On-Premises Enterprise AI
Enterprise AI that never leaves your company —it runs on your premises, not in the cloud
Enterprise AI that runs on your own infrastructure and works with your company’s documents. Sensitive information never leaves your network, AI costs are predictable, and if a more advanced model becomes available later, the system can be upgraded to use it.
The data remains within the company
We store and process it behind a firewall and in an encrypted format
Predictable cost
it is not determined by the number of tokens used
Interchangeable AI model
open license, no vendor lock-in
Enterprise-level support
Implementation, customization, and SLA-based operation
✎ New conversation
Search
Notes
Work Area
Models
● Qwen3 Coder (local)
● Llama 4 Scout (local)
● Gemma 4 (local)
● Frontier · as per the rules
Insights
Policies
Financial Reports
Project Documentation
Conversations
Quarterly revenue…
Timesheet Reconciliation
Excerpt from the HR Policy Manual
Quarterly Revenue by Region
•••
Import the arbevetel_2026.xlsx file and plot the quarterly trend by region.
Qwen3 Coder · local
Revenue in 2026 by Region
5,337 million HUF
▲ 46% Q1→Q4
EMEA
1 927
▲ 37%
AMER
1 801
▲ 35%
APAC
1 081
▲ 71%
LATAM
528
▲ 80%
Region
Q1
Q2
Q3
Q4
Year
EMEA
412
448
501
566
1 927
AMER
388
421
470
522
1 801
EMEA and AMER account for the majority, but APAC is growing the fastest (+71% Q1→Q4). The file did not leave the network.
Write a message…
Qwen3 Coder · local ⌄
Why now?
One foot on the gas, the other on the brake
Management wants operations to be natively AI-driven: a colleague asks a question, the agent does the work, and the report is generated. At the same time, management is also afraid of this—because data leaves the system, and the cost is unpredictable. Both of these feelings are true at the same time, and they counteract each other.
Cost
The cost of AI is rising dramatically
The unit price fell by 93% over two years—yet corporate bills more than tripled. When something becomes cheaper, we use more of it. Consumption-based pricing and agent-based tasks (5–30 times more tokens) further accelerate this trend.
unit price, $ per million tokens
Enterprise GenAI Spending, $ billion
Source: Ramp (unit price), Menlo Ventures (funding)
93%
It exceeds the budget for corporate AI teams
McKinsey FinOps, May 2026
Risk
The data goes out, but the responsibility remains
With every cloud-based AI query, corporate information leaves the company: contracts, customer data, financial spreadsheets. The responsibility for data security and compliance, however, remains with the user’s EU-based company—regardless of what the service provider promises.
Safety
Whatever the company doesn't allow, employees handle through their personal accounts. This is "shadow AI": you can't see it, you can't log it, and you can't trace it back. Today, this is behind every other data incident.
Compliance
GDPR, NIS2, and—starting in December 2027—the EU AI Act. During an audit, you must demonstrate where the data is located, who has access to it, and what has been logged. The promise that “we don’t train our models on your data” is not a sufficient response to this.
Addiction
Someone else replaces the model, and someone else shuts it down. It doesn’t speak, it doesn’t ask questions, and you can’t retest it—yet your processes are built around it.
Managers aren't afraid of AI. They're afraid of not knowing where the data is, who can see it, and how much it will cost next month.
The question isn't whether to hit the gas or the brakes—it's where the engine should be revving
If the model runs on-premises, no data leaves the premises, and the cost is a one-time, known expense. You can keep the gas and let off the brakes—that’s what Gloster Local AI does.
The solution
You're in control: decide which data to process locally and which to process using cloud-based AI
You don’t have to pay for an expensive cloud-based model for every task. The system automatically determines what can run locally and when it’s appropriate to call on external AI—the decision isn’t made by the user via a drop-down menu, but rather by the rules you’ve set.
Instead of using ChatGPT, Claude, and Gemini models, the local model handles smaller tasks.
Request
My colleague asks
AI Route Selector
decides where the request should be routed
confidentiality · cost · complexity · own rules
Firewall
Local
recurring, sensitive — stays in
DGX Spark · open model
Frontier
The hard work—only if you let it
external model · after DLP filtering
1 · Request
An employee asks
2 · AI Route Selector
The system decides
3A · Local
Recurring and sensitive
3B · Frontier
An exceptionally complex task
15 Key Features
What Turns a Chat into an Enterprise AI System
The colleague works in a standard chat interface; the company manages the models, knowledge, and access.
