CONCEPT

Protect your data.Put AI to work.

HYBRID PRIVATE AI

This page presents a deployment concept. Hardware specs, software, and scope of delivery will be confirmed through PoC and requirements analysis.

Consult on Deployment See How It Works

How It Works

Route each task to the right AI — automatically.

Employee requests pass through a policy engine that routes each task to in-house AI or approved cloud AI based on data sensitivity.

Employee Task

Policy Engine

Confidential / Internal

In-House AI

Approved / General Use

Approved Cloud AI

Access Control

Transmission Control

Audit Log

Employee Task

Policy Engine

Access ControlTransmission ControlAudit Log

Confidential / Internal

In-House AI

Approved / General Use

Approved Cloud AI

Confidential data is processed exclusively by in-house AI with external transmission blocked. Access to approved cloud AI is governed by policy, with audit logs recorded automatically.

Conceptual diagram. Implementation and validation of routing, transmission control, and logging are required.

Key Use Cases

Solving real workplace challenges with AI.

Enterprise

Internal Document Search & Knowledge

Streamline document creation and knowledge management with secure in-house AI.

Finance & Audit

Closed-Network AI for Audit & Risk

Analyse internal data and generate reports with AI under strict governance.

Healthcare

In-Hospital Closed LLM & Medical Docs

Support medical documentation and workflows within hospital compliance rules.

Start Small, Validate, Scale

8-GPU architecture and a 3-step deployment process.

GPU Allocation Overview (8-GPU Config)

GPU 01

GPU 02

GPU 03

GPU 04

GPU 05

GPU 06

GPU 07

GPU 08

The 8-GPU layout is a reference example. Actual role allocation and unit count are confirmed during PoC and requirements review.

3-Step Deployment

Requirements

Define target workflows and scope of use.

PoC & Evaluation

Validate under near-production conditions and confirm fit.

Deployment & Ops

Deploy based on PoC results and begin operational use.

Consultation

Build the AI foundation that fits your business — together.

Configuration and scope of delivery are confirmed during PoC.

PAGE CONTENTS

Overview & Strategy

EXECUTIVE SUMMARY

Not just hardware — an AI foundation for your enterprise.

Price

8 GPU / 768 GB

Planned Config

3 Domains

Use Cases

Hybrid AI Workspace

Auto-routes confidential vs. general queries. Staff use GPT, Claude, and local LLMs from one interface.

Internal Audit & Compliance

Cross-references contracts, invoices, journals, and emails to surface anomaly candidates with rationale.

Healthcare Closed-Network AI

Staff FAQ and in-hospital workflow support. Purpose is operational efficiency, not diagnosis.

Design Policy

Designed for 8-GPU configuration, with capacity for long-form confidential processing and headroom under peak load. Final configuration is confirmed through PoC.

HYBRID AI RATIONALE

Three reasons closed-network alone is not enough

Closed LLM Only — Challenges

Top-tier OSS models are heavy; one unit limits concurrent users

Handling all staff generation demand on-premises is cost-inefficient

High-performance models conflict with audit and background processing

Process safely and efficiently via the optimal route

Staff Request

Judgment AI (Kev)

Analyzes content and routes to the optimal destination

External AI

GPT, Claude, etc.

Closed AI

On-premises environment

Hybrid AI Advantages

General generation tasks use external AI such as GPT and Claude

Confidential, personal, and proprietary data processed on-premises only

Routing only confidential content to closed-network is optimal for large enterprises

Rather than keeping everything on-premises, protecting only confidential parts on-premises is the optimal design for large enterprises.

STRATEGIC SHIFT

From distributing the latest LLMs to enterprises controlling what information leaves.

EXIN mediates all AI access, centralizing confidentiality judgment, transmission control, and audit logging.

GPT

Claude

Gemini

Internal AI

AI Gateway (Kev + DLP + IAM)

Use familiar AI as-is

Staff continue using existing AI tools. No change in workflow.

Confidential data stays internal

Personal info, proprietary data, and unpublished information never leave the network.

No burden on staff to decide

The AI gateway auto-classifies. Staff can use AI without hesitation.

CUSTOMER PAIN

Confidential data leakage risk varies by department

Sales & Planning

Client info, next-year sales plans, proposal pricing

Product Development

Unreleased products, research plans, cost data

HR & Legal

Personal data, contracts, confidential documents

Don't use AI

Productivity drops

AI Gateway that enforces policy

A third option that solves both problems

Use without consulting

Risk of data leakage

Product & Design

PRODUCT CONCEPT

Confidential data stays internal. General tasks use world-class cloud AI.

EXIN automates routing to maximize utilization while maintaining data governance.

