Quernal Labs · Reports — The UK AI Infrastructure Map

PILOT — internal build, not published, editor gate pending. Not a subscriber edition; not free. A dated snapshot of the dataset as of 5 July 2026, not a live tracker. Where a commitment is recorded at programme level and not yet firmed against a primary source, it is marked in crimson.

Drill it Anything with a dotted underline opens a panel and drills to source. Follow the trail down, climb back up the breadcrumb. No dead clicks.

Section 1 — Executive read

Britain is building the ground. It does not own the stack.

The UK has committed tens of billions to AI capacity on home soil. The compute that fills it runs through a handful of foreign chokepoints — several of them single-source.

65 / 15

Ecosystem entities mapped, across 15 value-chain layers — labs to lithography to power

Chokepoints in the dependency graph — points where the whole chain narrows to one or few suppliers

£26.5bn+

GBP-denominated UK AI-infrastructure commitments, plus Microsoft's $30bn (2025–28). Mixed currency, as announced — not summed

The read

Read from the bottom up, the UK's position is strong and getting stronger: real facilities, real power connections, a national supercomputing spine, and a deep queue of committed capital. Read from the top down, it is exposed. The pieces that decide whether any of this compute can actually run — the lithography that makes the chips, the packaging that assembles them, the memory that feeds them — sit entirely offshore, and at three points narrow to a single company with no substitute.

This is the report's one structural conclusion: the UK owns the sites and the sockets, not the silicon.

Section 2 — The chokepoints

Nine places the chain narrows to one

The signature analysis. Nine dependencies where supply concentrates to a single or dominant provider. The substitutability column is The Quernal's classification of the sourced facts — not a proprietary score. Click any dependency to drill it.

Dependency	What flows	Layer	Substitutability	Source
ASML to Intel	EUV lithography systems	semicap to chipmaker	Single-source	Gotrade
ASML to Samsung Electronics	EUV lithography systems	semicap to memory	Single-source	Gotrade
ASML to TSMC	EUV lithography systems	semicap to foundry	Single-source	TechTimes
Bureau of Industry and Security to Nvidia	Export controls govern China-bound AI chips	gov / regulator to chipmaker	Policy lever	source pending
Nvidia to CoreWeave	AI accelerators (GPUs)	chipmaker to hyperscaler	Single-supplier dep.	CNBC
Nvidia to OpenAI	GPUs and up to $100bn partnership	chipmaker to lab	Concentration / interlock	source pending
Nvidia to xAI	Colossus GPUs (555k+ and growing)	chipmaker to lab	Concentration	source pending
SK Hynix to Nvidia	High-bandwidth memory (HBM)	memory to chipmaker	Oligopoly · sold out	Introl
TSMC to Nvidia	Leading-edge wafer fabrication and CoWoS advanced packaging	foundry to chipmaker	Single-source	INDmoney

How to read this

Three of the nine are single-source. ASML is the sole commercial supplier of EUV lithography — no substitute below the 7nm node — and its order book leans on just three buyers. TSMC's CoWoS packaging is the single tightest part of the whole hardware stack, sold out through 2026.

The memory node is an oligopoly running flat out. SK Hynix provides roughly 90% of the HBM Nvidia buys; HBM is the binding constraint on training and inference, sold out across all suppliers through 2026.

The Nvidia-downstream links are a different risk — not scarcity but concentration. Three of the nine carry source pending, shown not hidden: a paid report firms these against a primary before they carry weight.

Section 3 — The stack

The chain, layer by layer

The 65 mapped entities, upstream (the tools that make chips) to downstream (the labs that consume them). Click a layer, or any company, to drill in.

Semicap4ASML · Applied Materials · KLA · Lam Research — the machines that make chips

EDA2Cadence · Synopsys — the chip-design software duopoly

IP1Arm — the instruction-set architecture under almost every AI SoC

Foundry1TSMC — leading-edge fabrication and packaging

Chipmaker4AMD · Intel · Nvidia · Syntiant — the accelerator designers

Memory4Micron · SK Hynix · Samsung Electronics · Sandisk — HBM and storage

Networking3Arista Networks · Broadcom · Marvell Technology — the fabric between GPUs

Infra hardware5Dell Technologies · Eaton · Schneider Electric · Vertiv · Western Digital — power, cooling and kit

DC operator5DataVita · Digital Realty · Equinix · Pure Data Centres · SEGRO — the buildings

Hyperscaler8Alphabet · Amazon · CoreWeave · Meta · Microsoft · Nebius · Nscale · Oracle — the clouds

Power8Bloom Energy · Constellation Energy · EDF · GE Vernova · National Grid · SSE · Talen Energy · Vistra — the electricity

Capital3Blackstone · SoftBank · SpaceX — the balance sheets funding the build

Lab8Anthropic · DeepSeek · Google DeepMind · MiniMax · Mistral AI · OpenAI · Zhipu · xAI — the model-builders

Gov / regulator4Bureau of Industry and Security · DSIT · Ofgem · US Commerce Department — the state levers on both sides

Other5Apple · Palantir Technologies · Salesforce · ServiceNow · Tesla — adjacent AI-exposed majors

Where it thins

A bulge in the middle and a pinch at the top: many clouds, many power players, many labs — but one foundry, one IP licensor, two EDA houses, and a semicap layer that funnels into a single EUV supplier. Competition at the ends; single points of failure in the core.

