Galiark
Introducing Galiark ↓

Galiark.

Sovereign cognitive intelligence.
We build the cognitive intelligence layer underneath sovereign defence and regulated commercial AI. Two products, one engine.
Every regulated organisation owns data they cannot use. We are the keys.

Two founders. Thirty years combined in adversarial AI. One missing layer.

Galiark builds two products on one cognitive engine, designed to sit underneath the AI workloads where adversaries adapt and sovereignty is non-negotiable.

The takeaway ↓
Two founders. One cognitive intelligence layer. Two products on one engine, in two adversarial markets.
Where it started ↓

It started with a chance conversation.

An architect and a commercial operator at the seam between hard technology and high-stakes deployment.

01. The two of us, before this

Tom Bennett.

39 years across global financial markets, predictive AI, and proprietary AI software development. Architect of the Bankers Trust risk-system spin-out and the largest global trading-system rollout of its era. Built his own deep-learning variant after a decade decoding W.D. Gann.

Matt Tyler.

25 years across deep-technology commercialisation. Currently CEO of Herbi4, transitioning operationally to Galiark. Co-founder of Tech8 Digital. Founding Partner of Parzival Partners since 2018.

02. The chance conversation

A senior figure at a NASDAQ-listed semiconductor company we cannot name in this document hears one of us describe the cognitive architecture being quietly built. Approaches us privately. Says: we have a drone communications problem we cannot solve. Can your architecture solve it?

03. The realisation

We started building to solve a drone problem. Within months, we realised we had built something larger.

We had not built a drone communications layer. We had built the intelligence layer that is missing from every AI system deployed in the world today.

We named the original defence application TAGES, after the proving-ground deployment. We have since renamed the underlying cognitive engine Subcognit. Galiark is the company. Subcognit is the technology we build.

The takeaway ↓
We started solving a drone problem. We built the intelligence layer that is missing from every AI system in the world.
The problem ↓

Pattern matchers fail by design.

A thermal decoy that emits the right thermal signature.

A fraud ring that varies its transaction footprint.

A novel cyberattack that does not match any known signature.

Every AI system deployed today, Anduril Lattice, Palantir Maven, Shield AI Hivemind, Helsing Altra, every fraud system, every credit model, every diagnostic, is correlational. It learns what past examples looked like and matches new inputs against them. Adversaries change patterns. That is the definition of adversarial pressure. Pattern matchers fail by design under exactly the conditions where AI matters most.

And one structural constraint sits underneath all of this.
The most sensitive workloads, defence, healthcare, finance, law enforcement, cannot ship their data to a hyperscaler to be reasoned over. Sovereignty is a hard constraint, not a preference.
The takeaway ↓
Pattern matchers fail under adversarial pressure. Sovereignty is a hard constraint, not a preference.
Our products ↓

One engine. Two products. Two adversarial markets.

Galiark Sovereign and Galiark Commercial. Subcognit underneath both.

Galiark Sovereign
Sovereign cognitive intelligence for defence, national security, and government crime.
What we sellA deployable cognitive runtime, plus ingestion adapters, plus an operator console, plus integration services, plus annual support. Subcognit inside.
Who buysUK MoD AI procurement. DSIT Sovereign AI Unit. NATO ACT and allied national security agencies. NCA, SFO, HMRC, allied law enforcement. Channelled through BAE, Thales, Babcock, Five Eyes integrators where required.
Where it runsOn the customer's compute. Jetson at the edge. On-prem GPU in secure facilities. Sovereign cloud (UKCloud, sovereign Azure region, allied equivalents). Customer holds the keys.
BeachheadUK Sovereign AI Unit £80M competition opens July 2026.
Sales motionFramework agreement, pilot, scale rollout. 12 to 18 month procurement cycle.
Pricing directionPer-platform or per-node deployment licence, plus integration services, plus annual support. Tiered by deployment scale.
Galiark Commercial
Causally explainable, FHE-protected decision intelligence for regulated commercial buyers.
What we sellAn enterprise software runtime that sits between the customer's data plane and decision systems. Subcognit inside.
Who buysChief AI Officer, Chief Risk Officer, Head of Financial Crime at tier-1 banks first. Then large insurers. Then exchanges and fintechs.
Where it runsOn-prem or in the customer's own VPC in their chosen cloud. Customer holds the keys. FHE substrate ensures cryptographic sovereignty.
Lead verticalsFinance: fraud, AML, credit risk. Insurance: claims fraud, underwriting, novel risk. Financial Crime: AML at platform scale, sanctions evasion, beneficial ownership reasoning.
Beachhead and forcing functionEU AI Act August 2026 makes correlational AI legally non-deployable for high-risk classes. Galiark Commercial is the route to compliance.
Pricing directionEnterprise tenant licence, plus per-decision metering above contracted volume floor.
Sales-motion edge
FHE substrate collapses procurement cycles.
The 12 to 18 month sales cycle that kills most defence-tech and regulated-AI seeds is a compliance-review chain, not a commercial-decision chain. Subcognit's FHE-native architecture eliminates the largest compliance gates structurally, not contractually.
Data residency review
4 to 6 months at tier-1 bank, collapsed to weeks. Data never decrypts outside the customer's perimeter.
Cross-coalition data-sharing approval
9 to 12 months at MoD-level, structurally eliminated. Cryptographic per-tenant boundaries make coalition-data co-residence a non-issue.
GDPR Article 35 PIA
3 to 4 months at regulated commercial buyers, materially reduced. No plaintext exposure exists to assess.
The dual-stream GTM is a procurement-cycle advantage, not a sales-cycle drag.
One engine. Two products. Same Subcognit underneath. Capital-efficient at seed because we are not building two products, we are productising one engine for two buyer types.
The takeaway ↓
Galiark Sovereign and Galiark Commercial. One engine, two productised wrappers, two buyer types. Capital-efficient at seed.
Our unique proposition ↓

Subcognit gets smarter every deployment.

