PORTFOLIO · GLOBAL REACH & LATENCY · FIU-NET-001
UK compute, one hop from every market.
Modelled fibre round-trip times from the portfolio's three geographies — the Scottish cluster, the
Thames estuary and Pembroke — to the world's main interconnection markets, computed live in this page
from the same coordinates and methodology as the per-site Connectivity tabs (great-circle distance ×
1.35 route factor × 4.9 µs/km in glass, round trip). Production routes typically add 10–30%.
Training is latency-blind — the matrix below is not about where training runs; it is about
where inference and replication can live.
The Latency Matrix
Modelled fibre RTT from each portfolio geography to thirteen interconnection markets — every figure in this table is computed in-page from coordinates; nothing is hand-entered. Bands: green <5 ms (synchronous), teal 5–20 ms (interactive inference), amber 20–80 ms (async/DR), dim >80 ms (placement-irrelevant).
<5 ms — synchronous
5–20 ms — interactive inference
20–80 ms — async / DR
>80 ms — placement-irrelevant
Named-cable cross-check: New York over the named Yellow system (7,001 km, Bude → Bellport): ~74–78 ms — the named-cable calculation agrees with the model.
What Each Band Supports
Latency budgets map to workload classes, not to marketing tiers. The commercial reading: the portfolio's serving reach covers all of North-West Europe from every campus, and the workload its BTM power story targets — frontier training — never appears in this table at all.
<1–5 ms
Synchronous territory
Synchronous replication, active-active database pairs, real-time voice and interactive inference where round-trip is user-perceptible. Thames → London Docklands and Pembroke → Dublin both sit inside this band — metro-class positioning without metro land prices.
5–20 ms
GenAI inference at scale
Token streaming dominates the user experience, not the network — anything under ~20 ms is invisible inside a generation stream. Every site serves all of North-West Europe in this band, including the Scottish cluster: London, Dublin, Amsterdam and Paris are all in reach from every geography.
20–80 ms
Async · DR · distribution
Asynchronous replication, disaster recovery, checkpoint and weight distribution between campuses. This is the band of the Nordic training pairing — Narvik is ~22 ms (modelled) from the Scottish cluster: hydro-powered Nordic campuses train, UK sites serve from closer in.
>80 ms
Placement-irrelevant
At transcontinental distances placement is decided by power, land and jurisdiction — not the network. Training is latency-blind, and training is exactly the workload this portfolio's behind-the-meter power story targets: the >80 ms column costs the book nothing.
The Fleet Behind the Reach
Coupar Angus Combined-Domain Concept
Stirling + Keithick as one training domain
ENGINEERING CONCEPT — NOT A CURRENT DESIGN COMMITMENT
Stirling and Keithick sit a few hundred metres apart on one estate at Coupar Angus. Running the
scale-out fabric over short private dark fibre between adjacent buildings is standard
multi-building campus practice — at these distances the added latency is microsecond-scale,
inside a single fabric's tolerance. The pair could therefore present to a tenant as a
~52 MW combined training domain (40 + 12.5 MW IT) rather than two isolated small sites.
What it needs: a wayleave/duct between the plots and a fabric design that spans the two halls.
Until both exist, the sites are marketed as they are consented — separately.
40 + 12.5
Stirling + Keithick MW IT
µs-scale
Inter-building fabric latency
Wayleave + fabric
Required to realise
Named-System Reach
The reach story rests on named, dated cable systems — TeleGeography's published dataset, graded in the evidence base — not on adjacency hand-waving.
West · Bude Gateway
Transatlantic + Europe–Asia from UK soil
Yellow → New York (7,001 km, Bude → Bellport) · Amitie → Boston (Lynn MA, 2023) ·
EIG → Mumbai via the Med and Gulf (15,000 km system) — plus the 2Africa and Grace Hopper
systems per the dataset sweep. One Cornish gateway carries named routes to North America and Asia.
North · Nordic Corridor
The UK end of the corridor Microsoft is filling
Tampnet North (Aberdeen ↔ Norway) · NO-UK (Newcastle ↔ Stavanger, 2021) ·
Havhingsten (Newcastle ↔ Denmark, 2022) · FARICE-1 → Iceland. Far end:
Narvik — Nscale's 230 MW campus, Microsoft's 30,000+ Rubin GPUs (per the evidence file) —
the training-vs-serving pairing in physical form.
East · Kent / Channel Corridor
Continental Europe, still attracting new builds
IOEMA-1 (RFS 2029) — five-country North Sea ring landing at Dumpton Gap, ~40 km from
Littlebrook · the Channel trio south of the Thames campuses, including Q&E South
(RFS 2026), CrossChannel Fibre and Circe South. The corridor is gaining systems, not losing them.
Status & provenance. Every latency figure on this page is modelled, not measured:
great-circle distance × 1.35 route factor at 4.9 µs/km in glass, round trip — identical to the
methodology on the per-site Connectivity tabs. Production routes typically add 10–30%.
City and cable coordinates follow TeleGeography's published dataset where graded
VERIFIED-SOURCE in the evidence base; all other geometry is indicative. Fleet figures
(IT MW, GPU counts) come from each site's design data model in the portfolio registry. Last-mile
carrier confirmation is in flight — budgetary enquiries issued; see the per-site Connectivity tabs.
Nothing on this page is a committed network design.