Section 4 — Chokepoint dossiers

The four that decide it

The sourced detail on the load-bearing four, each with The Quernal's UK read.

ASML to TSMC · semicap to foundry

EUV lithography

ASML is the sole commercial supplier of EUV lithography machines; there is no substitute below the 7nm node. TSMC is the largest EUV customer, estimated at c.40% of deliveries. The single most critical chokepoint in the chip supply chain.

UK readNo British-made AI accelerator, and none the UK buys, exists without one Dutch company's machines. There is no domestic lever here at all — only the diplomacy of the export-control regime.

Single-sourceSource: TechTimes

TSMC to Nvidia · foundry to chipmaker

Wafers + CoWoS packaging

TSMC fabricates Nvidia AI GPUs on leading-edge nodes and packages them via CoWoS. Nvidia is the CoWoS anchor tenant, booking c.60% of 2026 capacity. CoWoS is the single tightest part of the AI hardware stack, sold out through 2026 (CEO C.C. Wei, 4 Jun 2026).

UK readEven with GPUs allocated, packaging capacity gates how many can actually ship. When UK operators say sites are ready, the real queue is often here — in Taiwan, not on the grid.

Single-sourceSource: INDmoney

SK Hynix to Nvidia · memory to chipmaker

High-bandwidth memory

SK Hynix is Nvidia primary HBM supplier and the HBM market leader (c.50-62% share). Nvidia accounts for roughly 90% of SK Hynix HBM supply. HBM is the binding constraint on AI training/inference and is sold out across all suppliers through 2026.

UK readThe memory pinch is the one most likely to cap 2026 delivery. All three HBM makers are now also equity partners of the labs — an interlock that concentrates the market further, just as demand peaks.

Oligopoly · sold outSource: Introl

Nvidia to CoreWeave · chipmaker to hyperscaler

GPU-anchored cloud

CoreWeave is an Nvidia-anchored AI cloud whose business is built almost entirely on Nvidia GPUs; a near-total single-supplier dependency.

UK readCoreWeave is among the operators committing UK capacity. A cloud that is single-supplier at its core inherits that supplier's scarcity — a UK customer's capacity is only as available as Nvidia allocation allows.

Single-supplier dep.Source: CNBC

Section 5 — The UK ground

Where it lands on British soil

Nine mapped UK sites — from Europe's largest operational cluster to the largest permitted campus in the country. The layer Britain does own. Click a site to drill it.

Facility	Location	Capacity	Status	Note	Source
Slough cluster	Berkshire	c.1,000 MW (c.675 MW hyperscale)	Operational	Europe's largest datacentre cluster and the UK de facto AI Growth Zone: 30+ operational datacentres totalling c.1GW, c.14,000 jobs, £30m+ annual business rates. Equinix operates the flagship LD4/LD5/LD6/LD7 campus.	Data Centre & Network News (Kao Data)
Elsham Wolds	North Lincolnshire	up to 1,000 MW	Permitted	Planning approved March 2026 for what is billed as the UK's largest AI datacentre campus: a 176-hectare former RAF airfield. Phase 1 construction from 2027, opening targeted 2029.	Royal Geographical Society (GOV.UK / PBC Today)
Culham AI Growth Zone	Oxfordshire (UKAEA)	100 to 500 MW	Planned	The first UK AI Growth Zone, confirmed 13 Jan 2025 under the AI Opportunities Action Plan, at the UK Atomic Energy Authority campus. A 400kV connection freed by the JET fusion decommissioning; AWS and CloudHQ reported developing nearby.	Oxford Calling
North East AI Growth Zone	Cobalt Park & Blyth	at least 500 MW required	Under constr.	The second UK AI Growth Zone, across two North East sites near Newcastle and in Northumberland. DSIT taskforce launched 25 Sep 2025. Cobalt Park already hosts three colocation datacentres on dedicated National Grid supply.	Computer Weekly
Nscale Campus, Loughton	Essex	50 to 90 MW	Under constr.	The UK's largest AI supercomputer when live (2027): built with Microsoft, housing c.23,040 NVIDIA GB300 GPUs, liquid-cooled. Part of Microsoft's $30bn UK AI push announced 16 Sep 2025.	Nscale
Pure / SEGRO, Park Royal	West London	56 MW	Under constr.	A £1bn, 56MW datacentre — a joint venture between SEGRO and Pure Data Centres formed March 2025, part of the shift of AI capacity into London-edge sites.	W.Media
Google Waltham Cross	Hertfordshire	—	Operational	Google's hyperscale datacentre, opened September 2025 — the anchor of a £5bn two-year UK investment spanning cloud, DeepMind research and infrastructure engineering.	W.Media
Dawn supercomputer	Cambridge	c.19 petaflops	Operational	The UK's fastest AI supercomputer (Phase 1), operational since November 2023 at the University of Cambridge's Open Zettascale Lab — 1,024 Intel Data Center GPU Max accelerators, built with Intel and Dell. Anchors the national AI Research Resource (AIRR).	University of Cambridge
DataVita	Scotland	—	Operational	A Scottish datacentre operator whose capacity underpins CoreWeave's £1.5bn commitment to UK AI compute — part of Scotland's growing role as a focal point for next-generation AI infrastructure.	Data Centre Insight

