TPEX 8299Public company (8299.TWO), global leader in NAND controllers and storage. Market cap over USD $18 billion.
Air-gapped GenBI with Phison
GenBI that never leaves your building.
Private models on the on-prem Phison AI Data Platform. Fully air-gapped: no cloud LLM, no data egress.
Solution Brief. For CIOs, CISOs, and data and AI teams in regulated industries
Get the reference architecture
Free PDF0
Data sent outside your network
10+× Faster
TTFT vs. recompute (aiDAPTIV+)
10 min
To a first answer on your schema
Why this partnership matters
TPEX 829916k+GitHubOpen-source GenBI agent and context layer for enterprise data, shared by your people and your AI agents. 17K+ GitHub stars.
The problem
GenBI touches your most sensitive data. Cloud LLMs send it out the door.
To answer one question, a GenBI agent reads your schema, writes SQL, and reasons over the results. On a cloud LLM, every step is data leaving your network.
Three questions no one will send to a public API
Hospital CFO
“Which departments ran over 90% bed occupancy last week, and how does that compare to the same week last year?”
Patient-level data. It cannot leave the hospital network.
Supply chain director
“Which suppliers missed SLA more than twice this quarter, and where is it concentrating by region?”
Supplier contracts and pricing. Confidential, on-prem only.
Bank risk officer
“Q1 loan defaults by region vs. our reserve targets. I need it for the risk committee tomorrow.”
Customer credit data. Residency rules keep it in-house.
The model goes to the data. Not the other way around.
Product highlights
Sovereign GenBI, from silicon to answer, in one air-gapped stack.
Air-gapped by default
Fully air-gapped, end to end.
Models, context layer, memory, and logs all run inside your network.
Private models
Private models, without a GPU cluster.
aiDAPTIV+ extends GPU memory into NAND, so smaller systems run larger local models.
Governed agent
Governed from the database up.
Row- and column-level access, validated SQL, and a full audit trail before any query runs.
How it fits together
From your hardware to business answers.
Vertical integration path
Wren AI × Phison AI Data Platform
SaaSAPI / AI Gateway
Audience
Humans and AI agents
Decision makersBusiness usersData analystsClaude CodeChatGPTEnterprise CopilotAgentic workflows
Wren AI: governed data agent (GenBI + MCP)
Native applications & intelligent context layer
Wren AISaaS
Chat-to-SQLGenBI dashboardsSlack and Teams embedEmbedded AI apps
Context Model (MDL)Memory & recallGoverned access (RLS, CLS)Eval & qualityAudit logs
Special vertical integration
Governed context meets local AI runtime.
Wren's MDL, access rules, and eval loop connect into Phison's API / AI Gateway so trusted questions can run against on-prem models and data without migration.
MDL through the API gatewayRLS and CLS on local dataEval loop on a monitored runtime
Vertical integration pathPhison AI Data Platform
↕ API / AI GatewayPaaS
L4
Reusable AI Modules Layer
Vision ModuleVoice ModuleReasoning ModuleData BridgeMLOps & LLMOps
L3
Reusable Open Source Components Layer
aiDAPTIVAI FrameworkMessage QueueDatabaseNetworkMonitoringAI Models
IaaS
L2
Hyper-Converged Infrastructure Software LayerPhison HCI
Software Deployment & Container MgmtCPU, GPU & VM Resource MgmtvGPUApp MarketplaceStorage MgmtCost MgmtMonitoring & AlertingSecurity & Access Control
L1
Infrastructure Layer
CPU ServerIntel, AMD, ARMGPU ServerNVIDIA, AMD, IntelCache & Storage ServeraiDAPTIV Cache Memory, Pascari SSD
API vs. fixed cost
Every question costs tokens on the cloud. On the appliance, it costs nothing extra.
A GenBI agent makes a dozen LLM calls per question, for every user, every day. On metered APIs that bill grows with adoption. On the appliance, it's hardware you already own.
Metered cloud APIsPhison appliance (fixed cost)
- 100 users
- Roughly a wash. Metered looks cheap while few people ask.
- 1,000 users
- Every new user adds tokens to the bill. The appliance costs the same at 1,000 as at 100, so the more you roll out, the further metered falls behind.
GenBI on metered cloud APIs
- Cost driver
- Per token: every question, retry, and agent step
- As adoption grows
- Spend climbs with success
- Budgeting
- Variable, hard to forecast
- Lifetime cost
- Pay again for every question, every year
GenBI on the Phison appliance
Fixed cost- Cost driver
- Hardware you size once, run at capacity
- As adoption grows
- Cost per question falls
- Budgeting
- One fixed line item, no usage caps
- Lifetime cost
- Paid once. Extra questions cost electricity, not tokens
The ROI math
The more your teams use it, the faster it pays back.
- Marginal cost per question
- Electricity, not tokens
- Usage caps
- None
- Cost per question
- Falls as adoption grows
A deployment, not a demo script
Seven days from install to production.
Day 1 morning
Connect your database
Postgres, MySQL, ClickHouse, Snowflake, Redshift, DuckDB, plus 20+ more.
Day 1 afternoon
Auto-MDL builds your model
Wren reads your schema; data teams review joins, metrics, and access.
Day 2–3
Seed real business questions
Capture what finance, ops, and leadership actually ask.
Day 3–5
Tune with the business
Refine the context layer until accuracy clears the bar.