Compound learning is structural. Every customer deployment makes every other customer's deployment more effective.

Block 01. The mechanism
The causal mechanism transfers across deployments without sharing data.

Every adversarial encounter Subcognit observes feeds into the cognitive architecture's causal model. Not the customer's data, Subcognit operates on encrypted observations and never sees plaintext. The causal mechanism transfer is what improves: Subcognit learns the structure of adversarial behaviour itself, and that structure transfers across deployments without sharing the underlying data.

A defence customer surfaces a novel adversarial pattern in their cognitive cascade. The architectural primitive that detects the pattern's mechanism, not its surface signature, transfers to the next deployment. The next defence customer's first day is the previous customer's last day.

In commercial deployments the same logic holds. A tier-1 bank's fraud ring evolves a new tactic; the architectural primitive that catches the mechanism uplifts every other commercial customer's fraud detection. Without ever sharing a transaction.

Every encounter compounds the architecture. Every customer benefits from every other customer's adversarial exposure. Without sharing data.

Block 02. The compute leverage
Adversarial pressure is the input that compounds. Defence and regulated commercial deployments produce it in unlimited supply.

The architecture is compute-bound. More compute, more deployments, more adversarial encounters, faster compounding.

This is structurally inverted to pattern-matching deep learning, where additional training data has diminishing returns and additional compute hits efficiency walls. Subcognit's causal architecture treats every encounter as additive: adversarial pressure is the input that compounds the system, and adversarial pressure is the one input that defence and regulated commercial deployments produce in unlimited supply.

The more customers Subcognit serves, the harder it is to compete with.

Block 03. The moat compounds
This is the structural moat amplifier the rest of the deck rests on.

A foundation-model-based competitor can match Subcognit's capability snapshot at a single point in time, with sufficient engineering investment, given the FHE substrate is solved. A foundation-model-based competitor cannot match Subcognit's compounded causal-mechanism library accumulated across every prior customer deployment. The moat widens with every customer.

This is why Subcognit is sold as an architecture, not a model.

Models depreciate. Architectures compound.
This is discovery-grade architectural engineering. Models are products. Subcognit is a substrate.
Every customer deployment widens the architectural gap and deepens the unlock value. Two compounds, one architecture.
The takeaway ↓
Subcognit gets smarter every deployment. Models depreciate. Architectures compound.
The proof ↓

Two adversarial proving grounds. Same architecture. Both passing.

Each product has its proof point already in hand.

Proving Ground 01. Galiark Sovereign
Defence, live deployment.
TRL7 certified February 2026, 13 of 13 criteria pass
  • 0real-world cognitive events on live Jetson 4-node federated swarm.
  • 0%adversarial decoy detection across 100 validated missions.
  • p=2.9 x 10⁻¹⁰compound learning significance.
  • Sub-0.1mshot-path cognitive latency.
Proving Ground 02. Galiark Commercial
Cross-domain adversarial validation in regulated financial markets.
The only non-military environment that matches defence for adversarial pressure, deception, and consequence speed

Used as architecture-portability validation, not as a productised trading capability. The same five-engine cognitive cascade, deployed in the most adversarially intense regulated commercial domain available for testing. The architecture's portability between two completely unrelated adversarial domains, defence ISR and financial markets, is the proof that the causal mechanism generalises.

Production-grade backtest under realistic execution constraints captured the following metrics:

  • 0+tests passing under FHE substrate.
  • 0%PnL over 284 day backtest window.
  • 0.00xwin/loss ratio core engine.
  • 0.00Sharpe ratio in backtest, supporting metric.
Methodology caveat. Sharpe ratio measured in production-grade backtest with realistic execution costs, slippage modelling, and walk-forward validation. The headline number reflects causal architecture's adversarial-domain transferability, not a productised trading capability. We are not a hedge fund and have no plans to operate as one. Detailed methodology available under NDA.
Structural correctness ↓
Detection invariance, structurally preserved.

Every architectural milestone the substrate has shipped has preserved the detection signal of every prior milestone exactly. Zero drift. Across every canonical scenario. Across every test run.

This is the strongest single architectural correctness claim in deep-tech AI: that adding cognitive sophistication does not corrupt detection reliability. It is git-verifiable, test-verifiable, and runtime-verifiable.

Not engineered to be impressive. Engineered to be correct.
The architecture's portability between two unrelated adversarial domains is itself the validation that it generalises. Anyone can win in one domain. Two domains, same underlying engine, is structural.
The architecture is correct. That is the only thing the proof is testing for. The numbers are downstream.
The takeaway ↓
Two unrelated adversarial domains, same architecture, both passing. Portability is the validation.
How it works ↓

Subcognit. The intelligence layer.

Not a model. Not a platform. Not a data fusion product. The reasoning layer that should sit underneath modern AI and does not, today, in any production system anywhere.

Bolt any model on top. Subcognit is the intelligence underneath.

