CONNECTED devices now regularly double as digital hoovers: equipped with a clutch of sensors, they suck in all kinds of information and send it to their maker for analysis. Not so the wireless earbuds developed by Bragi, a startup from Munich. They keep most of what they collect, such as the wearers’ vital signs, and crunch the data locally. “The devices are getting smarter as they are used,” says Nikolaj Hviid, its chief executive.

一些联网的设备配备一些传感器，收集各种信息，发送给制造商用于分析。它们好似一个数字的吸尘器，在以平均两倍的速度增长。慕尼黑的创业企业Bragi开发的无线耳机却并不如此。他们保留了大部分收集的信息（如穿戴者的生命体征），在本地数据处理。该公司首席执行官Nikolaj Hviid说：“这些设备会越用越聪明。”

Bragi’s earplugs are at the forefront of a big shift in the tech industry. In recent years ever more computing has been pushed into the “cloud”, meaning networks of big data centres. But the pendulum has already started to swing: computing is moving back to the “edge” of local networks and intelligent devices.

Bragi的耳塞可谓在科技行业重大转变中处于最前沿地位。近年来，越来越多的计算开始进入“云”状态，即大数据中心网络。不过，钟摆开始摆动：计算正趋于本地网络和智能设备的那一“边”。

As with the rise of the cloud in the early 2010s, the shift will cause upheaval. Many startups will try to ride the trend, as will incumbents such as hardware makers. But the real fight will be over who colonises the edge and, in particular, which firms will control the “internet of things” (IoT), as connected devices are collectively called. Will Amazon Web Services (AWS), Microsoft and other large cloud providers manage to extend their reach? Or will the edge be the remit of a different set of firms, including makers of factory equipment and other sorts of gear?

随着2010年代初云计算的兴起，这一转变将掀起浪潮。与众多现有的硬件制造商一样，许多创业企业也试图把握此次潮流。 不过，他们真正争夺的是本地网络和智能设备这一边——即“物联网”（IoT），众多相互连续的设备的统称——的控制。亚马逊网络服务（AWS）、微软和其它大型云供应商是否会设法扩大其影响力？ 或者说，这是否是包括工厂（及其他）设备制造商在内不同公司的职权范围？

Since emerging in the 1950s, commercial computing has oscillated between being more centralised and more distributed. Until the 1970s it was confined to mainframes. When smaller machines emerged in the 1980s and 1990s, it became more spread out: applications were accessed by personal computers, but lived in souped-up PCs in corporate data centres (something called a “client-server” system). With the rise of the cloud in the 2000s, things became more centralised again. Each era saw a new group of firms rise to the top, with one leading the pack: IBM in mainframes, Microsoft in personal computers and AWS in cloud computing.

商业计算自20世纪50年代诞生以来，在集中化和分散化之间摇摆不定。直到20世纪70年代，商业计算还局限于大型计算机。小型计算机在20世纪八九十年代问世后，商业计算变得更加分散：个人电脑可以访问应用程序，而应用程序位于企业数据中心（即“客户机-服务器”系统）中的电脑里。但随着2000年代“云”的兴起，商业计算又变得集中了。每个时代都有一批新公司跻身前列，但榜首只有一个：大型机为IBM，个人电脑为微软，云计算为AWS。

Better technology is one reason why computing is again becoming more distributed. Devices at the edge, from smartphones to machinery on the shop floor, are becoming more intelligent. Equipped with powerful processors, they can now tackle computing problems that a few years ago needed a fully loaded server. As for software, its increased flexibility means it can function well on the edge. Many applications are now “virtualised”, meaning they exist separately from any specific type of hardware: code can thus be packaged in digital “containers” and easily moved around within data centres—and, increasingly, closer to the edge.

计算再次变得分散化的一个原因是更先进的技术的出现。从智能手机到工厂车间里的机械设备，处于边缘地带的设备变得越来越智能。在配备了性能强大的处理器后，现在的设备可以解决几年前一个服务器满负荷运行才能解决的计算问题。软件变得更加灵活，也可以运行良好。现在许多应用都“虚拟化”了，即不依赖于任何特定类型的硬件，代码可以打包在数字“容器”中，在数据中心内便捷传输。这离“边缘”更近了一步。

Demand for computing at the edge is growing, too, often for non-technical reasons. Many countries have laws that require data to stay within their borders or even within the walls of a company. Firms want to use data but, worrying about leaks, often prefer to keep their own information inhouse. Consumers, for their part, care about privacy, which Bragi hopes to address with its self-sufficient earplugs.

