Private AI Cloud: Innovation in the Compliance Zone

Translated from the original article which appeared in IT Finanzagazin on 28 July, 2026

How can financial service providers use a private AI cloud to move from pure AI consumer to producer of their own intelligence? That move is a strategic decision, caught between dependence on the large AI providers and self-determination. But no single vendor delivers the full stack alone, which is why HPE, NVIDIA, deepset, QuantPi and secunet are pooling their capabilities into a shared platform rather than competing as individual vendors.

by Dennis Geisse, AI Enterprise Architect, HPE

‍

With a private AI cloud, financial services providers become producers of intelligence on their own terms. It’s an architectural decision that eases the balancing act between development speed and regulation. And it impacts the tokenomics more than many would expect.

‍

Suppose a bank wants to speed up its lending process. An AI application is designed to read balance sheets, connect risk data, check regulatory requirements and credit memos. Steps that cost hours before should now take only a few seconds.

‍

But the project gets stuck in pilot. An internal review demands complete auditability across the entire application lifecycle. Compliance asks about compounding risk in the supply chain of third-party AI providers. Risk control wants to know how the compounding risk from hallucinations can be contained when several AI agents build on one another.

Where potential profit meets regulation

This tension affects the whole financial sector. BCG sees 370 billion USD in additional annual profit potential from AI in retail banking alone.

‍

“Every AI application meets a regulatory framework that is getting tighter.”

‍

In December 2025, BaFin clarified that AI systems count as regular ICT assets, and belong in the risk management framework under DORA, with complete monitoring from data sourcing through to decommissioning. In addition, the EU AI Act classifies applications such as credit scoring and creditworthiness assessment as high risk.

‍

Then comes supplier risk, which became even clearer when Anthropic blocked two of its strongest models worldwide in June 2026.

A middle path between public cloud and building your own

Anyone who builds the AI stack themselves gains control, but pays for it in complexity, headcount and time. At Berlin’s Rise of AI Conference in May, we demonstrated a private AI cloud platform together with deepset and NVIDIA that addresses exactly this point. Private cloud, briefly defined: it can be used like a cloud but runs in your own data centre.

[Image: Dr Anna Gründler, Dennis Geisse and Dr Robert Friedrich on stage at the Rise of AI conference.] (Image source: Rise of AI, Dominik Tryba)

‍

The platform is highly integrated and delivered preconfigured, which necessarily limits freedom of hardware configuration. In return, you reach auditable operation in weeks, rather than months or years. And, decisively, it is designed as an open foundation. Models and applications can be swapped out without touching the platform.

Preconfigured foundation with open interfaces

We developed the platform, called HPE Private Cloud AI, together with NVIDIA. It bundles compute, storage, inference software and ready-made application blueprints into a ready-to-run environment. Specialised components can be added for specific requirements.

‍

Haystack, by deepset, handles the orchestration of the agents with fine-grained permissions and an audit trail. QuantPi tests the behaviour of the agents, monitors it in operation and shows early on when errors start to compound along a chain. Finally, the security layer of secunet’s SINA Cloud provides separation of tenants at the hardware level.

‍

What this looks like in practice can be demonstrated by the secure, decision support application “AIDA” (Artificial Intelligence Decision Agents). It consists of three, specialised agents orchestrated in Haystack that access locally running models. Through a simple interface, the system guides a credit analyst through balance sheets, internal risk data and regulatory requirements in seconds.

‍

Tokenomics in practice

That leaves the economics. At my own company, HPE, we run an agent-based customer support system. It initially ran in the public cloud, but the problem was that token costs rose to an unacceptable degree ­­­– mainly, because agents consume tokens with every observation, conclusion and action. We moved the system to the private AI cloud platform and cut costs by a factor of 30, a saving of around 100k USD per month. At the same time, customer data stays under our control and response times have improved.

‍

We stopped being AI consumers and started becoming producers of intelligence ourselves. At this point, the balancing act described at the start has disappeared. Sovereignty, auditability, economics and better customer experience are no longer competing goals. They follow from the same architectural decision.

Author: Dennis Geisse, HPE/dk

Dennis Geisse is an AI Enterprise Architect at Hewlett Packard Enterprise (HPE). He helps banks, insurance companies, and government agencies build robust AI platforms—from reference architecture and model selection to production operations. His areas of expertise include agent-based AI in regulated environments, data integration, and on-premises inference infrastructures.

‍