BRENNEN

Industries / Technology / Software and SaaS / Cloud and data centers

Planning capacity

This is the work of deciding how much computing capacity to build, and getting the most use out of what is built.

7

Milestones achieved, from 2001 to 2026.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2001

    Built ahead of demand

    Exodus Communications, a web-hosting company whose data centers housed more than 4,000 companies' websites, filed for bankruptcy on 26 September 2001 with $3.5 billion in debt. In November Cable & Wireless agreed to take 26 of Exodus's 44 operating data centers, plus four other fully equipped centers.

    LessonCapacity built on borrowed money for forecast demand can be left half used when the demand does not come.

    informationweek.comcomputerworld.com

    Credit

    Exodus Communications; reported by InformationWeek and Computerworld

  2. 2007

    Selling the same watts twice

    Xiaobo Fan, Wolf-Dietrich Weber and Luiz André Barroso measured the power drawn by groups of up to 15,000 Google servers for about six months. Whole data centers never reached their theoretical peak draw, with a gap of almost 40%. That gap, they wrote, could be used to put in more computers within the same power budget, a practice called oversubscription.

    research.google

    Credit

    Google; paper at ISCA 2007

  3. 2009

    Spare servers rented cheaply

    On 14 December 2009 Amazon launched EC2 Spot Instances, which let customers bid on unused Amazon EC2 capacity and keep it as long as their bid was above a changing spot price.

    aws.amazon.com

    Credit

    Amazon Web Services

  4. 2011

    A busy cluster made public

    Google published a trace of one Borg cell of about 12,500 machines for 29 days of May 2011, now under a CC BY license. Researchers at UC Berkeley, Carnegie Mellon and Intel Labs found the cluster booked for more than 100% of its CPU and more than 80% of its memory, while actual use stayed at about 60% of CPU and 50% of memory. Google's trace page lists 67 papers that used its traces.

    github.compdl.cmu.edu

    Credit

    Google (John Wilkes, Charles Reiss); analysis by Reiss, Tumanov, Ganger, Katz and Kozuch, ACM SoCC 2012

  5. 2017

    Bidding for servers dropped

    Technion researchers who studied Amazon's spot price histories wrote in 2013 that prices were usually "not market-driven" and more likely set at random within a narrow band by a hidden reserve price. In November 2017 Amazon dropped bidding: spot prices now "adjust slowly over days and weeks" according to supply and demand for spare capacity, and Amazon said interruptions averaged under 5% over the previous 30 days.

    LessonAn auction for spare capacity confused buyers, and a slow-moving posted price replaced it.

    doi.orgaws.amazon.com

    Credit

    Orna Agmon Ben-Yehuda, Muli Ben-Yehuda, Assaf Schuster and Dan Tsafrir (Technion); Amazon Web Services

  6. 2025

    Outside counts of AI capacity

    Since mid-2025 Epoch AI, an independent research group, has estimated the power, compute and cost of the world's largest AI data centers from satellite images and permits, and publishes the data under CC BY. As of 18 November 2026 its data covered 96 data centers, about 44% of deployed AI compute by its own estimate, with the largest at 1,037 MW of IT power (Anthropic-Amazon, New Carlisle, Indiana).

    epoch.ai

    Credit

    Epoch AI

  7. 2026

    Missed dates counted

    Sightline Climate tracked 777 data center projects over 50 MW announced since January 2024. It found that 26% of the 110 projects expected online in 2025 were delayed, and that of at least 16 GW due in 2026, only about 5 GW was under construction in February 2026.

    LessonA developer's announced opening date is a weak guide to when capacity arrives.

    latitudemedia.comsightlineclimate.com

    Credit

    Sightline Climate (Olivia Wang and co-authors); reported by Latitude Media

Watch for

  1. Racks sized to average power

    During AI training a rack's power draw can spike and drop. NVIDIA says the power shelves of its GB300 NVL72 racks store energy in capacitors during lulls and release it during spikes, and that when its engineers instrumented one rack training the Megatron language model, the peak power seen by the grid fell by 30% with the same GPU output. It says a data center can then provision power closer to average use instead of the peak, fitting more racks in the same power budget.

    Achieved whenan operator outside NVIDIA publishes metered peak and average power for a hall of such racks, and how many more racks it fitted under the same supply.

    developer.nvidia.com

    Claimed by

    NVIDIA, July 2025

Open

  1. 01

    An open ledger of promises

    A public, project-by-project record of announced data centers over 50 MW: the date announced, the megawatts and opening date promised, and the latest status with dated evidence (a permit, a satellite image, a utility filing), checked every quarter.

    The brief
    Credit · unclaimed
  2. 02

    Rack audits from a phone

    A tool that turns photos or a video of server racks into a rack inventory in NetBox, the open-source DCIM (data center infrastructure management) software. It reads each device's model and asset tag, the rack unit it sits in, and which cable goes into which port, then flags every difference from the records and the wiring plan.

    The brief
    Credit · unclaimed

Open 01 · the brief

An open ledger of promises

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public, project-by-project record of announced data centers over 50 MW: the date announced, the megawatts and opening date promised, and the latest status with dated evidence (a permit, a satellite image, a utility filing), checked every quarter.

Why it is still open

Sightline Climate's 777-project data is a paid product, and its free version is a summary. Epoch AI's open data covers 96 large AI sites. FracTracker Alliance's free US tracker lists about 1,415 sites with a status (proposed, under construction, operating, suspended, cancelled) but no promised or actual opening dates, and only for non-commercial use. Latitude Media reports that most developers, including the Stargate Project, Amazon, Meta and Vantage, do not share their timelines.

sightlineclimate.comepoch.aifractracker.orglatitudemedia.com

Built with what exists

Epoch AI's CC BY data center data; free Sentinel-2 satellite images through the Copernicus Browser; county permit portals; an AI model that pulls figures from press releases and filings, each checked by a person.

  1. First step

    For the 96 sites in Epoch AI's data, find each first public announcement and its promised opening date, then compare it with Epoch's estimated operating date.

  2. Done looks like

    200 projects, each with its announcement link, promised megawatts and date, and a quarterly status backed by dated evidence for a full year, plus a published slip rate set beside Sightline's 26% for 2025.

Open 02 · the brief

Rack audits from a phone

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A tool that turns photos or a video of server racks into a rack inventory in NetBox, the open-source DCIM (data center infrastructure management) software. It reads each device's model and asset tag, the rack unit it sits in, and which cable goes into which port, then flags every difference from the records and the wiring plan.

Why it is still open

Rack positions are still typed in by hand. OVHcloud's guide to its NetBox-based private cloud says that "today, the position of each node in the rack elevation is set manually", and that automatic placement from cabling is still being built. Roboflow, a computer vision company, says its software reads rack asset tags and can "Verify cabling against the build spec at integration, port by port", but it publishes no accuracy figures. No public test of photo-based audits against a hand-checked inventory turned up.

docs.ovhcloud.comroboflow.com

Built with what exists

NetBox (open source) and its API; a phone camera; a vision-language model; an object detector such as Ultralytics YOLO for ports and cable ends; open-source OCR such as Tesseract or PaddleOCR; a lab or university server room, with permission.

  1. First step

    With permission, photograph three racks in a university server room, audit them by hand, and see how many devices and ports an off-the-shelf vision model gets right.

  2. Done looks like

    20 racks photographed and audited by hand, with at least 95% of devices identified and placed at the right rack unit, and a published score for ports read correctly and for 5 to 10 deliberate wiring mistakes per rack caught. The images and the hand-made inventory are published so that others can score their own tools.