Day 5–6
Pilot with users
Turn on governance and validate with a live group.
Day 7
Roll out and connect agents
Go live for people and AI agents on one governed interface.
Customer voices
What leaders say.
With Wren AI on Phison aiDAPTIV+, enterprise AI deploys in days, not months. It supports 20+ databases with no migration and reaches up to an 80% cache hit rate in the on-prem AI infrastructure.
Wei, Lin
CTO, Phison Electronics (8299.TWO)
Wren AI powers our SaaS product DemandSense, so customers ask in natural language and build charts and dashboards in real time. Putting analytics in non-technical hands is a game changer.
Anna Khamitova
CTO, Impactable (US)
Wren AI changed how we decide. Real-time, natural-language insight lifted our decision speed by over 50% and made data a company-wide asset.
Shasta Ho
CEO, Nextlink (TPEX:6997)
Phison × Wren AI: air-gapped GenBI with private models
Bring your database. Keep it in the building.
See a private model answer questions on your own schema, fully on-prem, in under 10 minutes.
×Solution BriefFor CIOs, CISOs, and data and AI teams in regulated industries01 / 02
Air-gapped GenBI with Phison
GenBI that never leaves your building.
Private models on the on-prem Phison AI Data Platform. Fully air-gapped: no cloud LLM, no data egress.
*0*Data sent outside your network
10+× *Faster*TTFT vs. recompute (aiDAPTIV+)
*10* minTo a first answer on your schema
01The problem
GenBI touches your most sensitive data. Cloud LLMs send it out the door.
To answer one question, a GenBI agent reads your schema, writes SQL, and reasons over the results. On a cloud LLM, every step is data leaving your network.
Why this partnership matters
Phison brings proven enterprise AI infrastructure and deployment reach. Wren AI adds the governed context layer that turns that infrastructure into a practical business-answering system. Together, the partnership closes the gap between owning AI hardware and getting trustworthy answers from enterprise data.
For banks, hospitals, manufacturers, and government, that is not a risk to manage. It is a line that cannot be crossed, so GenBI has to run on-prem, on private models.
02The solutionProduct highlights
Sovereign GenBI, from silicon to answer, in one air-gapped stack.
Air-gapped by default
Fully air-gapped, end to end.
Models, context layer, memory, and logs all run inside your network.
Private models
Private models, without a GPU cluster.
aiDAPTIV+ extends GPU memory into NAND, so smaller systems run larger local models.
Governed agent
Governed from the database up.
Row- and column-level access, validated SQL, and a full audit trail before any query runs.
03The stack
Hardware layerPhison aiDAPTIV+
GPU memory that scales into your SSDs.
aiDAPTIV+ middleware extends GPU memory into an SLC NAND SSD cache, so smaller GPU systems run larger models.
- Extends GPU memory by up to 8 TB with aiDAPTIVCache SSD.
- Run larger models on smaller GPU footprints, no cluster required.
- Fully on-premises deployment for air-gapped environments.
- Practical economics for enterprise teams outside hyperscale budgets.
Agentic GenBI LayerWren AI GenBI
From data to answers in under ten minutes.
Open-source GenBI agent and agent-agnostic context layer that closes the analytics gap.
- Auto-MDL builds the context model from your schema.
- Natural-language queries return SQL-backed, governed answers.
- agent-agnostic: analysts and AI agents share one interface.
- RBAC, audit trail, and air-gap support built in.
×Solution BriefProof & rollout02 / 02
04Where teams start
Three use cases, one platform.
Plain-language answers, not dashboards-on-request.
Ask in plain English and get a governed answer with the SQL behind it. No ticket queue, no dashboard build.
Retail, e-commerce, and martechAI analytics that never leaves your perimeter.
Private models on Phison hardware, row-level access, and a full audit trail. Zero bytes leave the network.
Healthcare, finance, government, and manufacturingAnswers in Slack, Teams, and MCP agents.
One governed definition of every metric, served to people and agents alike. Nothing rebuilt per channel.
Ops, RevOps, and frontline teams05What powers every workflow
The agent layer behind every use case.
Tools, skills, and memory run on your hardware, inside your governance boundary.
More than a chatbot.
Agents reason over many steps in a sandbox: query data, build charts and dashboards, read PDFs, and save skills.
Reusable workflows. Compounding context.
Skills are your agent's standard procedures, so output stays consistent. Memory learns from past work and turns daily use into company knowledge.
Native Git and file system.
Skills, memory, and models are all files: versioned, branchable, and reviewable. Agents read and write them directly.
Custom dashboards from a single prompt.
Describe what you need and get a polished, governed dashboard in one click.
06Seven-day rolloutA deployment, not a demo script
01Day 1 morningConnect your database
Postgres, MySQL, ClickHouse, Snowflake, Redshift, DuckDB, plus 20+ more.
02Day 1 afternoonAuto-MDL builds your model
Wren reads your schema; data teams review joins, metrics, and access.
03Day 2–3Seed real business questions
Capture what finance, ops, and leadership actually ask.
04Day 3–5Tune with the business
Refine the context layer until accuracy clears the bar.
05Day 5–6Pilot with users
Turn on governance and validate with a live group.
06Day 7Roll out and connect agents
Go live for people and AI agents on one governed interface.
Bring your database. Keep it in the building.
See a private model answer questions on your own schema, fully on-prem, in under 10 minutes.