Data sovereignty
Training data, your documents, prompts and responses are processed on a network dedicated to you in Trabzon. No query is sent to a third-party model.
Powerful Infrastructure. Limitless Possibilities.
GPU & AI · END-TO-END PLATFORM
For AI applications that run on your own data, we build every layer under one roof, from data to model and from model to end user. Your private LLM, RAG pipeline and MLOps tools run in our Trabzon data center; your training data never leaves the country.
WHY YOUR OWN AI INFRASTRUCTURE?
General-purpose AI services give you a quick start. But in lasting applications that work with internal documents, customer records or production data, data location, cost and control become decisive.
Training data, your documents, prompts and responses are processed on a network dedicated to you in Trabzon. No query is sent to a third-party model.
As usage grows, the per-token bill grows with it. With dedicated GPU capacity the cost is fixed; as your user count grows, the cost per request falls.
You choose the model, pin its version and fine-tune it on your own data. A provider's model update will not change your application's behavior overnight.
The model runs right next to your data and application. With our backbone reaching your users in Istanbul and Ankara within milliseconds, responses flow quickly.
REFERENCE ARCHITECTURE
Pick a scenario and see which layers come into play. Click a layer to explore its components, the role of NuageDC and typical sizing.
A private language model that searches your internal documents and answers with source citations. The vector database and the inference layer are at the center.
LAYER 01
All the data that feeds AI: file servers, SharePoint and wiki pages, ERP/CRM databases, email archives, camera footage and production-line sensors. Data is transferred to Trabzon over a private line or encrypted VPN.
COMPONENTS
IN THIS SCENARIO
Text sources such as HR policies, procedures and technical documents come to the fore; each document's access permissions are transferred as well.
PLATFORM COMPONENTS
We build the platform on widely used open-source tools. That way your teams work with tools they already know and your workloads stay portable. For GPU hardware options, see the GPU Cloud page.
From single-GPU development environments to 8-GPU nodes and multi-node clusters. On Kubernetes, drivers, CUDA and monitoring components are managed automatically with the GPU Operator.
Local NVMe scratch during training, a file system shared across teams, and S3-compatible object storage for the data lake. Checkpoints are backed up automatically.
100G network between nodes and isolated VLANs. For large multi-node model training, the RDMA (InfiniBand or RoCE) link is designed per project.
A ready notebook environment for your data scientists, experiment tracking, a model registry and reproducible pipelines from training to release.
High-throughput LLM servers, an OpenAI-compatible API endpoint and automatic scaling as load grows. You can roll over gradually between model versions.
A pipeline that chunks your documents, vectorizes them and indexes them with permission information. Users get answers only from documents they have access rights to.
About model licenses: Each open-weight model family, such as Llama, Mistral, Qwen and Gemma, has its own license terms; these may include restrictions on commercial use, number of users or use of outputs. We choose the model by evaluating these terms together with you at the start of the project. Product and brand names belong to their owners.
PRIVATE LLM SIZING
Select your user count, request volume and model; see the GPU configuration you will need and your current token-based API spend. Our team will prepare a quote tailored to your configuration.
The number of employees or customers who will use the assistant regularly.
The calculation assumes about 25% of tokens are output (the generated answer) and the rest input (question + RAG context).
Model size
Precision
Response time target
How is it calculated? For requests spread over 10 business hours a day, peak load is taken as 2.5 times the average. Required memory = model weights + KV cache of concurrent requests; required speed = tokens/s to be generated at peak. For each GPU type, the smallest configuration that meets both conditions is found and the most efficient one is recommended. Results are indicative only.
Recommended configuration
2× NVIDIA L40S
Meets the memory and speed target with the fewest resources.
| GPU | Quantity | VRAM |
|---|
Total GPU memory
96 GB
Token-based API / month (your assumption)
₺52.800
Pricing for your configuration is provided as a quote.
—
The recommendation is a preliminary idea; exact sizing is settled with PoC measurements. API spend is calculated only from the assumption you enter and does not represent any provider's pricing.
HOW DOES RAG WORK?
RAG (Retrieval-Augmented Generation) makes the model search only your documents before generating a response. Follow it step by step through the HR policy of a fictional company.
STEP 1 / 7
NuageDC · TrabzonHR policies, the benefits guide and internal wiki pages on the file server are read. The information on which groups each document is open to is ingested with them.
source/HR-Policy-2026.pdf 42 pages access: hr-all source/Benefits-Guide.docx 31 pages access: hr-all wiki/hr/faq 45 pages access: all-employees ───────────────────────────────────────────────────── 3 documents · 118 pages read
The question is asked in step 5; the answer arrives in step 7 together with its source.
This is a fictional sample company policy; it does not belong to any real organization.
SERVICE MODELS
If you have a strong data science team, take delivery of the infrastructure; if you want to reduce the operational load, we can manage the platform; if this is your first project, let us go from idea to production together.
GPU cluster + Kubernetes, managed by you
GPU nodes, storage and network are delivered ready; Kubernetes and the GPU Operator come installed. Your own team installs and operates the tools and models.