All Staff

Unified AI Workspace

Local Policy Judgment (Kev + DLP + IAM)

Contains confidential / personal data

Processed by closed-network LLMs such as Qwen / GLM. Input, RAG, and responses never leave the network.

Approved general information

Sent via approved gateway to GPT / Claude API. Latest models remain available.

Same interface, same workflow. Users do not need to decide what can be sent.

SYSTEM DESIGN

Strict transmission control with the internal network as the boundary.

Staff AI Workspace

Internal App

SSO / RBAC

Body, attachments, history

Data classification

EXIN Judgment Layer

Zero-trust judgment

Policy DB + confidential registry

DLP + Kev judgment

Audit log / hold

Approved Route

Forced routing

Internal segment

Controlled API egress

External send blocked on failure

Processing Engine

Inference / generation

C: Local Qwen / GLM

A/B: GPT / Claude API

D: Block / approve

Direct external sends and individual browser use are controlled separately via MDM / proxy / SASE. Note: immersion cooling itself has no data-leakage prevention function.

8 GPU FULL ARCHITECTURE

Assumed configuration for this plan: all 8 GPUs assigned to real workloads.

Model names, versions, licenses, and GPU allocation to be confirmed in PoC

GPU 01

Kev + DLP

Judgment & send control

GPU 02

Qwen3.8-27B

Fast confidential processing

GPU 03

GLM-Flash

High-precision confidential inference

GPU 04

GLM-Flash

TP2 (paired with GPU 03)

GPU 05

GLM-Flash

Inference replica 2

GPU 06

GLM-Flash

TP2 (paired with GPU 05)

GPU 07

Audit / RAG

Long-form & audit batch

GPU 08

Audit / RAG

Concurrent processing & isolated slot

This concept assumes 8 GPUs. Not 100% utilization at all times — designed for peak headroom and isolation.

REFERENCE PERFORMANCE

Reference Performance (from documentation)

Measurement environments, metrics, and workloads differ; direct comparison is not valid. These are not measured or guaranteed values from IMMERSIA hardware.

1,246 queries/sec

Laya (PRO 6000) — short judgment queries

51.4 req/s

Kev-4B (L40S / 64 clients) — short text

Software Design

Primary DLP → Kev → uncertain cases: strict rules / LLM / human — 3-stage judgment

Validation Performance

Validates Japanese confidential miss rate, long-form, p95, and false positives. Metrics differ; direct comparison not valid.

CAPACITY PLANNING

Large-scale coverage is possible because cloud and closed-network share the load.

Design model: 10 requests/day per person, 8 hours, 10% confidential route, 10x peak arrival

Registered Staff	Total AI Judgments/Day	Closed Inferences/Day	Closed Peak (req/sec)
500	5,000	500	0.17
1,000	10,000	1,000	0.35
2,000	20,000	2,000	0.69
5,000	50,000	5,000	1.74

Important: This is a demand estimate for arrival volume. It does not guarantee that generation models can process without wait time.

USER EXPERIENCE

One interface — staff never need to think about which model is used.

EXIN AI WORKSPACE

Staff AccountConfidential Policy Active

Based on the terms with Company A and next year's sales plan, please draft a proposal.

This request contains 'internal confidential' data

Destination: Internal Qwen / GLM

External API send: None

Audit log: Recorded

Screen example — not actual logs

Use familiar high-performance AI

Maintain productivity while providing a safe AI environment.

Confidential data stays internal

Critical information processed on-premises to ensure data safety.

No burden on staff to decide

Proper governance so staff can use AI confidently without hesitation.

Use Cases & Economics

PRODUCTIZATION

One hardware unit. Three products differentiated by use case.

Application-level differentiation confirmed through PoC for each customer's requirements

Large Enterprise

Kev AI classifies confidentiality. Personal and proprietary data routed to internal Qwen/GLM; other data safely sent to GPT/Claude API.

Finance & Accounting

Cross-references contracts, invoices, journals, and emails to surface anomaly candidates with rationale. Final review by specialists.

Healthcare

Staff FAQ and in-hospital workflow support. Operational efficiency, not diagnosis. Subsidy eligibility confirmed individually.

Purchase / Lease / Installment + Setup & Maintenance

Terms subject to individual negotiation

BUSINESS PILLAR 02

Monitor finance and accounting with closed-network AI.

Payee changes, duplicate payments, invoice mismatches

Automatically surfaces candidates for abnormal payee changes and duplicate payments.

Invoice and journal auto-reconciliation

Cross-references invoice and journal data to present inconsistency candidates.

Internal email and approval analysis

Detects abnormal approval flows from internal emails and approval documents.

Contract and evidence reading

Closed-network LLM reads contracts and evidence to surface risk candidates.