01 Causal cognition
Reasons about why.
Adversaries must replicate the underlying mechanism to evade, not just the surface pattern. 100% adversarial decoy detection on validated defence missions.
02 Compound self-learning
Stronger under attack.
Six learning dimensions amplifying each other through cross-engine feedback. The system gets stronger under attack. +4.3% accuracy uplift, p=2.9 x 10⁻¹⁰ across 100 validated missions. The compounding mechanism that makes Subcognit a structural moat amplifier is detailed in Act 5.
03 Prescient inference
Predicts forward state.
Bayesian forward state estimation under causal mechanism constraints, performed inside the FHE substrate. Validated at Sharpe 14.46 cross-domain.
04 Sovereign substrate
FHE-native by design.
All of it runs inside the customer's cryptographic perimeter. Sovereignty is structural, not contractual.

The five-engine cognitive cascade.

Live in production today. Causation engine port queued. Six-engine cascade running in customer environments by Q3 2026.

causal trace fired
E1Distribution
⚷
E2Compression
⚷
E3Attention
⚷
E4Reactive
E5Causation, queued
encrypted observation → Distribution → Compression → Attention → Reactive → causal alert with full explanation

Five engines composed into the cognitive cascade. All internal computation runs under encryption. Output emerges at the customer's perimeter as a causal explanation. Live in production code with 223 tests passing under FHE substrate.

Substrate extension ↓
Where the substrate did not have what production required, we built it.

The available open-source homomorphic encryption substrate did not contain seven of the cryptographic primitives that production cognitive workloads need. We engineered them. Each primitive has been independently characterised under deterministic test conditions and integrated into the substrate-agnostic runtime.

This is the depth of foundational engineering that distinguishes architectures from products. UK patent counsel engaged Q3 2026 to formalise foreground IP.
AI-agnostic. Bolt any foundation model on top. Subcognit is the intelligence underneath.
Built to a standard the production deployment domains demand. The science is rigorous. The engineering is rare.
The takeaway ↓
Five-engine cognitive cascade live in production. AI-agnostic. Bolt any model on top.
The moat ↓

To our knowledge, nobody is building this.

Three movements on the conversion slide.

Movement A. The competitive landscape, layered.

Layer 04
Cognitive reasoning
Galiark (Subcognit). The only one in this layer. Causal, compound, prescient, sovereign, AI-agnostic.
Layer 03
Data fusion
Palantir (Maven, Gotham). Correlational. US sovereign. Documented bias. £240.6M MoD agreement Dec 2025.
Layer 02
Autonomy stacks
Shield AI (Hivemind, $12B Feb 2026). Helsing (Altra). GPS-denied flight intelligence and platform-AI.
Layer 01
Hardware platforms
Anduril ($60B Feb 2026). Helsing post-Grob acquisition. Quantum Systems. Tekever. Vertically integrated drone manufacturing.
FHE substrate
Zama, Duality, Inpher, Enveil. Encrypted compute, no cognition.
Confidential compute
Azure CC, Anthropic Managed Agents, Mistral private cloud. Plaintext-required AI.

Movement B. Why nobody can retrofit.

01. Architectural commitment

Building causal cognition requires a fundamentally different system design than pattern-matching deep learning. You do not add it on. You start over.

02. Composition discipline

A composed five-engine cascade with internal cryptographic boundaries is two years of architectural rulings. Not a refactor.

03. Validation in the wild

Parallel evidence in two unrelated adversarial domains is the only way to prove the architecture generalises. Nobody else has it because nobody else has the architecture that survives both.

04. Meta-learning that cannot corrupt detection

An architectural property unique to Galiark. The cognitive substrate continuously self-improves under encryption, while detection invariants remain mathematically isolated from the learning process. No competitor architecture supports this. No general-purpose homomorphic library makes it available.

Discovery-grade engineering, not iteration on prior art.
05. The data-trapper monopoly

Every regulated organisation in the UK, EU, and globally owns data that is simultaneously the most valuable asset they hold and the most legally radioactive. The compliance wall around it is built by GDPR, sector-specific regulation, and national security regimes. To Galiark's knowledge, no other architecture can lawfully access this data in any commercially deployable form. No federated learning system. No trusted execution environment. No conventional secure multi-party computation. Only FHE-native cognition. Only Galiark.

The competitive set is not other AI vendors. It is the customer's regulatory wall.

Movement C. AI-agnostic positioning.

Bolt any model on top. We are the intelligence underneath.

Bolt any model on top. We are the intelligence underneath.
To our knowledge, nobody is building this, because doing it correctly is harder than the field has currently solved for.
And the markets where this layer matters most are growing at 28 to 34% CAGR. The gap compounds.
And the data behind the wall is worth £25 to £50 billion per year in the UK alone. Galiark holds the keys.
The takeaway ↓
Layer 4 cognitive reasoning is empty. To our knowledge, nobody else is building this.
Why now ↓

Three forces converging in 2026.

A structural opening that did not exist twelve months ago.

