Blog | July 13, 2026

Bringing Snowflake power to Private AI: A hybrid architecture for privacy, compliance and cost efficiency

By Pawel Kowalski, Solution Architect, DXC's Data Driven Development Practice, and Piotr Frejowski, AI Solutions Architect and Product Owner of DXC Private AI

Snowflake is a leader in data analytics and cloud scalability, but for many organizations — such those in the banking, defense and healthcare Industries — moving sensitive data or proprietary AI models to the cloud is not an option. Regulatory constraints, privacy concerns and cost considerations often block full adoption.

To address these challenges, we designed a hybrid architecture that enables Snowflake to orchestrate analytics and AI workflows using on-premises resources. The solution ensures secure, reliable access to local AI models and storage, while maintaining compliance and minimizing cloud costs.

We presented this solution at the DXC booth during the Snowflake World Tour 2025 in Berlin. Our solution uses NVIDIA Jetson Nano as your edge device, with MinIO as Object Store and Ollama as the LLM server installed. You'll need to make sure your device has a public IP address and a domain name.

To ensure maximum flexibility, we designed the solution to operate from virtually any location with power and internet access. Our primary challenge was making the Jetson device reliably accessible to Snowflake, regardless of network environment.

We addressed this by deploying a lightweight virtual machine (VM) in the cloud with a static public IP address and domain. This VM acts as your secure gateway, utilizing a VPN and port forwarding to bridge external requests to the Jetson device.

Due to certain limitations in the Jetson Jetpack software and its kernel, we incorporated a compact router into the setup. This router connects to any Wi-Fi or Ethernet networks available to you and automatically establishes a VPN tunnel to the cloud VM. Port forwarding is configured to securely expose Jetson services as needed.

This architecture ensures that Snowflake can interact with the Jetson device securely and reliably, no matter where the hardware is deployed.

There are a few things you'll need to configure from Snowflake's perspective.

LLMs on Private AI

Snowflake calls REST APIs that interact with locally (Private AI) deployed LLMs. This means Snowflake queries can invoke AI models without sending data outside the organization — and without consuming Snowflake tokens for inference.

Technically, you'll need to use Snowflake Function with External Access Integration configured. Remember that Ollama by default works over http; Snowflake requires secure connection over https. To achieve that, we used an NGINX server with Reverse Proxy.

MinIO as S3-compatible storage

Snowflake connects to MinIO via an external stage using s3compat://. This allows secure file exchange while keeping all data on-prem.

Technically, make sure that:

• Your MinIO is:
  • configured to use virtual-hosted-style (virtual hosting of buckets)
  • compliant with S3 API and passes all tests from a public test suite
  • configured to use direct credentials
  • does not contain port number
• Your endpoint is accessible over https with a valid TLS certificate
• You create a Snowflake Support ticket and request to whitelist your endpoint in your account. In our case it took two days to complete.

Having accomplished all the above, your external stage creation is as simple as:

create or replace stage minio_on_private_ai

url = 's3compat://my_bucket_on_minio/'

endpoint = 'your-private-ai.enpoint.com'

credentials =

(

aws_key_id = 'this_is_my_key' aws_secret_key = 'this_is_my_secret'

) ;

The following business benefits are direct results of implementing the Snowflake + Private AI hybrid architecture. This solution enables organizations to achieve their goals for security, cost efficiency and operational flexibility — without compromising on innovation.

• Regulatory compliance and privacy: Sensitive data and models stay on-premises, meeting strict industry requirements.
• Cost efficiency: Local AI inference reduces cloud costs and maximizes hardware investments. No Snowflake token consumption for AI inference when you leverage your existing GPU hardware. This ensures full control over data and models while reducing cloud costs.
• Operational flexibility: Deploy anywhere with internet access, supporting hybrid and mobile use cases.
• Maximized ROI & hybrid flexibility: Combine Snowflake's power with your on-prem resources to extend Snowflake's value without new hardware investments.
• Innovation without compromise: Access advanced analytics and AI securely and efficiently.

• Private AI for the Enterprise
• Private Cloud+ Dedicated: Single-Tenant Cloud Services
• Private Cloud+ Government: Cloud Services for the Public Sector
• DXC and Snowflake