Infrabench

Published specs, corrected for reality.

A rack is rated for 600 kW, a drive for five years, an interconnection queue in gigawatts. None of it survives contact with a real site. Infrabench is a set of free tools that correct those numbers against public data, with every constant on the page.

177 racks on paper, 30 MW at 120 kWracks after PUE 1.2racks fit, of 250 on paper

  • Fits
  • 42 lost to overhead at PUE 1.2
  • 31 held back by a 15% margin

  1. Rack Budget

    You have the megawatts. What actually fits?

    Power, cooling and floor area for an AI hall. Start from a megawatt budget or from a fleet size and get the other side, plus the heat you have to reject.

    Data: NVIDIA platform disclosures, Uptime Institute PUE survey

    Worked example
    Power at the fence
    30 MW
    Overhead
    PUE 1.2, 15% margin
    Rack
    GB200 NVL72, 120 kW
    Racks that fit
    177
    Heat to reject
    6,039 tons
    White space
    7,080 sq ft
    Open tool: Rack Budget

  2. Crash Cart

    You cannot pull them all. Which ones this week?

    Drive failure triage on thirteen years of public fleet telemetry: a live pull list, ranked by risk times blast radius, that replays week by week.

    Data: Backblaze Drive Stats, about 300,000 drives

    Worked example
    Fleet
    3,000 drives, 4 to 22 TB
    Redundancy
    Erasure-coded 17+3
    Crew capacity
    12 pulls a week
    Pull this week
    12 × Seagate 12 TB
    Still in the queue
    344 drives
    Queue clears in
    34 weeks
    Open tool: Crash Cart

  3. Where the power is

    The chips are the easy part. Where is the power?

    Siting under a power constraint. Ten US grid regions ranked by the earliest realistic date a utility can energize your load, drawn against your deadline.

    Data: LBNL Queued Up, EPA eGRID, EIA

    Worked example
    New load
    300 MW
    Needed by
    2030
    Power source
    Grid only
    Earliest window
    ERCOT, 2028–2030
    Regions that make 2030
    1 of 10
    Next fastest
    Southeast, 2029–2031
    Open tool: Where the power is

  4. Outage replay

    The dashboard was green. How long had it been wrong?

    Six public cloud incidents replayed against the provider’s own retrospective: what the status page said, and how long after customers started failing.

    Data: retrospectives from AWS, Azure, Google Cloud, Cloudflare and Fastly

    Worked example
    Incident
    Google Cloud, 12 June 2025
    First customer impact
    17:51 UTC
    First status post
    19:05 UTC
    Status page stayed green
    74 min
    A 15-minute alert beats it by
    59 min
    Median gap, all six
    48 min
    Open tool: Outage replay

How the tools are built

Free tools for the people who build data centers, on public data with every constant on the page.

  • Public data onlyevery input is something you could download yourself
  • Every constant on the pagethe assumptions and the arithmetic behind every number
  • One decision outeach tool ends by naming what will actually stop you

Built by Adrian Mucha

A product designer for the software most designers avoid: data center tools, AI tooling and operator consoles where a wrong click costs real money.