Force 01. Regulatory
Correlational AI becomes non-deployable.
EU AI Act August 2026 makes correlational AI legally non-deployable for high-risk classes (credit, law enforcement, healthcare, employment, biometrics). Mandatory explainability, bias auditing, human oversight. Correlational AI on plaintext data cannot satisfy. Causal AI under encryption can.
Force 02. Geopolitical
Sovereignty supercycle.
UK Sovereign AI Unit £500M envelope, £80M competition opens July 2026. EU Office.eu coalition March 2026. EU BuyFromEU directive. Germany €12B drone arsenal. Anthropic-Pentagon clash February 2026 made explainability a contractual requirement for EU procurement.
Force 03. Technical
FHE crossed the practicality threshold.
FHE crossed the practicality threshold in 2024 to 2025. Production cognitive workloads now feasible. Multiple paths to further acceleration through silicon (Cornami) and photonic (Optalysys). Substrate-agnostic CryptoProvider trait targets next-gen accelerators without rewrites.
The window is open now. It closes when one of the established platforms locks the European sovereign-AI default for the next decade.
The takeaway ↓
EU AI Act, sovereignty supercycle, FHE practicality threshold. The window opens 2026 and closes when an established platform locks the default.
The market ↓

£965M UK mature addressable. £6.5B+ international ambition.

One product family, two GTM motions, two distinct addressable markets.

Pillar 01. Galiark Sovereign UK
UK sovereign cognitive intelligence.
£0M
Y5 UK addressable

Five operational domains: Air, Land, Sea, Space, Cyber. Plus law enforcement and government crime: NCA, SFO, HMRC, police forces, allied agencies.

Doctrine: defence, not attack. Aligned with EU and NATO requirements. Not autonomous lethal weapons.

Beachhead: UK Sovereign AI Unit £80M competition opens July 2026.
Pillar 02. Galiark Commercial UK
UK regulated commercial intelligence.
£0M
Y5 UK addressable

Eight verticals across regulated UK AI. Lead three at seed: Banking (£135M Y5), Insurance (£185M Y5), Financial Crime (£50M Y5).

Expansion verticals: Healthcare (£65M), Legal (£170M), Energy and CNI (£112M), Central government beyond defence (£31M), Life sciences (£110M).

Catalyst: EU AI Act August 2026 forces high-risk AI to be explainable, bias-audited, sovereign-deployable.
£0M
Y5 UK mature addressable across both products
Tier 02. International ambition layer
At full international scale (US, EU, Five Eyes, AUKUS), addressable expands to £6.5B+ at Y5.
United States
£2.4B
Y5 addressable. FedRAMP and IRAP equivalents. $100B+ regulated AI market.
European Union
£1.8B
Y5 addressable. EU AI Act enforcement creates the largest single regulatory forcing function globally.
Five Eyes plus AUKUS
£400M
Y5 addressable. Canada, Australia, New Zealand, Japan, South Korea sovereign-AI initiatives.
Trapped value, quantified ↓

Every regulated organisation owns data they cannot lawfully use. The trapped value is not theoretical. It has been quantified.

Bottom-up across four UK regulated-data categories, the latent value locked behind compliance walls that current cloud AI cannot lawfully cross is on the order of £25 to £50 billion per year. The methodology is independently sourced, peer-published, and verifiable.

NHS data
£12 to £15bn
trapped value per year
What unlocks it. Cohort-scale clinical AI on identifiable records inside the NHS Secure Data Environment, processed without plaintext exposure to the AI platform.
Ernst & Young 2019 baseline (£9.6bn) scaled to current NHS spend (£204.9bn in 2024/25); UK Treasury values UK public sector knowledge assets at £150bn with £5bn unrealised annual return.
Fraud and AML
£2.5 to £4.5bn
trapped value per year
What unlocks it. Cross-bank consortium intelligence sharing in FHE-protected form, with no individual bank's customer data exposed. Currently impossible under data residency rules without cryptographic mediation.
UK Finance Annual Fraud Report 2025 (£1.17bn stolen plus £1.45bn prevented); BioCatch Trust Australia benchmark (200 percent uplift in APP fraud detection through cross-bank consortium intelligence) scaled to UK Tier-1 size.
Government data
£8 to £15bn
trapped value per year
What unlocks it. AI on identifiable records inside HMRC, DVLA, ONS, Companies House, and Border Force perimeters, without exposing personal data to AI platforms.
HMRC tax gap (£35.8bn, with 5 to 15 percent AI-addressable share); UK Treasury public sector knowledge assets valuation; departmental AI strategy estimates.
Defence and intelligence
£1.28bn
immediately addressable
What unlocks it. Cross-coalition signals fusion, classified intelligence analysis, and FHE-protected processing of cross-jurisdictional defence data, satisfying coalition information-sharing constraints.
UK Sovereign AI Unit £80m procurement competition launching July 2026, broader £500m SAU envelope, £100m National Data Library, £600m Health Data Research Service.
£25 to £50 billion of trapped value per year in the UK alone. Multiply by the EU at the same proportional shape and you reach £100 to £200 billion. Multiply by mature data-protection jurisdictions globally and you reach $300 to $600 billion.
Markets compound. Architectures compound. Regulatory walls do not move. Galiark holds the only architecture that operates inside them.
Market velocity ↓

Galiark sits at the intersection of four of the highest-growth markets in enterprise software.

The TAM figures above show the size of the addressable opportunity. The figures below show how fast it is compounding.

