Summary

  • Across the thirteen documents read for this piece on 22 November 2026, Akamai prints its edge place count nine times in six different forms — 4,000+, more than 4,300, more than 4,350, almost 4,400, over 4,400 and 4,400+ — and only the annual filing attaches a measurement date.
  • The unit that count denotes is defined once, in a glossary, as data centres; the same filing admits dependence on third-party co-location to house its servers, and its partner programme has network operators supply the rack space, power and addresses at no cost to Akamai.

Akamai has spent twenty-five years selling reach, and reach is the one thing a distributed platform cannot fake. So the company prints a number for it, and has printed the same number for years: 4,000+ edge PoPs on its infrastructure page, beside 1,200+ networks and 130+ countries, with no date anywhere in the text.

Then the same figure appears in six other forms across the other documents read for this piece. A blog published on 11 May 2021 puts it at more than 340,000 servers in more than 4,000 locations, across more than 1,000 cities, in 135 countries, inside more than 1,400 networks. A 2024 impact report prints 4,300+ edge PoPs. The 2025 impact report prints 4,400+. A press release dated 16 March 2026 headlines 4,400 edge locations and says over 4,400 in the same document. A blog dated 14 May 2026 says almost 4,400 points of presence and, a few paragraphs later, almost 4,400 global edge locations. The newest of the nine, published on 10 November 2026, says more than 4,350 edge locations — twelve days before this piece, a smaller number than the March release.

Six forms of one quantity in five years is not a scandal. It is what happens when a number is copied onward from deck to page to filing without anyone restating the counting rule beside it. The interesting part is not the spread. It is that nine printings across thirteen documents produce exactly one sentence with a date attached to it: in the Form 10-K for the year ended 31 December 2025, filed on 20 February 2026, Akamai writes that as of that date its infrastructure comprised core and distributed compute sites, more than 4,300 edge points-of-presence in over 130 countries and approximately 700 cities, and a global network integrated with roughly 1,200 network partners.

That sentence is a genuine improvement on the marketing page, and it is the only one of the nine that a reader can date. Everything else floats: a publication date is when someone wrote the page, not when anyone counted the network.

Now the unit. Search the thirteen documents for a definition of the thing being counted and you find one, in a cloud CDN glossary: points of presence are data centres distributed around the world that serve content to nearby users. As definitions go it is honest and generic, and it tells a reader almost nothing about the 4,300. A data centre is a building, a hall inside a building, a cage inside a hall, or a brand name on a partner's floor plan. Nothing in the reviewed set says which, and nothing says whether one hall holding three separate Akamai clusters counts once or three times, or whether a city with forty sites counts as forty or as one.

That question would be academic if Akamai owned every site. It does not, and its own documents say so. The FY2025 filing's risk factors record dependence in part on transmission capacity provided by third-party telecommunications network providers and co-location facilities to house its servers and equipment. The network partner programme FAQ, whose footer reads 09/16, is plainer still: partners agree to colocate Akamai servers within their own networks, and they provide the rack space, the power, the IP address space and occasional hands-and-eyes support at no cost to Akamai.

So part of the footprint is other operators' real estate. Whether those deployments are inside the 4,300+ is a question no reviewed document answers. Two readings are possible and the documents do not choose between them: either the count includes partner-hosted clusters, in which case the number measures reach across other people's buildings, or it excludes them, in which case the number understates the platform. A count whose unit cannot be resolved either way has stopped being a measurement and become a slogan with an integer attached.

Here is my own arithmetic, clearly labelled as mine and not the company's. On the 2021 pair of figures, more than 340,000 servers over more than 4,000 locations gives 85 servers per printed location — both operands are floors, so the quotient is a floor too, and it is my division, not Akamai's. On the May 2026 pair, roughly 300,000 servers over almost 4,400 points of presence gives about 68 servers per printed place. Same method, five years apart, and the servers per place falls while the places rise. Neither document performs that division. Neither document explains the direction.

Whether it reflects larger sites, a different counting rule for what a location is, or two unrelated marketing vintages, the reviewed set does not say, and I will not guess on the company's behalf.

The capacity layer is worse, and more useful. The 2025 impact report prints a 2025 peak traffic event of 395.84 Tbps. It also prints a figure of 33% for network capacity growth year over year, with no base quantity printed anywhere beside it. And it prints an energy intensity of 0.59 MWh per Gbps of capacity, in a report that never states the capacity total that intensity divides.

Three numbers, three different things, and only one of them is complete. A peak traffic event is a record of demand met at a moment; it is not provisioned capacity and must never be treated as such. A growth percentage without a base is a ratio with one operand withheld: 33% of what remains a question the report declines to answer. An intensity stated per Gbps of a total that is never printed cannot be inverted by a reader who wants the total. Here is a comparison I can make and the report cannot stop me making: the same company's 2024 impact report prints a 339 Tbps peak event, so 395.84 over 339 is 16.8% higher. That is my arithmetic.

It is not a refutation of 33% capacity growth — a peak event and a capacity total are different dimensions and I will not join them. But it does show what happens when only one of two numbers is dated and bounded: readers build their own comparisons out of the pieces that were left lying around.

