The Infrastructure File — The Architecture of the Data Center
The Distributed Machine: Redundancy Becomes the Architecture
Telephony and computing began in single rooms, served through hierarchies. A century of automation, copying, and routing moved the service into the network itself. The record runs from the La Porte exchange of 1892 to the availability zones of 2026, and it lands on a buildout now contested site by site.
The News in the File
The newest data-center plan in the record is not a campus. It is a scatter. Trade accounts filed on Friday, November 6, reported that Nvidia is examining a plan with the fiber carrier Brightspeed to place small data centers in more than 130 former telephone buildings across the United States—sites that already hold power connections and fiber—for a combined capacity reported at up to one gigawatt. Inside the project, according to the same accounts, the arrangement is described as an inference network: computing placed in many small facilities near the people requesting it, rather than in a few large ones far away. The companies have issued no announcement; the desk files the item as reported, and grades it again in the cocked hat below. The plan is filed here because of its shape. The shape is the subject of this report: the industry’s unit of construction is changing from the building to the network.
The counts describe the same change at the largest scale. Synergy Research Group counted 1,297 operational hyperscale data centers worldwide in late 2025, nearly triple the number of early 2018, with 770 more in planning, construction, or fit-out. Amazon Web Services operates 38 regions; Microsoft closed 2025 with more than 70 Azure regions and more than 400 data centers; Google Cloud stood at 42 regions and 127 zones. The facilities differ in size and owner. The architecture they share is the point: capacity held in many places at once, data copied between them as a matter of routine, and no single address at which the service lives.
The Single Room: Computing as It Began
Commercial computing began as one machine in one room, and the room was the service. IBM announced the System/360 family in April 1964, a line of compatible mainframes on which a customer’s programs could run across models. The machines stood in air-conditioned rooms, attended by operators, and the work arrived as decks of cards and left as printed paper. Time-sharing, added to the line with the Model 67 in 1966, let many users work at the same machine in the same hour. The user’s furniture was a terminal: in the 1970s, characteristically, an IBM 3270 display—a screen and keyboard with no computing of its own, which collected a full screen of entries and sent it to the mainframe when the operator pressed Enter. The arrangement was efficient and entirely dependent. A mainframe of the period could serve thousands of terminal users, and every one of them was served from the single room, or not at all.
Telephony began the same way, in a single room of a different kind. In the manual exchange, every call passed through an operator seated at a cord board, who took the request by voice and completed the connection by hand, plugging one cord into the jack of the calling line and another into the jack of the called line. The network above the local board was arranged as a hierarchy of five classes of office, from the Class 5 end office that served subscribers up through toll and primary and sectional centers to the Class 1 regional centers; tandem offices connected end offices to one another and to the levels above. The system’s research arm, Bell Telephone Laboratories, began moving to Murray Hill, New Jersey, in 1941; the transistor was demonstrated at Murray Hill in 1947. The hierarchy was orderly and it was a chain: each level depended on the level above it, and the operator’s hands were the switching fabric of the whole.
The Operator Leaves the Path
The first break in the chain was mechanical. Almon Strowger’s step-by-step switch, patented in 1891, let the caller’s own dial pulses move the switch directly, digit by digit, with no operator in the path. The first commercial step-by-step exchange opened at La Porte, Indiana, on November 3, 1892. The Bell System’s crossbar switch followed: the No. 1 Crossbar was installed at the Troy Avenue office in Brooklyn in 1938, and the No. 5 Crossbar, first installed at Media, Pennsylvania, in 1948, became the workhorse of the system. In 1965 the No. 1 Electronic Switching System put a stored program—a computer—in charge of the connections. Each step removed a pair of hands and a single point of attention from the path of a call. Switching became a function the network performed on itself, in every office at once, and the five-class hierarchy flattened as the offices grew able to reach one another directly.
The lesson the telephone network taught was structural, and computing learned it twice. A service concentrated in one room fails with the room. A service performed by a hierarchy fails at the level that fails. The systems that replaced both arrangements distributed the work itself—first the switching, then the content, then the computing—until the service no longer had a single address at which it could be interrupted.
The CDN Turn: Copies Near the Reader
The distribution of content began as an answer to a crowd. In 1995 Tim Berners-Lee challenged colleagues at the Massachusetts Institute of Technology to solve the congestion of a popular web site—the flash crowd that arrived at one server at one address and overwhelmed it. Tom Leighton, an MIT professor of applied mathematics, and Daniel Lewin, his graduate student, answered with consistent hashing, a method of spreading content across many servers so that the set could grow, shrink, and lose members without re-mapping the whole. Akamai Technologies was incorporated on August 20, 1998, by Leighton and Lewin with Jonathan Seelig and Randall Kaplan. The name is the Hawaiian word akamai: intelligent, clever—colloquially, cool.
