- Dell Technologies reported a 30% surge in inquiries for its Apex Private AI solutions in Q2 2026, a concrete signal that enterprises are actively moving AI workloads off public cloud.
- Regulatory pressure, GDPR, HIPAA and the EU AI Act’s August 2026 transparency requirements, is the primary driver, as private infrastructure makes audit trails and data residency far easier to demonstrate.
- For latency-sensitive applications like warehouse robotics, and for firms where trained models are core IP, private infrastructure removes constraints that public cloud cannot reliably solve regardless of network quality.
Dell Technologies reported a 30% jump in inquiries for its Apex Private AI solutions in Q2 2026, and the pattern behind that number is worth examining. Enterprises are pulling AI workloads out of public cloud for reasons that are structural, not stylistic: tightening data sovereignty rules, hard latency ceilings and the competitive risk of running proprietary models on shared infrastructure.
Ensuring Uncompromised Data Sovereignty
For multinationals operating in heavily regulated regions, particularly the EU, keeping data inside specific geographic boundaries is a legal requirement, not a preference. Private infrastructure gives enterprises direct control over where data lives and how it moves, which matters considerably when GDPR and national data localisation rules apply and the infrastructure is shared with other tenants.
Enterprises in sectors like German automotive manufacturing are deploying dedicated AI supercomputing clusters inside their own data centres to handle sensitive R&D data and customer driving patterns. Keeping everything in-house simplifies compliance proof considerably. The direct payoff is lower exposure to regulatory fines and a cleaner compliance story for customers who want to know exactly where their data sits.
Meeting Stringent Regulatory Compliance
Finance, healthcare and government agencies face compliance frameworks that go beyond GDPR. For a U.S. healthcare provider deploying diagnostic imaging AI, HIPAA required that patient records used to train and run those models never leave a controlled environment.
Public cloud environments can be secured, but auditors often require granular proof of physical and logical isolation that is harder to produce in a shared-tenancy model. The provider invested in on-premise NVIDIA DGX systems, giving its security team direct control over physical access, simplified audit trails and cryptographic isolation of patient data. Compliance burden dropped; deployment of AI diagnostic tools accelerated. The trade-off was capital expenditure upfront, but the regulatory clarity it bought was the deciding factor.
This dynamic is not unique to healthcare. For any sector where regulators require documented chain of custody over sensitive data, private infrastructure removes a layer of uncertainty that public cloud cannot easily eliminate. The EU AI Act’s August 2026 transparency requirements are adding further pressure on enterprises to demonstrate exactly this kind of control.
Optimising Performance for Low-Latency AI
Cloud latency is measured in milliseconds. For warehouse robotics, that is too slow.
Autonomous sorting robots and real-time route optimisation often demand sub-millisecond response times, a threshold that even geographically optimised public cloud connections can struggle to meet. Enterprises deploying these systems have moved compute on-premise, with sorting throughput and route adjustment responsiveness both improving during peak hours. The infrastructure cost was offset by operational savings.
The engineering argument for edge AI is straightforward: if an application cannot tolerate round-trip latency to a remote data centre, the compute needs to be local. Public cloud architectures cannot always solve that constraint regardless of network quality.
Gaining Granular Control and Customisation
Standard cloud instances are designed to serve a wide range of workloads. That breadth is a feature for most customers, but a limitation for teams with highly specific performance requirements.
An aerospace engineering firm working on computational fluid dynamics models and material simulations for aircraft design found that ceiling quickly. Rather than force its workloads into available cloud instance types, the firm built a private cluster using AMD Instinct accelerators and open-source AI frameworks. That gave its engineers direct control over hardware topology, memory configuration and software libraries for maximum performance on those specific simulations. Iteration cycles shortened and simulation accuracy improved. For a firm where design efficiency is a direct competitive factor, that level of control justified the infrastructure investment.
Protecting Proprietary Models and IP
A trained AI model can represent years of development and significant compute spend. For some enterprises, it is the core competitive asset. Deploying it on shared public infrastructure carries real risk.
FinTech companies specialising in high-frequency trading algorithms and fraud detection are moving AI development and deployment to private infrastructure built on dedicated NVIDIA H200 GPUs. Their trading models are continuously retrained on proprietary market data, and these firms have concluded that keeping both the models and the training data entirely off shared infrastructure is worth the operational overhead. The goal is to eliminate any exposure path for model weights and training data, whether through misconfiguration, insider access or infrastructure-level vulnerability.
For firms where the model architecture itself is the product, this calculus is clear. The risk being managed is not just data leakage but competitive intelligence: a sufficiently detailed reconstruction of a trading model’s behaviour could inform a rival’s strategy. Private infrastructure closes that surface entirely.
Dell’s reported demand spike is a useful data point, but the underlying drivers are structural. Data sovereignty rules are tightening, particularly as the EU AI Act adds new documentation requirements. Latency constraints are a hard physical limit for edge applications. And for firms where AI models are core IP, the risk calculus favours keeping everything in-house. Public cloud will remain the right answer for many workloads. For the ones where it is not, the case for private infrastructure is increasingly well-defined. For more coverage of AI chips and infrastructure, visit our AI Hardware section.