Sovereign AI infrastructure
0%
CAGR through 2035
$177B by 2035 (Precedence Research, Apr 2026). McKinsey: sovereign AI could be a $600B market by 2030, with up to 40% of AI workloads moving to sovereign environments.
First-mover in the cognitive reasoning layer that runs on sovereign infrastructure.
Precedence Research, Apr 2026; McKinsey & Company, Dec 2025
AI governance and explainable AI
0%
CAGR through 2035
$5.9B by 2035 (Precedence Research, Apr 2026).
Architectural compliance with EU AI Act August 2026 by construction. Causally explainable outputs are not a feature; they are how Subcognit reasons.
Precedence Research, Apr 2026
AI in defence
0%
CAGR through 2031
$32.8B by 2031 (Knowledge Sourcing, 2026); $211.8B by 2035 in the broader AI-in-defence segment (Cervicorn, 2026).
TRL7+ certified February 2026, validated proof, the only architecture in this segment with parallel cross-domain validation.
Knowledge Sourcing, 2026; Cervicorn Consulting, 2026
Fully homomorphic encryption
0%
CAGR through 2030
$1.4B by 2030 (Strategic Market Research).
The first FHE-native cognitive substrate. We are not a consumer of FHE; we extend it where production cognitive workloads require it.
Strategic Market Research; Business Research Insights
Three of these four markets are growing faster than the cloud computing boom of the 2010s. The fourth (FHE) is at parity with the mobile economy of the early 2010s. All four are growing roughly five times faster than enterprise software broadly (~6 to 8% CAGR).
Markets compound. Architectures compound. Galiark sits at the intersection of both.
The intersection ↓

Each of these markets needs the cognitive reasoning layer. None currently has it.

Sovereign AI infrastructure is being built worldwide, but the layer that runs on it remains pattern-matching deep learning.

AI governance frameworks are being legislated, but the architectures that satisfy them remain correlational and unexplainable.

AI in defence is scaling rapidly, but at the architectural level the cognition deployed is the same correlational approach adversaries have learned to defeat.

Fully homomorphic encryption has crossed the practicality threshold, but no production cognitive system runs on it. The substrate exists. The cognitive layer that uses it does not.

Galiark is the first cognitive reasoning layer that satisfies all four conditions simultaneously. Sovereign by architecture. Causally explainable by construction. Validated under adversarial conditions in defence. FHE-native from day one.
First-mover at this intersection compounds. Every customer deployment widens the architectural gap between Galiark and any potential competitor by months of foundational engineering they have not begun.
Methodology. Y3 and Y5 UK addressable derived bottom-up across nine sectors and 45 sub-verticals. 89% of cells anchored to public registers (NHS England Trust register, FCA firm register, ABI membership, College of Policing list, Companies House FTSE constituents, NDA site list, Genomics England public reporting) plus benchmark deals (NHS Federated Data Platform £330M/7yr, NICE Actimize bank refs, Palantir MoD £240.6M, Quantexa HMRC). 11% LOW-confidence cells flagged.
The takeaway ↓
£965M UK addressable bottom-up. £25 to £50bn per year trapped value behind the wall. Four adjacent markets compounding at 28 to 34% CAGR. First-mover at the intersection. Holder of the keys.
Lighthouse customers ↓

Five lighthouse customers. £6.0M Y1 anchor pipeline.

Named accounts. Anchor contract values. Regulatory drivers. Warm-intro paths in motion.

Lighthouse 01. Tier-1 UK bank cross-desk surveillance
Cryptographic Chinese-wall reasoning that the incumbents cannot architect.
Anchor account
Tier-1 UK retail or investment bank.
Y1 contract value
£1.83M (LTV £6M+ over 3-year SaaS tenure).
Workload
Cross-desk market-abuse surveillance under FCA enforcement window 2022 to 2024 pressure.
Why us
NICE Actimize, Behavox, Eventus cannot solve the Chinese-wall constraint that Galiark resolves cryptographically. FHE substrate enables cross-desk reasoning without information barrier breaches.
Path
Warm introduction via Galiark commercial network. Procurement runs through the bank's Head of Financial Crime or Chief Risk Officer.
Lighthouse 02. NHS Foundation Trust
Diagnostic pattern recognition under sovereign data residency.
Anchor account
NHS Foundation Trust (215 Trust addressable population).
Y1 contract value
£400K (LTV £1.3M over 3-year tenure).
Workload
Cross-modality diagnostic pattern recognition under NHS Data Security and Protection Toolkit.
Why us
Gates DSPT compliance by construction. Cloud-native AI cannot deploy under NHS sovereign data residency rules. Galiark's FHE substrate is the only architecturally compliant path to AI-assisted diagnostics at Trust scale.
Path
NIHR and NHS AI Lab adjacencies, Genomics England Discovery Forum networks. NHS Foundation Trusts procure independently under the Provider Selection Regime.
Lighthouse 03. UK MoD coalition signals
Cross-coalition signals analysis without exposing classified ally data.
Anchor account
UK Ministry of Defence (DSIT Sovereign AI Unit envelope).
Y1 contract value
£1.5M (LTV £15M over 10-year framework).
Workload
Cross-coalition (Five Eyes plus NATO ACT) signals intelligence analysis.
Why us
Verbatim Sovereign AI Unit policy match. Palantir cannot solve the coalition data-sharing constraint without exposing each ally's classified data to the others. Galiark's per-tenant cryptographic boundaries do.
Path
Defence and Security Accelerator (DASA) inbound route, DSIT direct engagement, defence prime channel partners (BAE, Thales, Babcock).
Lighthouse 04. Genomics England
Encrypted genomic-scale analytics across federated NHS sites.
Anchor account
Genomics England Ltd (with NHS Genomic Medicine Service expansion path).
Y1 contract value
£1.2M (federated NHS GMS expansion adds projected £3M+ over 5 years).
Workload
Encrypted genomic-scale analytics across federated NHS sites without exposing individual genomic data.
Why us
Existing federated-analytics market (Aridhia, Lifebit) has primed buyer willingness to pay. Galiark removes the residual privacy constraint those vendors cannot solve. Sovereign-AI policy directly aligned.
Path
Genomics England Discovery Forum, Health Data Research UK (HDR UK), NIHR.
Lighthouse 05. Nuclear sector operations
Cross-site encrypted operational analytics in NCSC sovereign-tech environments.
Anchor account
Sellafield Ltd or EDF Energy nuclear fleet.
Y1 contract value
£1.0M+ (multi-year operational frameworks typical).
Workload
Cross-site encrypted operational analytics under ONR, JSP 440, Nuclear Industries Security Regulations.
Why us
Regulatory floor uniformly high. NCSC sovereign-tech preference. Cloud-native AI is structurally locked out of UK nuclear environments. Almost no incumbent competition.
Path
Nuclear Decommissioning Authority innovation programme, Dstl nuclear partnerships.
Y1 anchor pipeline aggregate: £0M. Realistic Y1 close rate: 60 to 80 percent of pipeline (£3.6M to £4.8M lighthouse revenue). Plus IUK Phase 2 grant £1.0M and advisory revenue ~£0.4M. Y1 total revenue mid-case: £5.0M.
The takeaway ↓
Five named anchors, £6.0M Y1 pipeline, regulatory drivers per cell, warm-intro paths in motion.
Commercial model ↓