The ground read

The strongest column in the report: c.1 GW operational at Slough alone, c.1 GW permitted at Elsham Wolds, two Growth Zones scaling toward 500 MW each, a national supercomputing spine, capacity from the London edge to Scotland. Real, dated, mostly operational or under construction.

But every site needs the GPUs, packaging and memory from Section 2. The ground is sovereign. What runs on it is not.

Section 6 — The money

The commitment ledger

The headline UK AI-infrastructure commitments, in their announced currency — deliberately not summed across currencies. Click a committer, amount or provenance to drill.

Committer	Commitment	Announced	Detail	Provenance
Microsoft	$30bn (2025–28)	16 Sep 2025	Largest-ever UK commitment; c.$15bn capex incl. the UK's largest supercomputer with Nscale (Loughton). UK–US Technology Prosperity Deal.	source pending
Blackstone	£10bn	—	AI campus at Cambois / Blyth, Northumberland; within the North East AI Growth Zone.	programme-level
Amazon (AWS)	£8bn (2024–28)	—	UK datacentre fleet expansion; a steady programme rather than a single flagship site.	programme-level
Alphabet / Google	£5bn (2 yr)	—	Cloud, DeepMind research and datacentre capacity; anchored by Waltham Cross (opened Sep 2025).	date not vs primary
Nvidia	£2bn + c.60k GPUs	16 Sep 2025	To catalyse UK AI startups; the largest single Nvidia deployment in the country. Tech Prosperity Deal.	source pending
CoreWeave	£1.5bn	—	UK AI compute anchored on DataVita capacity in Scotland.	source pending

The money read

£26.5bn in GBP-denominated commitments, plus Microsoft's $30bn (2025–28). Shown separately and not converted, because an FX-blended total would be a manufactured figure — and this report doesn't manufacture figures.

Two of the six are pinned to a hard date; the other four are programme-level, not yet firmed to a primary. Here they carry their provenance state on their face. Firming them is the first item any paying version clears.

Section 7 — The Quernal read

What the map means

Britain has bought the ground and is wiring the sockets. On this dataset the UK's physical AI build-out is genuine, large and accelerating. On the layer the country can own — land, grid, planning, sites — the position is strong.

The exposure sits one layer up, and it is concentrated. Every site depends on hardware whose critical nodes are entirely offshore and, at three points, single-source: ASML's EUV lithography, TSMC's CoWoS packaging, SK Hynix's HBM. None has a domestic substitute; two are sold out through 2026. The UK's lever is not industrial but diplomatic — the export-control regime (BIS, Commerce).

Why this sits inside The Quernal's question. The infrastructure on this map is the physical answer to what happens when we don't have to work — the compute that automates the work is built on it. Who controls that compute, and where it pinches, is who sets the pace and terms of that change. Offered as analysis, not a settled house verdict; the sourcing above is where a reader checks our working.

Method & provenance

How this report is built — the glass box

Every figure is drawn from The Quernal's AI-ecosystem dataset (65 entities across 15 layers; the dependency graph and its 9 chokepoints; 9 UK facilities; the commitment ledger). Each carries the real date it was recorded and a separate date for the fact it describes. There is no proprietary composite: the substitutability classifications are The Quernal's labelled reading of the sourced facts beside them — the inversion of the market convention of all-scores-are-composites-methodology-in-appendix.

Where a commitment is recorded at programme level and not yet firmed against a primary source, it is marked in crimson. This is a dated snapshot as of 5 July 2026, not a live tracker. No scraped sources; enrichment only from legitimately public and licensed material. No synthetic inputs — no modelled, composite or simulated attitudes are used, quoted, or counted anywhere.

Conflict of interest. Anthropic — whose models write The Quernal — appears on this map as a model lab and in the surrounding capital flows. This is the deepest conflict on the map and is disclosed wherever the AI-lab stack is named. (Drill any COI-flagged company for the specific disclosure.)