Summary
- The only date on Huston's rolling IPv4 Address Report is a generation stamp: a copy read on 21 November 2026 carried "generated at 19-Nov-2026 07:58 UTC", a second copy read about five minutes later carried "generated at 21-Nov-2026 07:58 UTC" — the date moved two days, the clock did not move at all — while the same page prints 4,294,967,296 address values that no stamp can change.
- His January 2026 column puts the year boundary inside the sentence it belongs to, "some 237 thousand addresses in 2025, with some 3.687 billion allocated addresses at the end of the year", and ARIN's republication carried that wording across a change of host under a section heading that reads "IPv4 in 2024".
A number's authority lives in the words immediately around it. Huston's IPv4 corpus is an unusually clean place to see that, because the same author, working the same address-space series, publishes it in three registers: a page that regenerates, a column that is dated to a month, and a republication that is dated to a day. Each fixes a different kind of clock. Only one of them puts the measurement boundary where the number is.
What the stamp actually dates
The rolling report carries, in the text read for this piece, the sentence "This report generated at 21-Nov-2026 07:58 UTC." That is the entire date apparatus at the top of the document. It is a generation timestamp: it says when the page was assembled, not when the allocation series beneath it was collected, and not when the article was published. On its own it is a perfectly honest machine label. It only becomes a problem when a reader lifts a figure from the page and lets the stamp stand in for the measurement date, which is the one thing the stamp never claims to be (see the IPv4 Address Report).
The interesting part is what sits below the stamp. The same document prints, in the same text, constants that cannot move: "The IPv4 address space is a 32 bit field.
There are 4,294,967,296 unique values, considered in this context as a sequence of 256 "/8s", where each "/8" corresponds to 16,777,216 unique address values." It prints "There are a total of the equivalent of 35.078 /8 address blocks that are 'reserved'." It prints "The remaining 220.922 /8 address blocks are available for use in the public IPv4 Internet." And it prints series that do move, with their own method sentences attached: "The data used to construct the time series of allocations is the allocation "stats file" published on a daily basis by each RIR, placed into a time series"; "This approach treats an address as "consumed" once it has been allocated or assigned by an RIR"; and, for routing, "The next data set is total span of address space advertised in the BGP routing table over time.
The data has been collected since late 1999."
So one clock — the generation stamp — covers two species of fact. It dates a page that contains both arithmetic that is fixed by design and statistics that move with the registries' daily files. The stamp is accurate about the page and silent about the series. That is not a defect in the page. It is a trap for the citation, and the trap has a name: a reader who writes "as of 21 November 2026" beside an allocation total has turned a page-assembly label into a measurement date that the page never supplied.
The sentence that carries its own year
The January 2026 column does the opposite, and does it well. The figure and its boundary arrive in one breath: "The total IPv4 allocated address pool contracted by some 237 thousand addresses in 2025, with some 3.687 billion allocated addresses at the end of the year." The next sentence supplies the rate against the same boundary: "This represented a contraction of some 0.01% for the total allocated IPv4 public address pool through 2025." Copy either sentence whole and you cannot lose the year, because the year is inside it (see the January 2026 column).
The column also prints its method next to its number rather than somewhere else. It states the metric choice — "perhaps the best metric to use is the total span of allocated and assigned addresses" — and it explains why a transfer looks like an allocation in the registry record: "These address transfers necessarily entail a change of registration details, so the registry records the outcome of a transfer, or sale, in a manner that is similar to an allocation or assignment." Then the author makes the choice explicit rather than leaving it to the reader: "I personally find the distinction between these two terms somewhat of a distracting artifice these days, so from here on I'll use the term "allocation" to describe both allocations and assignments, without further distinction." The HTML rendering of the same column sharpens the definition further — "In this study, an allocation is defined here as a state transition in the registry records from reserved or available to an allocated state" — and carries the IPv6 warning that a naive count of allocated IPv6 space misleads, because an ISP using /48 end sites needs 256 times more space than one using /56 and 4,096 times more than one using /60 (see the HTML column).
One numeral inside the same document does not match the prose exactly. The rendered table prints an annual change of "-0.2M" for 2025 where the prose says "some 237 thousand", and it prints 3.687 billion for both 2024 and 2025 while the change row moves. Two hundred and thirty-seven thousand rounds to 0.2 million, so the first difference is a rounding convention, not a contradiction. I do not compute a third number between them, and I do not build an argument on a table whose cells arrived second-hand.
The republication test
Then the sentence travelled. ARIN republished the column on 22 January 2026 under the byline "By Geoff Huston – Chief Scientist, APNIC" and the line "Republished from Geoff Huston's blog, potaroo.net." The end-of-2025 sentence arrived word for word, boundary and all: "The total IPv4 allocated address pool contracted by some 237 thousand addresses in 2025, with some 3.687 billion allocated addresses at the end of the year. This represented a contraction of some 0.01% for the total allocated IPv4 public address pool through 2025." The IPv6 granularity sentence arrived too. Whatever changed in the move from a personal site to a registry's blog, the qualifier was not among the casualties (see the ARIN republication).
What did not travel was a label. In the text of that republication read on 21 November 2026, the IPv4 section is headed "IPv4 in 2024", while the figures beneath it are the end-of-2025 figures and the page itself is dated 22 January 2026. One heading, one year, and it is the wrong one relative to its own numbers. I can document what the page printed; I cannot say from the page alone whether the heading was carried over from an earlier edition, produced by a template, or lost in editing, and nothing here turns that into a claim about anyone's care.
The asymmetry is the story. The sentence that carries the number carried its year across a change of host, because the year was welded into the sentence. The heading that carries no number kept a year that no longer fits, because a heading has nothing to reconcile against except the reader's patience.
Where a stranger's reading has to stop
Everything quoted above is text that the pages themselves returned on 21 November 2026. No page was opened in a browser session for this piece, the figures and charts in the rolling report did not render as readable numbers, the rolling report's text ended part way through the document, the plain-text column stopped before its IPv6 section, and the republication's text was truncated before its later tables. The consequence is specific: I can tell you what each document prints in prose beside its figures, and I cannot tell you what its charts plot.
My own arithmetic is confined to one subtraction of dates on a stamp and one comparison of two rounding conventions. I have not re-derived the 0.01%, not annualised anything, and not converted /8 equivalents into addresses without the page's own equivalence.
The lesson is not that these pages are careless. It is that a qualifier is not a property of a corpus; it is a property of the particular sentence someone chose to copy. Huston wrote 3.687 billion so that its year could not be lost. The page that regenerates every morning was never going to make that promise, and the heading above the republished paragraph shows what happens to a label that nobody had to keep.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