From £5M Y1 revenue to £95M Y5 revenue. Exiting Y5 at £125M ARR run-rate.

7.8% UK mature addressable capture.

Pricing benchmarked against public sector deals. Forecast built bottom-up from named lighthouses. Financials recalibrated for £5M seed velocity.

01. Pricing.
Galiark Sovereign pricing
Per-deployment licence: £750K to £2M annual base.
Integration services: £200K to £1M one-time.
Annual support: 20% of licence.
Multi-year framework agreements: £5M to £20M+ lifetime contracts.
Reference benchmarks: Palantir MoD £240.6M Dec 2025, NHS FDP £330M/7yr, DASA Phase 1/2 envelope, MoD Multi-Source Procurement Framework awards.
Galiark Commercial pricing
Enterprise tenant licence: £750K to £2M annual base.
Per-decision metering: £0.001 to £0.01 per decision above contracted floor.
Implementation: £300K to £800K one-time.
Premium support: 18% of licence.
Reference benchmarks: NICE Actimize bank deals, Quantexa HMRC and bank refs, Darktrace customer references, ABI fraud-spend benchmarks.
02. Five-year revenue trajectory.
Y1
£0M
Y2
£0M
Y3
£0M
Y4
£0M
Y5
£0M
In-year revenue, Galiark Sovereign
In-year revenue, Galiark Commercial
ARR exit run-rate at year-end
Methodology. Revenue is in-year recognised P&L, including partial-year ramp on new customers. ARR exit is the annualised run-rate value of contracts in place at year-end. Series A funded scale takes effect from Q3 2028, separating the two trajectories from Y3 onward.

Pricing methodology. Published numbers anchor on the conservative comparable-product floor (public sector AI procurement benchmarks, £500K to £2M ARR per enterprise tenant licence). Lighthouse customer engagements anchor on a sliding scale toward unlock-share pricing, a small percentage of value Galiark unlocks for the customer (typically 1 to 5 percent of measurable annual unlock value, NDA-protected per cell). Floor case is what the deck shows. Ceiling case is materially higher per cell, available under NDA.
YearCustomersAvg ACVIn-year revenueARR exit (run-rate)Sov / Com split
Y1 (FY26-27)4£1.4M£5.0M£5.0M60 / 40
Y2 (FY27-28)7£1.5M£11.0M£14.0M55 / 45
Y3 (FY28-29)16£1.6M£24.0M£32.0M50 / 50
Y4 (FY29-30)32£1.7M£52.0M£68.0M45 / 55
Y5 (FY30-31)56£1.8M£95.0M£125.0M40 / 60
Y1 to Y2 UK-led, smaller seed funds slower velocity. Series A trigger Q1 to Q2 2028 on £14M ARR.
Y3 sees Series A close in 2028 funding international expansion. UK now 65% of revenue.
Y5 customer mix: 22 Sovereign (12 UK MoD / SAU / NCA / nuclear, 10 international defence), 34 Commercial (20 UK regulated, 14 international regulated).
03. Unit economics (steady state, Y3 to Y5 onward).
Galiark Sovereign
ACV
£2.0M
LTV
£15M
CAC payback
18 to 24 mo
Gross margin
80%
Galiark Commercial
ACV
£1.7M
LTV
£5.5M
CAC payback
12 to 18 mo
Gross margin
76%
Blended at Y5
Average ACV
£1.8M
Blended LTV-to-CAC
4.5x
Blended gross margin
78%
04. ARR milestones.
£1M
Q3 2027
£5M
Q1 2028
£14M
Q4 2028
£32M
Q4 2029
£68M
Q4 2030
£125M
Q4 2031
05. UK capture cross-check.
UK ARR / Y UK addressable
Y3: £20M UK ARR / £312M Y3 UK addressable = 6.4% capture
Y5: £75M UK ARR / £965M Y5 UK addressable = 7.8% capture
Conservative versus typical category-leader Y5 capture range of 10 to 15%. International revenue (US, EU, Five Eyes) layers on top. £5M seed velocity assumed; faster capture available with Series A acceleration.
The takeaway ↓
£5M Y1 to £125M Y5 ARR. 7.8% UK capture at maturity. Pricing anchored to public sector benchmarks.
Team and build state ↓

Two senior operators. Two confirmed advisors. Production code live.

