A review of the lecture by Christoph Streit (ScaleUp Technologies) at the SCS Summit 2026 by Lisa Seifert

The AI revolution is taking place largely within ‘black boxes’ – on the servers of American hyperscalers, behind proprietary APIs, and under jurisdictions that fundamentally contradict European data protection standards. At the SCS Summit in May 2026 said Christoph Streit, CEO of ScaleUp Technologies, has impressively demonstrated that there is another way. One that does not require any compromise between performance and digital sovereignty.

Who is behind it?

ScaleUp Technologies is not a fledgling start-up riding the wave of current trends. Christoph Streit founded the company in Hamburg in the late 1990s together with his best school friend – back then still using a bit of Windows, but with a clear focus on open source from the early 2000s onwards at the latest. Today, as a managed hosting, cloud and colocation provider, ScaleUp operates OpenStack-based cloud infrastructures, including managed Kubernetes, with teams in Hamburg, Berlin and Nuremberg.

Since 2020, the topic of „sustainable digitalisation“ has also been added to the agenda – and, for some time now, the Open Compute Project (OCP), in which ScaleUp is active as a solution provider. This is precisely what a key part of the presentation will focus on.

The problem: the same trap, only deeper

For most people in the SCS community, the dependence on US hyperscalers and big tech companies is nothing new. Proprietary APIs, opaque software stacks, the US Cloud Act hanging over them like the sword of Damocles – these are all well known.

But when it comes to AI, according to Streit, the problem becomes considerably more acute. With a large language model, you’re not simply sharing the computational load – you’re sharing data. And potentially very sensitive data at that. Streit sums it up: an employee who asks the chatbot a question could inadvertently reveal a password or disclose confidential business information. With an external provider, this is virtually impossible to control.

In his view, the sovereignty gap in AI is therefore even more serious than in the cloud sector in general.

The basis: openness right down to the firmware

Streit begins his proposed solution in a place that many find surprising: hardware. And in doing so, he draws attention to a problem that is often overlooked in the debate.

„One of the biggest problems these days isn’t really at the operating system level, or the hypervisor level, or the cloud stack level, but actually the firmware running on every server in the world. That’s still 99 per cent closed source.“

This is where the Open Compute Project (OCP) comes in – a project that originally arose from a simple realisation at Facebook: the way servers are built is completely inefficient. So, over 15 years ago, Facebook didn’t just set about building better servers internally; it shared its ideas and blueprints with the community and turned them into an open-source project.

The result: servers without power supplies (power is supplied centrally via direct current through the cabinet), with no power cables attached to the chassis, fewer fans, and no casing or logo. Anyone who has ever seen a conventional server cabinet will immediately understand the logic behind this: Instead of 80 individual power supplies in a 40-U rack, the OCP standard requires only six to twelve. And anyone with a bit of knowledge about power supply efficiency will know what that means – the optimal load point of a power supply is rarely reached in traditional servers.

Added to this is OpenBMC as an open firmware standard – which means that even the hardware itself is no longer a ‘black box’, and backdoors at chip level are ruled out.

Another advantage highlighted by Streit is that all major GPU infrastructures from Nvidia and AMD are now based on OCP rack and server infrastructure. This makes OCP not only a standard for sovereignty, but also a standard for AI.

Network: Disaggregation with SONiC

In addition to its server infrastructure, ScaleUp relies on network disaggregation using SONiC – the open-source operating system for network switches. The model: no more proprietary switches, where hardware and software must come from a single source, but rather an open networking model. Ethernet as the standard, open source as the operating system.

SCS as the link: from hardware to application

Building on this open hardware foundation, ScaleUp deploys the Sovereign Cloud Stack (SCS) using OpenStack. The logic is as follows: OCP hardware is abstracted via OpenStack; on top of that sits SCS as a standardising layer; above that is Kubernetes; and on top of that is the AI application.

Streit addressed a topic of practical relevance: if SCS is also to be used for AI workloads, standardised GPU flavours across providers will eventually be required. The standardisation of VM flavour names is already established in SCS – but this does not yet exist systematically for GPU-based nodes. This discussion continued immediately after his presentation during the Q&A session: the technical basis already exists in the flavour specification, but it remains to be seen whether the resulting letter-number combinations are also suitable for marketing departments.

Gardener: GPU clusters at the touch of a button

For infrastructure management, ScaleUp relies on Gardener as its orchestration layer. Gardener automates the deployment of Kubernetes clusters – including GPU-enabled worker nodes for computationally intensive AI applications. Auto-scaling, self-healing, uniform abstraction: the complexity of operations is eliminated without relinquishing control.

The aim here is to create a standardised infrastructure on which AI workloads can be rolled out regardless of the provider – another issue which, according to Streit, needs to be given even greater attention within the SCS community.

Privately hosted LLMs: data stays where it belongs

The most practical aspect: running open-source models on your own infrastructure, for example via vLLM or KServe. All data – that is, both the model itself and all documents integrated via RAG – is stored in your own data centre, on local Ceph-based S3 storage. Optional end-to-end encryption or a fully isolated private environment completes the picture.

The result: powerful AI without the need to share data with a third party.

Conclusion

At a time when AI is dominating the technological debate and more and more companies and public authorities are coming under pressure, it is refreshing to hear a talk like this. Christoph Streit does not offer visions, but reality: a well-thought-out stack that can be implemented today, which is open and controllable from the firmware right through to the language model.

The key takeaway: sovereignty is not an either/or issue. You don’t have to choose between powerful AI and data control – if you combine the right building blocks. The challenges are real, particularly when it comes to GPU availability and cost-effectiveness. But the technological foundation is in place. This is an important message at a time when many still believe that genuine AI expertise cannot be achieved without relying on the major American providers.

Christoph Streit’s full presentation is available at YouTube.