The content delivery network changed where a web page lives. Under the Akamai model, copies of a publisher’s content are held on servers placed inside many networks, near the readers. A request is answered by a nearby copy. The publisher’s origin server remains, but it is no longer the place the public visits; it is the place the copies are made from. The failure arithmetic changed with the geography: a server under a flash crowd is one copy among many, and a server that fails removes one copy, not the publication. Within its first years Akamai was, by the accounts of the period, effectively the content delivery industry; the architecture it sold—many copies, served locally, refreshed from an origin—became the default shape of the public web.
The Modern Form: Zones, Regions, and One Address Everywhere
The hyperscalers applied the same principle to computing itself, and wrote it into their operating instructions. Amazon Web Services divides its infrastructure into Regions, each physically and logically independent, and each Region into three or more Availability Zones. An Availability Zone is one or more discrete data centers, physically separated from the zones beside it, with independent power, cooling, and network connections. AWS’s published reliability guidance instructs customers to deploy production workloads across at least two zones and to replicate data between them; the zones sit close enough for synchronous replication at single-digit-millisecond latency, so a copy is current, not eventual. Several of the company’s largest storage and database services replicate across zones by default. The design objective, in the document’s own terms, is the avoidance of single points of failure, including the underlying physical infrastructure.
Cloudflare carries the principle into the network’s addressing. Its network runs on anycast: every location announces the same IP addresses, and the internet’s routing delivers each request to the nearest location able to answer it. Cloudflare’s own documentation states the failure behavior plainly: because multiple nodes advertise the same address, if one node fails, requests are routed to another node in close proximity. There is no failover address to publish and no single facility whose loss changes the address of the service. The company reports data centers in 335 cities. The service is the address; the address is everywhere the network is.
The storage industry had named the underlying principle a generation earlier, and the facility now follows the disk. RAID arrays, standardized in the late 1980s, write data across multiple disks so that the failure of one disk is an event, not a loss; backup practice added copies held in different places against the failure of the building itself. The desk prints the analogy as an analogy, because it is exact at the level of design: the unit of engineering is no longer the single store of the data but the set of copies, the rules that keep them in agreement, and the routing that finds a surviving copy when one is gone. A modern data center is built the way the array was: replaceable, replicated, and one member of a set.
The Contested Ground, and the Structural Answer
The construction program is contested at the level of the individual site, and the record of the contest is counted. On November 3, voters in ten states decided 41 local measures on data-center construction; in Ballotpedia’s complete tracker, 40 went the way of tighter limits. Data Center Watch, as reported by Data Center Knowledge, counted more than 36 projects representing $162 billion of investment blocked or significantly delayed as of June 2025. Site selection has adjusted its vocabulary to the record: an industry account of 2026 practice describes the discipline as deliverability—the overlap of entitled land, available power, and a local community willing to partner on the project—and describes the market splitting into two tracks, hyperscale campuses of hundreds of megawatts and hyperlocal sites of one to twenty megawatts close to users.
Distribution is the architecture’s answer to a refusal, and the answer is mechanical rather than rhetorical. A workload deployed across zones and regions under the operators’ own published guidance does not have a hometown. Capacity refused in one jurisdiction is capacity placed in another; a site that is delayed delays one member of a set whose other members continue. The reported Nvidia–Brightspeed plan is the limiting case of the same logic: capacity distributed across more than 130 existing buildings, each individually small, each already holding the power and fiber connections a new campus waits years to obtain, and none of them individually the network. The data center began as a palace built to hold a unique machine. The record of 2026 shows it completing the telephone network’s migration: from the single room, to the hierarchy, to the network of copies—a service that persists because no one place is the service.
The Cocked Hat
Three bearings never give a point; they give a small triangle, and this paper prints the triangle. The scale figures carry one. The desk’s sources put Amazon’s footprint at 38 regions with 120 availability zones in one October 2026 accounting and at 38 regions with more than 100 zones in another; the counts were taken at different dates against a buildout adding zones continuously, and they agree on the structure while differing on the count. The Friday item carries a wider triangle: the Nvidia–Brightspeed plan rests on trade reporting that the companies have not confirmed in any announcement the desk can file, and its figures—130 buildings, one gigawatt—are the reported figures of a plan under examination, not of a construction program under way. The historical bearings are single-valued: the La Porte exchange opened on November 3, 1892; Akamai was incorporated on August 20, 1998; the No. 1 Crossbar went in at Brooklyn in 1938. The architecture those dates produced is the single-valued fact underneath the triangle of counts: the copies are the system now. The log proceeds to the Midday Watch.