Founder team plus governance bench plus evidence base.

CTO, co-founder
Tom Bennett

35 years across global financial markets, predictive AI, and proprietary AI software development.

Started in 1987 trading FX and money markets at Westpac New Zealand. Joined Dow Jones London 1994 to set up the markets group across all asset classes. Headhunted to Bankers Trust New York 1997 as CTO of the Risk Group, the only bank-owned risk modelling platform regulators permitted banks to run against their own books. Led the platform's spin-out as Global Head of Trading Group at IQ Financial Systems. Acquired Rolfe and Nolan's Lighthouse, rebranded it Trade IQ, led the global rollout: Credit Suisse First Boston, Merrill Lynch, the four largest Japanese banks. The largest global trading-system deployment of its era.

Independent financial consultant 2001 to 2015. Beat every major management consultancy to win the contract building the predictive model for the World Economic Forum at Davos.

Built a proprietary deep learning methodology at Deep Pattern Learning (2015 to 2019). Co-founded Parzival Partners in 2020 on more than twenty years of proprietary AI software development.

Architect of the cognitive ontology that underpins Subcognit. Foreground IP contributed to Galiark Ltd at incorporation under documented assignment. Co-author of all live homomorphic cognitive engines (Reactive, Attention, Distribution) on Galiark's proprietary OneGrail build platform. Advisory Board member at HGP Technologies advising on the QadraAI enterprise AI platform for institutional investing.

CEO, founder
Matt Tyler

25 years across deep-technology commercialisation: project management, property, property development, construction technology, fintech, agritech, defence technology and executive search.

Currently CEO of Herbi4 Ltd, commercialising a patented metabolic nutrient backed by twenty years of research that cuts livestock methane by approximately 65% while increasing yields. Structuring international project finance across the UK, EU, US, Brazil, and Argentina. Operational handover to Herbi4's incoming CEO is in motion. Galiark becomes Matt's full-time focus on completion.

Co-founder and former Chief Commercial Officer of Tech8 Digital, where he built the commercial strategy, brand positioning, and go-to-market for a next-generation banking platform bridging traditional banking with digital-asset infrastructure. Now Co-Founder and Strategic Advisor to Tech8.

Founding Partner of Parzival Partners since 2018.

Architect of the Galiark commercial strategy: sovereign cognitive infrastructure positioning, services-first to enterprise tenant commercial model, and regulated sector target taxonomy. Established the project management discipline (sprint cadence, ADR governance, build board, deterministic runtime evidence) that underpins the current evidence base. Co-author of live engine implementations through parallel co-development with the CTO on the OneGrail build platform.

Senior advisors in active dialogue
Three senior advisors in active dialogue, expected to confirm Q3 2026.

Three senior advisors in active dialogue with Galiark, each bringing a distinct strand of governance, scaling, or regulatory pedigree directly relevant to the dual-stream go-to-market and the regulated buyer environments.

Defence-tech main board
Currently on the main board of a NASDAQ-listed US defence group.
Cybersecurity scaler, ex-C-suite
Former C-suite at one of Europe's largest commercial cybersecurity exits.
UK regulated financial services, board governance
Senior independent director and Risk Committee Chair across multiple FCA-authorised entities.
Names available under NDA on request. Confirmation expected Q3 2026 on conclusion of current contracts.
Senior bench, activating with seed close

Planned hires.

  • VP Engineering (September 2026). Senior Rust engineer with production systems experience in regulated environments. Familiarity with applied cryptography, fully homomorphic encryption, agentic systems, secure multi-party computation, or confidential computing. UK based. Owns the performance hardening pass, foundation model integration spike, and Phase 2 demonstrator hardening.
  • Defence Senior Advisor (August 2026 retainer). Sovereign AI Unit / DSIT or UK Defence Prime archetype.
  • Commercial Lead (Q1 2027). Regulated-sector enterprise sales.
  • Three FHE engineers phased October 2026 to April 2027.
  • End Y1 headcount: 6 FTE.
External specialist engagements (already scoped)
  • Independent Rust and cryptography code reviewer (Month 4). External due diligence on the homomorphic cognitive architecture, with sign-off requirement before white paper publication.
  • UK patent counsel (Months 2 to 4). Foreground IP capture. Drafting and filing of UK patent application covering the trait-based homomorphic composition pattern and the single-decryption-gate discipline.
Cadence proof. Two-person team that ships at multi-engineer cadence is the delivery model. Parallel engineering operating system delivers measured 2.2x to 12x throughput multipliers across sprint cycles.
Production state
TRL position
Defence application certified TRL7+ February 2026, 13 of 13 criteria PASS, 0 cognitive events on live federated swarm, 0% adversarial decoy detection. Integrated FHE-native Subcognit stack at TRL 6 to 7, hardening funded by this seed.
Build discipline

Five architectural milestones shipped in five days. Every milestone clean. Every performance forecast hit within band. Every architectural ruling held under post-hoc audit. 0 tests passing under FHE substrate at the substrate's current architectural state.