01 Usage
Chat, models, AI running on the user's own computer, and task-based route selection.
01
Knowledge
Assistant
Template
Law
shared corporate context
Company Background
A shared corporate context underlies every conversation
Knowledge bases, assistants, templates, and permissions all in one place: every colleague works from the same corporate knowledge base, and that knowledge isn't scattered.
02
request
→
local model
request
⇢
external API
It depends on the task
Model Switch
For each task, we choose from among several models
The choice of model may depend on the task at hand and company policies; adding a new model does not require a new employee interface.
03
Corporate Network
DGX Spark · local model
⇢
✕ cloud
You don't have to go out
In-House AI
The AI runs on the company's own DGX Spark machine
With the on-premises model, model calls do not need to go through an external AI provider. The company itself determines where the model runs and which data paths are allowed.
04
request
→
rule
→
Local
External API
Smart Router
The system decides which model to use
Expensive modeling capacity can be reserved for tasks that truly warrant it, while the data flow can be controlled.
02 Corporate Knowledge
Document search, up-to-date resources, expert assistants, and collaborative work methods.
05
"What expenses can I claim when traveling?"
→
Source: Section 4.2
Knowledge Answers
AI provides answers based on company documents and cites the source
Search by report in company documents; the source is listed next to the result, for example: Travel Policy, Section 4.2.
06
Rules v2
→
↻
→
+ new
~ amended
− deleted
Knowledge Sync
Any changes are automatically incorporated into the company's knowledge base
The response is always based on the most recent version of the document. New files are added, modified files are updated, and deleted files are removed.
07
Sales
HR
Finance
Back Office
shared knowledge base
own instructions · tool · model
Custom Assistants
Each area will have its own agent, built on the same corporate knowledge base
Sales, HR, finance, back office: each has its own instructions, tools, and model, tailored specifically to its own team.
08
Skill: Offer
✓ —————
✓ ————
✓ —————
→
the entire company achieves
Skill Library
Once we record something, the entire company uses it
If someone comes up with a method that works well, the entire company will start using it the very next day, working faster while maintaining the same level of quality.
03 Control
Integrations, enterprise onboarding, permissions, data management policies, metrics, and automation.
09
ERP
SQL
CRM
Local AI
MCP · OpenAPI
Docs
Jira
API
System Connect
The AI integrates with the company's existing systems
MCP, OpenAPI, and proprietary tools: AI can request data from enterprise systems and, upon approval, initiate an action.
10
colleague
→
Corporate Login
→
AI
No new password, no new account
Identity Link
The colleague logs in using their existing Microsoft or Google account
You can access the system using the same corporate login credentials you use for all other corporate systems. The departing employee's access will be revoked.
11
HR-AI
HR Knowledge
Sales-AI
Sales Skills
HR
✓
✓
✕
✕
Sales
✕
✕
✓
✓
Admin
✓
✓
✓
✓
Access Guard
Everyone sees only the AI resources allocated to them
Each employee has access only to the models, knowledge base, tools, and skills appropriate for their role.
12
→
✓
Data Guard
→
Local
✕ External
Data Guard
Confidential data is not shared with external AI models
Truly internal, confidential information does not make its way into the Frontier models: it recognizes such information in text, documents, and images.
13
Monthly AI Cost
412,000 Ft
local external
sample data
Usage Insights
You can see who is using AI and how much it costs
By user and by group, broken down separately by local and external models. AI costs will be brought under the company’s control.
14
a question
A
★★★☆☆
B
★★★★★
C
★★★☆☆
blind evaluation → ranking
Model Arena
Several models answer a question, and it becomes clear which one is better
We're getting better and better answers because the models running on the company's knowledge base are becoming increasingly well-tested.
15
Monday
8:00 a.m.
↻ every week
start
collect
summarize
person checks
share
Task Scheduler
The routine task runs automatically at the specified time
The weekly summary is generated automatically on the company's own server and is shared only after human approval.
The figures are schematic; the example data are for illustrative purposes only.