Starting from ₺12,260 per GPU per month
View GPU OptionsMLOps tools installed and monitored
JupyterHub, MLflow, the vector database and the inference server are installed and updated by our team and monitored 24/7. Your team focuses on the model and the application.
GPU capacity used + platform management fee · by quote
Get a QuoteFrom discovery to production, together
Use-case selection, data preparation, model and RAG design, evaluation (eval) and go-live; carried out project by project together with our solution partners.
Project-based pricing according to scope
Schedule a Discovery CallUSE CASES
The projects that organizations working with sensitive data most often start with; in each of them, data location and access control are at the center of the design.
An assistant that answers with source citations across legislation, regulations and internal correspondence; data stays in the country and access is restricted per unit.
Approach: Private LLM + RAG
Fast answers for branch staff across procedure, product and internal control documents; every answer comes with its source and all queries are in the audit log.
Approach: RAG + access control
Pre-assessment and prioritization models for medical images; systems that support the physician's decision and are trained on anonymized data.
Approach: Vision · fine-tune
Defect detection in images from line cameras; the model is trained in Trabzon, runs in the field in a lightweight version and is updated regularly with new samples.
Approach: Vision · MLOps
Transcription and summarization of calls, plus topic/sentiment classification; voice-of-customer data is processed without leaving your organization.
Approach: Speech recognition + LLM
Personalized product recommendations from behavioral data, smart search and demand forecasting; auto-scaling inference during campaign periods.
Approach: Classical ML + vector search
PROJECT PROCESS
Each stage is evaluated against measurable success criteria before moving on to the next. This way the investment grows in step with proven value.
Scenario selection, data inventory and success criteria; a quick proof of concept with limited data.
2–4 weeksTrial with a real user group, collecting feedback, measuring quality and cost.
4–8 weeksGo-live with scalable inference, authorization integration (AD/SSO), monitoring and a redundant architecture.
4–6 weeks24/7 monitoring, model and data updates, quality evaluation and capacity planning.
ongoingSECURITY, KVKK & RESPONSIBLE AI
What matters is not only what the model knows but also who can access what. We design the platform around the principles of data minimization, access control and traceability.
The platform runs on a VLAN dedicated to you; model endpoints are not exposed to the internet and are reachable only over VPN or a private line.
In RAG search, documents the user is not authorized to see are never retrieved; authentication integrates with your corporate directory (AD/SSO).
Model, data and administration operations are logged; logs can be forwarded to your SIEM system and retention periods are set according to your policy.
Personal data not needed for training and indexing is filtered out or masked; anonymization steps can be added to the pipeline.
Which version of which model is used with which data and under which license is kept traceable in the model registry.
Response quality is measured regularly with test sets; for uses that affect critical decisions, human approval is included in the process.
FREQUENTLY ASKED QUESTIONS
We have answered the most frequently asked questions about private LLMs, RAG, data location and sizing.
AI solutions team
0212 951 00 61GPU Cloud is the rental of GPU virtual servers or bare-metal servers; you install and manage the software on top. AI & ML Infrastructure, on the other hand, is an end-to-end platform covering data storage, a vector database, a training cluster, a model registry, an inference layer, monitoring and security, and it can be managed by our team if you wish.
No. Training data, the vector database, model weights and inference traffic are processed in our Trabzon data center on an isolated network dedicated to you. No data is sent to an external AI API; connections abroad are used only in cases that you explicitly define.
Open-weight models such as the Llama, Mistral, Qwen and Gemma families, as well as models you train yourself, can be run. Each model has different license terms; we evaluate restrictions on commercial use, number of users or use of outputs together with you at the start of the project.
RAG means the model searches your own documents and uses the relevant sections before generating a response. It is usually the first choice when information is updated frequently and sources are expected in the answers. Fine-tuning, on the other hand, is used to change the model's style, format or domain-specific skill; the two methods can also be applied together. You can follow the flow step by step in the RAG demo.
Model size, precision (FP16, FP8, INT4), the number of concurrent users, context length and the response-time target are evaluated together. The estimation tool gives a first idea; exact sizing is settled with real values measured during a PoC on your own data.
For multi-node training of large models, an RDMA link between servers (InfiniBand or RoCE) is designed per project. Workloads of up to 8 GPUs on a single node and inference clusters run on our standard 100G network infrastructure. For high power density deployments, see the Space Rental options.
It depends on the service model you choose. In the Infrastructure model, Kubernetes and the GPU Operator are delivered ready and you install the tools. In the Managed platform model, JupyterHub, MLflow, the vector database and the inference server are installed, updated and monitored by our team.
The first discovery call is free of charge. Then a time-limited PoC environment is prepared, with scope and success criteria defined for your use case. PoC terms and resource usage are stated clearly in the quote; after the PoC, you move on to the pilot and production stages.
GET A QUOTE
Briefly share your needs; our solution architects will review your current infrastructure and get back to you with a recommendation that balances performance, security and budget.
RELATED SERVICES
Virtual and bare-metal servers with NVIDIA GPUs; hourly or monthly.
Learn MoreA dedicated resource pool that runs alongside your AI applications.
Learn MoreHigh-density cages and private suites for your own GPU cluster.
Learn MorePrivate line and high-speed connectivity from your data sources to Trabzon.
Learn More