Cross-reference

Anomaly candidates & rationale

Final review by specialists

FINANCE POC

Finance & Accounting Joint PoC Concept

This is a planned concept for a joint PoC with Tsuji & Hongo Tax Corporation. Not a confirmed contract or track record.

Contract & evidence reading

Closed-network LLM reads contracts and evidence to surface risk candidates.

Accounting data reconciliation

Cross-references journal and invoice data to present inconsistency candidates.

Payment audit & anomaly detection

Surfaces candidates and rationale for duplicate payments, payee changes, and approval inconsistencies.

Proof-of-Concept Roles (Planned)

WOODMAN: Hardware, immersion cooling, maintenance
iore: Corporate sales, distribution, commercialization
SNAFTY: AI UX, DLP, data integration
Tsuji & Hongo: Accounting practice evaluation, specialist review

MEDICAL POC

For healthcare: staff phone inquiries answered by on-premise AI.

Planned concept for a joint PoC with Yukaria and Howa-kai Hospital. Purpose is staff FAQ and workflow support, not diagnosis. Deployment on IMMERSIA requires separate validation.

Staff mobile authentication

Secure access via phone number and PIN authentication.

Voice recognition, policy RAG, LLM

SIP/SBC → voice recognition → policy RAG → LLM → voice synthesis flow.

Transfer to admin & effect measurement

Transfer to admin when needed. Continuous effect measurement via logs.

Implementation Support

Subsidy eligibility confirmed individually. Subsidy rates and amounts not listed.

Cost

Please contact us.

Planned for initial validation on Mac Studio. Deployment on IMMERSIA requires separate validation.

WHY 8 GPUS

Protection, generation, and audit coexisting

Security Management

High-speed judgment by Kev + DLP. Prevents external transmission of confidential data.

AI Generation

Closed-network LLM processing by Qwen / GLM. Latest models available internally.

Continuous Monitoring

Long-form and audit batch processing by Audit / RAG.

Full configuration is not 100% utilization at all times — designed for peak headroom and isolated slots.

ECONOMICS

Evaluate as an enterprise-wide AI foundation, not just hardware

Hardware & Initial Setup

Please contact us.

Operations & Maintenance

Please contact us.

TCO Estimate Scope

Please contact us.

TCO Estimate Scope

API fees	Contact Us
Power & cooling costs	Contact Us
Hardware maintenance	Contact Us
Financial costs (lease, etc.)	Contact Us

Pricing, monthly fees, and per-user costs are not listed. Please contact us individually.

Structure & Criteria

EXECUTION PLAN

Execution Plan

Party	Responsibility	Initial Activity
WOODMAN	Hardware, immersion, maintenance	Hardware loan & config validation
iore	Corporate sales, distribution, commercialization	PoC support & deal flow
SNAFTY	AI UX, DLP, integration	PoC app development
Business PoC Partner	Accounting & healthcare practice evaluation	Practice validation & feedback

0–3 months: PoC

4–6 months: First unit

7–12 months: Finance & healthcare rollout

Estimated schedule. Does not guarantee delivery dates or availability.

Above are planned roles and estimated timelines. Not confirmed contracts or organizational structures.

THE PROPOSAL

Joint Productization Concept

WOODMAN

Hardware, immersion cooling, maintenance. Provides and maintains IMMERSIA hardware.

iore

Corporate sales, distribution, commercialization. Customer development and deal structure.

SNAFTY

AI UX, DLP, data integration. Application development and integration.

To be explored as a standard package through joint PoC. Not a confirmed contract or track record.

ACCEPTANCE CRITERIA

Success is not GPU utilization — it is customers using AI with confidence.

Verify accuracy and response with the 8-GPU hardware before production deployment.

01 Confidential Miss Rate

Missed sends of personal info, contracts, unreleased products

02 Routing Accuracy

Accuracy of Cloud / Local / Redact / Deny

03 Response Performance

Kev p95 target 500ms; Local LLM SLA per model

04 Real-world Load

Equivalent to 1,000–2,000 staff, long-form 8K/32K (PoC target)

05 Audit Effectiveness

Anomaly candidate extraction, false positives, staff time saved

06 Operations & Security

Egress control, audit logs, fail-safe behavior

These are PoC target metrics. They do not guarantee performance or accuracy.

5-PART SERIES

Hybrid Private AI Explainer Series

View All Columns

The content on this page represents a deployment concept. Hardware specifications, software, scope of delivery, and reference figures are subject to confirmation during PoC and requirements review. No pricing, cost, or subsidy information is provided.

Immersion cooling functions under appropriate installation and operating conditions. It does not guarantee prevention of liquid leakage.