对于边缘地带的计算的需求也在增长，而且往往出于非技术层面的原因。许多国家的法律要求数据不得出境，甚至是企业外部。企业既希望使用数据又担心泄漏，往往倾向于将信息保留在自身内部。就消费者而言，他们关心隐私，这正是Bragi希望通过“自给自足式”耳机解决的问题。

The dominant narrative in the tech industry—that most data are best crunched centrally in the cloud—is also undermined by the fact that many new applications have to act fast. According to some estimates, self-driving cars generate as much as 25 gigabytes per hour, nearly 30 times more than a high-definition video stream. Before so many data are uploaded, and driving instructions sent back, the vehicle may well already have hit that pedestrian suddenly crossing the street.

科技行业的主流观点是大部分数据最好在云端集中处理。然而，由于许多新应用程序运行迅速，这一观点也难以立足。据估计，无人驾驶汽车每小时可产生25G的数据，比高清视频流多出近30倍。再加上发回的驾驶指令，在如此庞大的数据上传之前，汽车可能早就撞上了某个窜出在马路上的行人。

Changing economics are another consideration. The faster adjustments can be made—for instance, to optimise the operations of a machine in a factory—the bigger revenue gains tend to be. That means data are often best analysed as they are captured, which needs to be done locally. The costs of transferring, storing and processing data in the cloud can be avoided too.

另一个考虑是经济性的调整。调整越快收入就会越多，如优化工厂机器运行。这意味要在获取数据时第一时间进行分析，而这一环节要在本地完成。这也能省去云端传输、存储和处理数据的成本。

Car-boot brains

无人驾驶汽车的控制大脑

These constraints explain why services using artificial intelligence (AI) are increasingly split in two, much like client-server applications, explains Pierre Ferragu of Bernstein Research. The algorithms of autonomous cars, for instance, are first trained in the cloud with millions of miles of recorded driving data; only then are they deployed on powerful computers in the boot, where they steer the car by interpreting live data. Similarly, many video cameras used for surveillance now ship with face-recognition software trained in the cloud, as does Apple’s latest iPhone model. In November, Google announced an addition to TensorFlow, its AI technology, which allows developers to deploy algorithms to mobile devices.

伯恩斯坦研究公司（Bernstein Research）的皮埃尔·费拉格（Pierre Ferragu）解释称，这些束缚表明使用人工智能（AI）的服务越来越多地分裂成两个部分，就像客户端和服务器程序一样。例如，无人驾驶汽车的算法首先要在云端数百万英里的驾驶数据中进行训练，才能部署在汽车里性能强大的电脑里，通过分析实时数据来驾驶汽车。同样，许多监控摄像机现已配备云端训练的面部识别软件。苹果最新款的的iPhone也是如此。 去年11月，谷歌宣布对其人工智能系统TensorFlow进行升级——这个系统能允许开发人员将算法部署到移动设备上。

But in many cases even the training of algorithms must happen locally for AI applications to make commercial sense, argues Simon Crosby, chief technology officer of Swim, a startup. For instance, sending the four terabytes of data generated daily by traffic lights at intersections in Palo Alto, in Silicon Valley, to a cloud provider for processing would cost thousands of dollars a month. Swim has built a system that does the equivalent job for few hundred dollars by learning from the data on the fly as they are generated.

不过，Swim首席技术官Simon Crosby认为，在许多情况下，算法的训练必须在本地进行，人工智能程序才能具有商业价值。例如，在硅谷Palo Alto的十字路口，交通信号灯每天产生4TB的数据并发送给云计算提供商进行处理，每月将花费数千美元。Swim新研发的系统，在数据产生时就对其进行分析，完成同样的任务只需几百美元。

Although a shift to the edge is now generally acknowledged to be under way, opinions are divided over how it will change the technology industry. Nobody expects the “end of cloud computing”, to quote the provocative title of a podcast by Peter Levine of Andreessen Horowitz, a leading Silicon Valley venture-capital firm. He himself predicts that centralised clouds, in particular those of Amazon, Google and Microsoft, will continue to grow.