Security and Control
It's not a new security paradigm— it falls under your existing controls
Directory, role, quota, logging. AI traffic passes through the same gateway and can be monitored using the same tools as your other enterprise systems.
The Journey of a Question — Four Checks, One for Each Request
01
Only those you've approved can ask
Sign in using your existing company directory. You decide, based on role and group, which users can access which models, knowledge bases, and assistants.
02
You decide what can leave the company
DLP and routing rules follow your specifications. Confidential information remains secure, even if the user selects an external model.
03
You decide where the task takes place
By default, on your own computer, within the company network. Requests to external models are only allowed if your rules explicitly permit them.
04
Every step can be traced
All questions and answers are logged and integrated into the existing SIEM. In the event of an audit or incident, you can see exactly who asked what and what they received.
Behind a firewall
closed VLAN, no direct Internet access
Encrypted storage
both on the album and during the recording process
SSO and Directory
using the existing corporate ID
Audit Log
Sent to SIEM, with searchability
Backup and Failover
Second tool: saved model weights
Let’s be clear: just because it runs locally doesn’t automatically mean it’s compliant. Local operation simplifies the management of data handling and compliance issues, but the work still needs to be done—we’ll set up the authorization and data management layer at your site.
Technological Independence
Corporate knowledge is not tied to any single AI model
If a more advanced model becomes available, it can be replaced without having to rebuild your documents, permissions, integrations, and workflows. The value of the implementation lies in this layer, and this layer remains with you.
Third-Party Service Providers
No vendor lock-in
✕
Corporate Knowledge
This layer stays with you
documents
permissions
integrations
workflows
→
Replaceable
Our own internal model
You aren't dependent on large external service providers
Open-Licensed Models
Open-Licensed Models
Our tests were run on Qwen3.8-27B, licensed under Apache 2.0. But it can run anything that fits within the 128 GB—and even more than that with compression.
You have the model weights
Saved and versioned along with the configuration. A model can't just disappear overnight because the service provider discontinued it.
The transition is scheduled
It doesn't update automatically. First, we'll check on the secondary machine to see how it handles previous tasks, and only then will it go live.
Cloud-based models can also be used
The system does not exclude Frontier models: where it makes business sense and company policies allow it, they are also accessible through the route selector.
Tested in-house
Our in-house developers refactored an actual client project
Egy legacy kódbázis refaktorálását futtattuk a Gloster Local AI-on, éles kódon. Az alábbi ábra végigviszi a folyamatot a bemenettől az új rendszerig.
Software Development · Live Client Project
Legacy migráció valós ügyfélprojekten, egy dobozon
Az ügyfél kódjából, dokumentációjából és tesztrendszeréből új, modern rendszer készült, nem csak refaktorált kód. Egy DGX Sparkon 4–5 fejlesztő dolgozhat párhuzamosan.
Input
Ügyfél kódja
legacy kódbázis
Dokumentációk
specifikációk, leírások
Tesztrendszer
hozzáférés, futtatás, ellenőrzés
One box, on-site
DGX Spark
Qwen3.8-27B
4–5 fejlesztő
párhuzamosan, egy eszközön
Output
Új, modern rendszer
nem csak refaktor · a kód helyben marad
One box, multiple developers. A single DGX Spark can accommodate 4–5 developers at once.
Új rendszer, nem foltozás. A régi kódból modern rendszer készül.
Sonnet Level 5 encoding. Using the Qwen3.8-27B open model, locally.
„Legacy projekt migráción próbáltuk. Részletes projekttervhez jobb az Opus, de a mindennapi kódolási feladatokhoz ez is elég.”
Internal Gloster measurements based on our own data, run on a DGX Spark using the Qwen3.8-27B model. The implementation was completed by two developers in approximately 10 engineer-days. This is not a general performance guarantee.
Gloster's Services
Customized Services for Seamless Operations
A hardvert és a nyílt modellt bárki megveheti. A Gloster beállítja a helyi környezetre, és utána üzemelteti. Két csomagból indulhat.
Eszköz + alapbeállítás
Local AI Core
Az eszköz, üzembe helyezve és alapszinten beállítva. Ezzel már az első naptól dolgozni lehet.
