Flopper.io Flopper.io
Data quality

Corrections

What we got wrong, what the page says now, and why. Newest first. Spotted something? Tell us and it lands here.

How figures are sourced and labelled is described on the methodology page. To report an error, use the contact form with the page URL and the source you are comparing against.

  1. Systems ranked on one basis

    /systems
    Before
    The headline FLOPS column ranked a mix of sparse figures (where a vendor published one) and dense figures (where it did not) under a single "Peak (sparse)" label.
    Now
    The default ranking and label are dense only. A Dense | Sparse toggle switches the column; systems without a published sparse figure show "not published" rather than their dense number.
    Why
    Dense and sparse throughput are not comparable, and a dense-only system was being labelled sparse.

  2. Homepage no longer derives system totals the vendor never claimed

    /
    Before
    System and rack FP16 totals were always the per-accelerator peak multiplied by the accelerator count, which published 560 PF FP16 for the Huawei Atlas 950 SuperPoD although Huawei publishes no 16-bit total.
    Now
    Stored vendor aggregates take precedence; a figure is derived only when no stored row exists, and derived cells carry a dagger and an "aggregate theoretical peak" footnote. The Atlas 950 SuperPoD shows no FP16 total.
    Why
    A multiplied figure is a theoretical ceiling, not a vendor claim, and must be labelled as such.

  3. Estimated maximum power labelled as an estimate

    /gpus
    Before
    Parts with no published maximum board power showed TDP x 1.15 under the same "Max Power" label as vendor figures (147 of 302 parts).
    Now
    Those parts read "Max power (Flopper estimate)", render in muted text with an "est." suffix, and show the formula.
    Why
    An estimate must never sit beside an official specification without distinction.

  4. One definition of a live price

    /pricing
    Before
    The homepage counted every active listing of any pricing type as a live price (244, including one quarantined outlier); /pricing and the GPU table each used a different rule.
    Now
    A live price is an active on-demand listing above $0 from a cross-listable provider, not flagged as an outlier, confirmed within the last 30 days (167 at the time of the change). Every headline count, "from" price and median uses this rule.
    Why
    Spot, community and reserved rates are not the price a buyer gets on demand.

  5. Median takes one price per provider

    /pricing
    Before
    Medians were taken over every listing, so a provider with eight configurations counted eight times.
    Now
    The median is over each provider's cheapest live on-demand listing, and the methodology is shown beside every median.
    Why
    A provider with many SKUs should not dominate a market figure.

  6. Unreported power is no longer shown as 0.0 MW

    /datacenters
    Before
    Country and owner pages summed unreported power as zero, so /datacenters/country/Mexico showed "0.0 MW" in the page, its title and its meta description.
    Now
    Pages show "No verified power data" when nothing is published, and "n of m report power" when coverage is partial.
    Why
    A missing total is not a zero.

  7. Country names made consistent

    /datacenters
    Before
    "United States of America" (184 facilities) and "United States" (52) were two separate countries; "Korea (Republic of)", "Philippines (the)" and the long United Kingdom form leaked into titles, and the countries total overcounted.
    Now
    Common names throughout (United States, South Korea, United Kingdom, Philippines); the old URLs redirect permanently.
    Why
    One country, one facet.

  8. Datacenter index title read 786 clusters; database held 808

    /datacenters
    Before
    The page title hard-coded a count from an earlier snapshot while the description beneath it was computed live.
    Now
    Every public total on the site comes from one catalog-stats service and carries a "Database totals as of" date.
    Why
    Counters were computed independently in eight places and had drifted.

  9. Datacenter index was invisible to crawlers

    /datacenters
    Before
    The page rendered only in the browser, so search and AI crawlers received a 4 KB shell with no heading, no table and no FAQ markup.
    Now
    The page is server-rendered in full.
    Why
    The flag disabling server rendering was a misreading of what it does.