Two-person engineering team, working through a parallel engineering operating system, advanced the cognitive substrate through five consecutive architectural milestones with zero detection-signal drift. UK patent counsel engaged Q3 2026 to formalise foreground IP across the architectural patterns and the substrate extension primitives.

The cadence is not extraordinary because it is fast. It is extraordinary because it is correct under speed.
What this round funds
Productisation of Galiark Sovereign for first defence-track lighthouse. Productisation of Galiark Commercial for first regulated commercial lighthouse. Engineering team scale to 6 FTE end of Y1. Advisor bench engaged. Compute infrastructure procurement. Path to 6-engine Subcognit cascade running in customer environments by Q3 2026.
The takeaway ↓
Two founders, three senior advisors in active dialogue, production code live, TRL7+ defence application certified.
Risk and mitigation ↓

The risks. The mitigations.

We have thought about what could go wrong. Here is the list, and what we are doing about each.

Risk
Mitigation
TRL slippage on integrated FHE-native stack between current TRL 6 to 7 and customer-deployable TRL 8.
Hardening funded by this seed round. External Rust and cryptography code review at Month 4 of seed. UK patent counsel filing covers the integration pattern. The performance hardening pass on synthetic workloads of 100, 500, 1000 observations is the integration-readiness gate.
Defence procurement timing. UK SAU competition, framework agreements, and pilot conversion take 12 to 18 months and have material slip risk.
Dual-track GTM is the structural mitigation. Galiark Commercial procurement (6 to 12 months, enterprise software pathway) runs in parallel and does not depend on defence procurement timing. And the FHE substrate compresses both procurement cycles materially: data residency review collapses from 4 to 6 months to weeks, cross-coalition data-sharing approval is structurally eliminated, GDPR Article 35 PIA materially reduced. The dual-stream GTM is a procurement-cycle advantage, not a sales-cycle drag. Worst-case defence slip preserves Y2 revenue trajectory through Commercial expansion.
EU AI Act enforcement timing or scope drifting from the August 2026 trigger that drives Galiark Commercial demand.
Galiark Commercial value proposition (causal explainability, FHE substrate, sovereign deployment) holds independently of EU AI Act timing. Commercial buyers face FCA, EBA, and SEC explainability pressure that does not rely on the EU AI Act schedule. Forcing function delays soften the timeline; they do not soften the demand.
Key-person concentration on Tom Bennett as the architect of the cognitive ontology.
VP Engineering hired Month 1 of seed. Architecture is documented in 44+ ADRs. External code review at Month 4 produces a third party who can validate the architecture under independent inspection. UK patent counsel filing crystallises the foreground IP at Galiark Ltd, not at any individual. Architecture is reproducible from the documentation; this has been engineered for from day one.
Foundation model commoditisation changing the AI competitive landscape underneath us.
Galiark is AI-agnostic by architectural commitment. Foundation model commoditisation accelerates Galiark's positioning rather than threatens it: as foundation models become cheap, the differentiated value sits in the reasoning, sovereignty, and explainability layer underneath. That is exactly what Subcognit is.
The takeaway ↓
We have thought about what could go wrong. Each risk has a structural mitigation.
The ask ↓

Galiark is raising a £5M Series Seed.

Lead anchor £2 to £3M. 18-month runway funding the productisation of both products, the first lighthouse in each stream, and the path to Series A trigger event. Terms in market.

Aggressive, structured, targeted growth toward early market dominance in the cognitive intelligence layer.

Use of funds.

Engineering team scale (6 FTE end of Y1)
45%
£2.25M
Lighthouse customer pursuit, both products
25%
£1.25M
Defence advisor bench
10%
£0.5M
Compute infrastructure
8%
£0.4M
G&A, legal, IP, runway buffer
8%
£0.4M
IUK Phase 2 plus EU funding match
4%
£0.2M
Path to Series A. Q1 2028. £14M ARR.

£5M raised, £5M revenue achieved, two lighthouse pilots signed. Conditions met for Series A close:

  • At least one Galiark Sovereign lighthouse pilot signed by end Y1.
  • At least one Galiark Commercial lighthouse pilot signed by end Y1.
  • IUK Phase 2 won (£1M).
  • 6-engine Subcognit cascade live with Causation Engine ported.
  • Production code reviewed and signed off by independent Rust and cryptography reviewer.
  • UK patent application filed on the trait-based homomorphic composition pattern.
  • Two senior advisors formally appointed (Chair and Senior Advisor confirmed Q3 2026).

Series A target: £15 to £25M raise at £150 to £250M pre-money. Defensible 8 to 12x ARR multiple at category-leader Series A in defence-tech and regulated-AI. Funds international scaling, US and EU lighthouses, expansion of engineering and commercial bench.

Anchor candidates
NATO Innovation Fund. Paladin Capital. Shield Capital. Razor's Edge Ventures. Lux Capital. Point72 Hyperscale. NSSIF UK.
Defence-tech and sovereign-data specialists with thesis-aligned mandates.
The takeaway ↓
£5M Series Seed. Lead anchor £2 to £3M. Aggressive structured targeted growth toward early market dominance.
Two products. One engine. Two adversarial markets.
Subcognit gets smarter every deployment.
Structurally different. AI-agnostic.
The category nobody else has built. Aggressive, structured, targeted growth toward early market dominance. The window is open now. And closing.
Book an investor briefing.
matt.tyler@galiark.com