Eszköz: alapesetben 2 DGX Spark + 1 szerver
Installation and Commissioning
Alapbeállítások
céges branding
modellek
smart router
company context (RAG)
skill library
promptsablonok
Identity Link
Active Directory
hozzáférés-kezelés
token-alapú user insights
Core + oktatás + testreszabás
Local AI Advance
Minden, ami a Core-ban van, és a cég saját folyamataira hangolva.
A teljes Core csomag
Felhasználói oktatás: általános használat és a saját use case-ekre szabott képzés
Egyedi testreszabás az ügyfél igényei szerint
Például
custom asszisztensek
custom konnektorok
részletesebb monitoring
erősebb security-beállítások
Mi része, és mi kérhető egyedileg
Introduction
Installation and Commissioning
Része
RAG és tudásbázis
limitált dokumentumszámig · bővítés egyedi igény szerint
Része
Routing
Része
Identity Link + Active Directory
Része
Custom asszisztensek
2 asszisztens (Back Office és Coding) · további egyedi igény szerint
Része
DLP
mélyebb, bővebb beállítás egyedi igény szerint
Alapszinten
Task scheduler
komolyabb megoldás egyedi igény szerint
Alapszinten
Üzemeltetés és support
Hibajegy-alapú üzemeltetés
Része
SW support, SLA
Része
Controlled Model Update
Része
Regressziós tesztek
modellcserekor, igény szerint
Része
A hardver-support nem része a szolgáltatásnak.
Csak egyedi igény szerint
egyedi konnektorok
természetes nyelvű adatbázis-lekérdezés
magas rendelkezésre állás (HA)
költségalapú usage insights
audit log
backup
Business Model
Vásárolható vagy havidíjas: CAPEX és OPEX is
Ugyanaz a rendszer, kétféle finanszírozással. Ha nem akartok induló beruházást, havidíjjal is elérhető.
CAPEX · Egyszeri beruházás
Own equipment, with full service
You purchase the device, and Gloster handles the customization and operation.
Tool
Yours, from Gloucester
Customization
Gloster
Operations
Gloster
Egyszeri beruházás + support díj
OPEX · Havidíj
Gloster Local AI Managed
Nincs induló beruházás. Egy havidíj, amiben az eszköz, a testreszabás és a support is benne van.
Tool
Gloster guarantees
Customization
Gloster
Operations
Gloster
Induló költség nélkül, supporttal, minimum 1 éves szerződéssel
Both packages include Gloster's customization and operational services.
Start
In four steps, starting small
01
Selecting a Use Case
We select the process that yields the greatest profit.
a well-defined process
measurable goal
02
PoC in small test projects
We'll use our own data to prove that it really works.
firewall, VPN
proprietary hardware
the data remains
03
Expanding Integrations
We'll connect it to the existing systems and configure the security settings.
SSO
DLP
routing rules
audit log
04
Scaling and Operations
Additional use cases, continuous operation.
additional use cases
regressziós tesztek
controlled model update
SLA
Frequently Asked Questions
Amit a döntés előtt meg szoktak kérdezni
The person standing behind him
A bevezetés mögött tőzsdén jegyzett IT-csoport áll
An on-premises AI environment is a long-term operational commitment. It matters who takes it on—and what kind of vendor support they have.
Hardware
NVIDIA
The DGX Spark platform and the manufacturer's hardware support.
Enterprise Platform
Microsoft
Directory, SharePoint, Business Central — integrated with your existing environment.
Frontier model
Anthropic
The outer side of the hybrid route, if the rule allows it.
Networking and Security
Cloudflare
Edge network and access protection layer.
Technology partnerships. Gloster has more than 20 years of experience in corporate IT operations.
8,65
Revenue in billion Ft (2025)
91% recurring revenue
751
M Ft EBITDA (2025)
52% exports
200+
colleague
HU · DE · UK
GLSTR
A group listed on the BSE
Budapest Stock Exchange, Xtend
Source: Gloster Digital Group Plc., 2025 Annual Report, April 2026.
Let's move on
Which process should we use to demonstrate what our company's AI entails?
Tell us where your organization is losing time, and together we'll see how Gloster Local AI fits into your operations. We'll bring our own metrics and real-world use cases to the demo.
Gloster Digital Group Plc. · Park Atrium, Budapest
BÉT: GLSTR · Microsoft • Anthropic • NVIDIA • Cloudflare partner
I'd like a demonstration
Fields marked with an asterisk are required.