虽然人们普遍认为物联网是大势所趋，对于其将如何改变科技行业则各有己见。引用硅谷风险投资公司Andreessen Horowitz的Peter Levine在其播客中使用的一个耸动标题，没有人期待“云计算的终结”。他自己预测，集中式云计算，特别是亚马逊，谷歌和微软的云计算将继续增长。

But smaller and more local data centres are springing up everywhere. Firms such as EdgeConneX and vXchnge have built networks of urban data centres. Vapor IO, a startup, has developed a data centre in a box that looks like a round fridge and can be quickly put in any basement. Makers of telecoms equipment, including Ericsson and Nokia, as well as network operators, talk a lot about “mobile edge computing”, which amounts to putting computers next to wireless base stations or in central switching offices. Some also speculate that one reason why Amazon last year bought Whole Foods, a chain of grocery shops, for nearly $14bn, was to accumulate property for local data centres.

不过，各地正涌现出许多小型的本地型数据中心。EdgeConneX和vXchnge等企业建立了城市数据中心网络。创业企业Vapor IO开发了一种便携数据中心，外型像一个圆形冰箱，能迅速部署在地下室。爱立信和诺基亚等电信设备制造商以及一些网络运营商都在研究“移动边缘计算”，这相当于把计算机放在无线基站旁或中央交换局内。也有人推测，亚马逊在2017年以近140亿美元收购连锁商店Whole Foods的原因之一就是为当地数据中心积累资料。

Computer makers see the shift as a chance to regain lost territory. Dell EMC and HP both want to sell more gear to firms keen to crunch data locally. But they are limited in how far they can move to the edge, says George Gilbert of Wikibon, a consultancy. These firms know how to sell commodity hardware to IT departments, but most IoT gear will be more customised, requires special software and is sold to people managing machinery. Cisco, which sells all kinds of internet equipment, seems well placed.

计算机制造商将这一转变视为收复失地的机会。戴尔/EMC和惠普（HP）都希望将更多的设备卖给热衷于在本地收集数据的企业。不过，咨询机构Wikibon的乔治·吉尔伯特（George Gilbert）表示，这些IT企业继续向前边缘领域进发的能力有限。这些企业懂得如何向IT部门销售商品硬件，但大多数物联网设备将更加定制化，需要特殊的软件，并销售给管理机器的人员。销售各种互联网设备的思科（Cisco）似乎处于优势地位。

Big cloud-computing providers are also trying to colonise the periphery. In May Microsoft changed its slogan from “mobile first, cloud first” to “intelligent cloud and intelligent edge”. It sells services that dispatch software containers with AI algorithms to any device. AWS’s portfolio now includes a service called Greengrass, which turns clusters of IoT devices into mini-clouds. In buying the Weather Company for $2bn in 2015, IBM wanted weather data, but also thousands of “points of presence” for edge computing.

大型云计算提供商也试图占领周边市场。今年5月，微软的宣传语由“移动优先，云优先”改为“智能云和智能端”，推出用AI算法将软件容器分派到任何设备的服务。AWS的产品组合现在也包含一项名为Greengrass的服务，用于将物联网设备的群集变成微型云。IBM在2015年以20亿美元的价格收购天气预报公司的时，希望得到不仅是天气数据，还有成千上万的边缘计算“存在点”。

Whoever prevails, computing will become an increasingly movable feast, bits of which can be found in even the smallest devices. Processing will occur wherever it is best placed for any given application. Data experts have already started using another term: “fog computing”. But the metaphor is a bit, well, foggy. Better, and more poetic, would be “air computing”: it is everywhere and gives things life.

无论谁占上风，计算能力将愈加移动化，即使在最小的设备中也能发现这一趋势。可以根据程序选择最佳位置进行数据处理。数据专家开始使用另一个术语：“雾计算”。 但是，这个比喻略显模糊。更好、更诗意的一个词是“空气计算”：它无处不在，给物以生命。

编译：morningvicky

审校：钟楠

编辑：翻吧君

来源：经济学人