electricity supply is not enough to support the expansion of AI's computing power, Peterson Corio made different points.
"Is there enough electricity? Of course it is." She said. The real bottlenecks are the approval process required for project networking and the lengthy assessment study. The Tapestry project was launched at Alphabetical X Laboratory, which is working with PJM, the largest power grid operator on the east coast of the United States, to use AI to accelerate the relevant assessment process.
In Texas, the county ' s electricity regulatory agency has slowed down some data centre project approvals and strengthened controls to avoid new computing facilities that place additional grid burdens and higher electricity prices.
Other industry executives also noted that the power grid itself was far from being used at the desired level. In a recent interview with Brian Janos, co-founder of the data centre land and power service, Coverleaf Infrastructure, it was mentioned that on average about 45 per cent of the grid capacity was permanently idle in response to peaks. For the vast majority of the day, the grid has excess capacity to cover the needs of the data centre.
It is worth noting that Google has an advantage over cloud operations counterparts in terms of the Power Use Efficiency (PUE) indicator, which measures the data centre ' s computing capacity efficiency and represents the ability to extract more from the same power.
With regard to the rise of anti-data-centre waves in the United States, Peterson Corio attributed it to “a large number of new players who lacked industry experience in collecting land and building their calculus,” but did not name specific businesses.
In recent years, there has been a large influx of beginner developers into the AI data centre track, but more than just new players, and Oracle, a well-known enterprise, has experienced similar resistance from the local population.
Higher industry standard
In any case, she supported industry regulation and higher standards. She did not propose specific provisions, but gave an example of Google ' s increasing input to the local communities of the project, while setting up a plan to avoid the risks posed by the tailings of the project.
This included a commitment to cover the cost of grid upgrades on a continuous basis over a period of 15 to 20 years, and to avoid “forced billing” by local communities after projects were discontinued. Such cooperative models, which pre-commit the costs of grid infrastructure, have now landed in more than nine states in the United States and are “extensively expanding”.
The supporting criteria also include: full payment of electricity fees and property taxes by Google; payment of bonds for land restoration if the project is abandoned; and cancellation of confidentiality agreements with local governments.
In the context of her presentation, investors are revisiting the valuation and viability of a group of new AI cloud manufacturers: Most of these enterprises are heavily supported by British Wida, which is based on large-scale borrowing and demented access to data centres in the absence of long-term business records.
There are companies in Peterson Corio that need to tighten standards, and the probability is that they include them. At the same time, Google has revealed that it will increasingly rely on third-party computing resources for its own computing needs, without disclosing specific partners. SpaceX is one of them. Google requires external calculations because self-contained data centres cannot keep pace with demand.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.