Then the physical layer, which is where this record becomes strange. The only figures in the thirteen documents that describe actual ports and actual transport sit in a LINX conference slide deck dated November 2024, prepared by an Akamai network architect for a London Internet Exchange meeting: about 19,000 servers in the United Kingdom, UK daily peak traffic of 17-20 Tbps spread across 12 cities, global daily delivery of over 300 Terabits per second, backbone bundles of n100G and n400G with most bundles above 800G, and total long-haul capacity above 100 Tbps. If the UK fleet were spread evenly across those twelve cities, that is about 1,580 servers per city — my conditional division, and the slide claims neither evenness nor that all UK servers sit in those cities.

That deck is unaudited, dated, hosted by the exchange rather than by the company, and by some distance the most informative document in the set. It is also the only one. The audited filing prints a place count and no ports; the engineering slide prints ports and no audit; and the marketing pages print neither, faster.

PeeringDB's self-reported record for AS20940 fills part of the remaining gap with a band rather than a number: traffic level 100+Tbps, heavy outbound, 20,000 IPv4 prefixes and 6,000 IPv6 prefixes, open peering policy. It is self-reported, explicitly not independently verified, and it is one network record out of the several Akamai operates. A band is not a capacity measurement. It is a self-declaration that a threshold has been crossed.

What this record cannot answer, then, is precise and short. No reviewed document prints a provisioned capacity total in Tbps, a port count, a wavelength count or a fibre-pair perimeter. No reviewed document states whether partner-hosted deployments count inside the headline totals. No reviewed document reconciles the opposite directions of the place count and the printed server count. And no reviewed document explains why a report that can measure peak traffic to two decimal places cannot print the base under its own growth rate.

None of that is concealment. It is the ordinary condition of a company whose infrastructure claims serve marketing, sales engineering and securities disclosure at once, with three different standards of proof and no shared footnote. But the consequence is real, and it belongs to whoever is deciding whether to buy, peer with, or depend on the platform: the place count that gets quoted in every comparison of distributed platforms is a number without a unit, and every such comparison silently inherits it. A buyer can price a contract. A peer can measure a port. Nobody can check 4,400 of anything until someone prints what one is.

Boundaries

  1. No reviewed document states whether a partner-hosted or operator-hosted deployment counts as one PoP inside the 4,300+ or 4,400+ totals, and the partner programme FAQ relied on here carries a footer date of 09/16, so the mechanics it describes may have changed.
  2. The text of Akamai's own pages was read through a rendering of the same URLs rather than opened directly on 22 November 2026. An as-of date may exist in an element or script that the retrieved text did not carry, so every undated-page absence here is an absence in the text retrieved, not a finding about the live page.
  3. The FY2025 filing sentence was read through a provider rendering of the sec.gov accession URL, not a full-document parse; no section other than the Item 1 overview and the quoted risk-factor passage was read.
  4. Both impact reports and the partner FAQ were read through renderings of the PDFs rather than as rendered documents; figures elsewhere in either report were not examined.
  5. The LINX deck is dated November 2024, is unaudited and is hosted by the exchange; no later Akamai engineering deck was read, so those bundle sizes and that long-haul total stand as one dated engineering statement, not a series.
  6. PeeringDB is operator-self-reported and explicitly not independently verified; 100+Tbps is a coarse band, and no PeeringDB record for any other Akamai ASN was read here.
  7. No port count, wavelength count, provisioned capacity total or fibre-pair perimeter was located in any reviewed document. That is a count over thirteen named documents read on 22 November 2026, not a claim about what Akamai publishes elsewhere.
  8. No reviewed document explains the opposite directions of the place count and the printed server count, and none states why 33% capacity growth arrives without a base.
  9. No reviewed document reconciles the different place counts; the spread may be rounding, update cadence, differing measurement dates or a different counted object, and this piece does not choose among them.

Readings I rejected

X01 A peak traffic event is not provisioned capacity, and the two are never added or compared as equals. X02 A growth percentage whose base is unpublished is not a measured increase in capacity. X03 An energy intensity per Gbps of capacity cannot be inverted into a total by a reader who is not given the total. X04 4,400+ locations is not 4,400 buildings: a glossary calls a PoP a data centre, the filing admits third-party co-location, and the partner FAQ has operators supply the rack space. X05 A falling printed server count is not proof of a shrinking fleet; it is a change in a printing.

X06 PeeringDB's 100+Tbps band is a self-reported bucket, not a capacity measurement. X07 One ASN's prefix count is not the edge network's surface. X08 A conference slide's long-haul total describes backbone transport, not edge capacity, and it is unaudited. X09 A place count covering 130+ countries does not divide into a per-country or per-city count. X10 A footprint count is not a redundancy guarantee, and a traffic peak is not an availability record.

Sources read on 22 November 2026

Akamai Global Infrastructure · Form 10-K for the year ended 31 December 2025 · Form 10-K for the year ended 31 December 2024 · 2025 Impact Report · 2024 Impact Report · Akamai Content Delivery in the UK, LINX123 · PeeringDB AS20940 · The Internet Has a Front Door · Akamai Functions business case · AI Grid press release · Why Developers Are Writing Apps on Our Edge Platform · What Is a Cloud CDN? · AANP Program FAQ · Akamai in the BTW directory