UK exits for British corpus, Cambridge-Oxford research, eval

Why a British English corpus is its own collection problem

British and American English diverge in spelling, idiom, domain vocabulary (legal, medical, academic), and content hosting patterns. Training data pulled from a US-routing gateway under-samples .ac.uk academic content, UK-centric knowledge bases (gov.uk, NHS.uk, BBC News/iPlayer metadata, Hansard), and the British press tier that geoblocks US-origin cloud traffic. The SquadProxy UK pool exists to collect those surfaces cleanly and to run eval workloads from a credible UK-resident vantage point.

What the UK pool routes through

• Datacenter — AWS eu-west-2 (London), GCP europe-west2, Azure UK South. Primary path for bulk corpus work against open or tolerant targets.
• Residential — BT, Sky, Virgin Media, TalkTalk exits with postcode-level selection in the top 12 conurbations. Used when the target geoblocks cloud ASNs or when eval workloads need a credible home-broadband IP.
• Mobile — EE, Vodafone, Three, O2 4G/5G. Narrow but real; useful for cellular-anchored evaluation of services whose mobile rendering differs from desktop.

Research cluster proximity

Cambridge (home to Microsoft Research Cambridge, the Alan Turing Institute, DeepMind's Cambridge presence) and Oxford are both within ~60ms RTT of our London edge. That matters if you are running evaluation workloads from a university affiliate and need low-latency paths to external model APIs, or if you are scraping academic repos like the arXiv mirrors at Cambridge.

Good fit

• Training-corpus collection for British English: gov.uk, Hansard, BBC News, Guardian, FT (public tier), UK academia
• Regional evaluation against model APIs from a UK-resident IP
• RAG ingestion of UK-specific knowledge bases and government data
• Benchmark scraping of UK-hosted papers, datasets, and leaderboards

Not a good fit

• Circumvention of paywalled UK press — our AUP prohibits it
• NHS.uk content involving patient identifiers under any circumstance
• Any scraping that would constitute unauthorised access under the Computer Misuse Act 1990

ASN depth in the UK residential pool

UK residential broadband concentrates across a short list of carriers and the pool is sized against those:

• BT (AS2856) — largest UK residential footprint, including Openreach-delivered services on top of FTTC and FTTP.
• Virgin Media O2 (AS5089) — cable footprint, strong in urban areas, generally higher downstream bandwidth than BT on equivalent tiers.
• Sky / NOW (AS5607) — significant subscriber share in urban England and Scotland.
• TalkTalk (AS13285) — secondary residential, budget tier.
• Vodafone UK (AS25135) — fiber and FWA, growing share post-2023.

On the mobile side, EE (BT-owned), Vodafone, O2, and Three carry essentially all UK 4G/5G. The 4G/5G pool samples across all four.

Top UK metros

• London — dominant. Financial services data, dense CDN infrastructure, AWS eu-west-2 home region. Most UK-anchored AI eval runs use London as the default origin.
• Manchester — Northern England anchor, MediaCityUK creative industry cluster. Useful for non-London UK regional diversity.
• Edinburgh — Scottish regional content, financial services (RBS/NatWest historic base), and academic (University of Edinburgh) research-adjacent surfaces.
• Birmingham, Bristol, Leeds, Glasgow — secondary metros with useful regional news and content footprints.

UK legal landscape for AI data collection

• Computer Misuse Act 1990 — UK equivalent of CFAA. Unauthorised access (behind authentication or technical controls) is criminal; scraping publicly-accessible content is not.
• UK GDPR + Data Protection Act 2018 — personal data handling obligations equivalent in substance to EU GDPR. The ICO is the supervisory authority; its published guidance on AI and automated decision-making should inform training-data practices.
• TDM (Text and Data Mining) regime — UK's CDPA s29A permits TDM for non-commercial research; the proposed extension to commercial research was dropped in 2023-2024. UK commercial training on copyrighted content remains legally uncertain and has driven licensing deals rather than permissive-default use.
• AI Safety Institute (AISI) remit — AI safety testing in the UK has an explicit regulatory footprint via AISI. Safety red-team workloads coordinated with AISI have specific considerations; see the safety red-team use case.

Latency and edge routing

UK edge routes through AWS eu-west-2 (London). Latency from edge to major AI inference providers:

• OpenAI API: ~15-30ms p50
• Anthropic API: ~20-40ms p50
• Google Gemini (europe-west): ~25-45ms p50

Residential adds 40-150ms; mobile adds 70-200ms. BT and Virgin Media footprints are the most consistent for latency-sensitive workloads.

Related

• UK residential proxies for AI training
• UK ISP proxies
• UK 4G mobile proxies
• Residential vs datacenter routing