__MK__
tabaqat.
Telecom
Underserved Gap & Take Rate

Two maps both call themselves
“the broadband gap”.

One counts the places where service cannot be bought. That is the map public money follows. The other counts the places where it can be bought and nobody does. That is the map a build’s payback follows. In California they differ by 20×.

The Gap Matrix Live CPUC data 293 assertions · 146 against live services
Built on strata-core · open, MIT, runs on Strata or ArcGIS Verified 2026-08-09
__MK__
tabaqat.
The question
“We can get funded to build in the unserved blocks. But the households are in the served blocks where nobody subscribes. Which gap are we solving, and how many customers does each one actually contain?”
$42.45B
BEAD, allocated on unserved and underserved location counts
46.5%
US average fiber take rate on primary passings
23M
households on the ACP when it ended, 1 June 2024

Every party in this market has been handed a map of availability and asked to make a decision about revenue. The industry says so itself: “a home passed is a unit of infrastructure — it is not a subscriber, not a satisfied customer and not a return on invested capital.”

Sources: NTIA/CRS · Fiber Broadband Association 2025 · Pew 1 / question
__MK__
tabaqat.
The finding

California, EOY 2024. Same blocks, two classifications.

Both layers published by the CPUC. Joined block-by-block on GEOID20, 655,000 rows, fully paged.

Can’t buy — unserved blocks
142,732
Won’t buy — served, ≤60 % take it
2,904,667

The revenue gap is 20.4× the funded one, in households — and it is somewhere else on the map.

Fresno County alone: 3,818 households can’t buy; 64,137 can buy and don’t.

CPUC EOY 2024 · availability ⟕ adoption, outer join on GEOID20 verified live
__MK__
tabaqat.
The row nobody prints

Of California’s 34,046 unserved blocks, 29,641 have no take-rate measurement at all.

87.1%
of the funded gap is absent from the adoption layer
7.7×
how far wrong an inner join reports the funded gap — silently

The take-rate map is blind precisely where the money goes. Most tools would inner-join the two layers, drop those blocks and report a tidy number. We keep them as a fourth row of the matrix and colour it amber: an absence, not a zero.

Fresno reads 86 % — so this is systemic, not a local quirk 29,641 of 34,046
__MK__
tabaqat.
The instrument

If two institutions disagree about a classification,
put both on the axes.

Availability down, take-rate band across, cell area ∝ households. Click any cell and the whole application adopts that population — map, reading, provider list, block table, URL. Fresno County, EOY 2024, at 100/20 Mbps:

0 / ? ⚠≤20%≤40%≤60%≤80%>80%
Served
2,508
11,742
16,316
36,079
82,105
146,040
Unserved
119
298
103
108
·
17
Priority Unserved
3,038
50
85
·
·
·
No take-rate record ⚠
5,063 blocks — no take-rate published at all

The widest thing on the screen is a measurement nobody took.

Households per cell · hover to preview · click to adopt Fresno · EOY 2024 · 100/20
__MK__
tabaqat.
Credibility

The credibility is in the refusals.

We never state a take rate

CPUC publishes it in six 20-point bands. The app says “≤ 60 %”. It will not say “47 %”.

Band 0 means “0 % or no data”

One code, two facts. Amber, excluded from headlines by default, and a toggle that shows you exactly what including it costs: 16.8× → 17.5×.

Advertised ≠ measured

Availability is provider-reported. Ookla’s open tiles could flag a disagreement; they are a sample of speed tests, not a census of subscribers, so they never move a number.

No build-cost model

No public cost-per-passing exists for California. A dial with an invented default would make the map a function of a number we made up.

Every figure on screen carries a printed derivation — service, field name, vintage, join type, caveat. Seven lines, on the page, not in a log.

Honest scope is published in the app’s own splash and footer 2 / instrument
__MK__
tabaqat.
Field notes

Five things we did not expect to find.

FindingWhat it means
California publishes no broadband data All 2,569 endpoints in the state’s known ArcGIS estate: zero availability layers, zero adoption, zero provider registers. Maryland publishes three. Virginia publishes two.
An eighth server, found in a Leaflet config The state broadband map is Leaflet — but it loads esri-leaflet, and its config.js points at a CPUC ArcGIS server nobody had crawled. 36 services, CORS-open, adoption and availability at two vintages.
The federal source is unreachable broadbandmap.fcc.gov returns 403 to any non-browser client; its data API returns 401. The only reachable FCC series is Form 477, June 2021 — the dataset the BDC replaced because it over-reported.
A point layer that returns no points 215,804 broadband-serviceable locations — the unit BEAD money is counted in — are countable but not mappable. Verified across every parameter combination.
Two services, one dataset “25/3 Mbps” and “any available speed” are band-for-band identical. Publishing both implies two measurements; there is one.
17 traps documented, each reproduced rather than inferred G1 – G17
__MK__
tabaqat.
Why us

Every incumbent in this market sells one of the two maps.

Plant record

VETRO, IQGeo, 3-GIS, Render, Comsof. They answer where is my cable — not who would buy from it.

The location file

CostQuest is the FCC’s Fabric contractor. It cannot be the tool that questions the availability map.

Measurement

Ookla, M-Lab. They sell the meter reading.

The platform

Esri’s BEAD dashboard and Telecom Management — an enterprise agreement priced by organization size.

We have no map to defend, so we can hold both up at once — and show you the part of the first one that nobody has measured.

Coexistence, not replacement — runs on Strata or ArcGIS MIT · sovereign · on-prem
__MK__
tabaqat.
How it was built

Described, not coded.

A GIS admin describes the application; the authoring commands write the map spec, the symbology, the popups and the panels. The research, the data catalogue and this application were produced in a single working session.

293
assertions, green
146
against the live services — no fixtures
147
offline, driving the shipped page
6
bugs the harness caught before release

The trap assertions are inverted on purpose: they pass while the trap is still there. If CPUC ever fixes one, the suite fails and tells us the research is stale. That is the difference between a demo and an instrument.

strata-core · React + MapLibre · ESRI Web Map JSON in and out one session
__MK__
tabaqat.
Next
The offer

Bring us one county and the
gap number you are about to fund.

We will show you the same ground under the other definition, how many households sit in each cell, and how much of your funded gap has never been measured. If the two agree, you have lost an afternoon. They have not agreed yet.

Deploys on-prem — your subscriber book never leaves Strata or ArcGIS
tabaqat · Underserved Gap & Take Rate CPUC EOY 2024 · verified 2026-08-09
1 / 10