HPE’s first server with the NVIDIA Vera CPU, record Dell orders and a memory shortage, Huawei storage for AI inference, and the first obligations under the AI Act. What affects planning for 2027.
Alongside its third-quarter results on 2 September, HPE introduced the HPE ProLiant Compute DL394 Gen12 – a server built on the NVIDIA Vera CPU and positioned as a compute foundation for agentic AI workloads and demanding data processing. General availability is planned for autumn 2026. The results themselves showed revenue of roughly 12.2 billion dollars, up by a third year on year, and a record backlog of AI systems.
What this means for you: the DL394 is interesting for organisations that want AI inference without investing in large GPU systems – the Vera CPU has high memory bandwidth and handles smaller models and agents well. It can go into a 2027 tender now; for ordinary workloads (virtualisation, databases), however, the standard Gen12 models with AMD or Intel remain the cheaper and more proven choice.
Source: Tech Insider, 4 Sep 2026
On 1 September Dell raised its revenue outlook for the current fiscal year to 192 billion dollars and reported a record backlog of AI server orders worth 95 billion dollars. Management also named openly what is holding them back: the availability of DRAM and NAND memory, which every server and every storage system needs.
What this means for you: this is not just news for investors. When the largest manufacturers cannot keep up because of memory, lead times get longer and configuration prices rise for ordinary enterprise servers too – not only for AI. If you have a refresh or expansion planned for the fourth quarter, order early and expect longer delivery times; in tenders, set a realistic delivery deadline, otherwise fewer bidders will take part.
Source: 24/7 Wall St., 2 Oct 2026
At the Huawei Connect conference in Shanghai, Huawei presented OceanStor M900 Context Memory Storage – storage designed for the KV cache used in AI inference, due in early 2027 in a 2U air-cooled and a 1U liquid-cooled version; a cluster is to scale to 64 PB. The predecessor, the M800, is available today. Huawei also announced a partnership with ThinkParQ around the BeeGFS parallel file system for HPC and AI workloads.
What this means for you: for most Slovak organisations the M900 is oversized, but the direction matters: manufacturers (Huawei, just like HPE with its Alletra X10000) are starting to address the fact that GPU memory is expensive during AI inference and that it pays to offload it to storage. When choosing storage for the next three years, ask about RDMA support and low latency, not only about capacity.
Source: StorageNewsletter, 30 Sep 2026
On 30 September HPE announced a 1.2-billion-dollar contract with the cloud provider Vultr for AMD Helios AI racks, including HPE networking, liquid cooling and services. At its investor day it also raised the long-term outlook for its networking division, with a target operating margin in the 25 to 30 percent range.
What this means for you: HPE is making clear that networking is a growth priority for it – after the Juniper acquisition this shows in the pace of innovation in Aruba Central and Mist. For customers it means long-term support for the platform, but also faster changes in licensing models; when replacing network equipment, ask for licence terms covering the full five years, not just the first one.
Source: 24/7 Wall St., 30 Sep 2026
After the August wave (Qwen3.8-Max with 2.4 trillion parameters, Kimi K3), September brought smaller, practically deployable models. On 10 September DeepSeek released V4.1 Flash under the MIT licence – a sparse model with 8 to 16 billion active parameters and a context window of up to one million tokens. Z.ai published the weights of GLM-5.3, which holds a leading position among open models in independent evaluations. The most usable option for a single server remains Qwen3.8-27B (Apache 2.0), which runs on a single H100 card or on two consumer GPUs.
What this means for you: an internal assistant over documents or form processing no longer needs a model that requires a whole rack. A 27B dense model on a single server with one or two GPUs handles most such tasks in Slovak and English. With Chinese models, check the licence before deploying in a public institution (MIT and Apache 2.0 are unrestricted; some custom licences have revenue-based clauses) and the origin of the training data if your internal policy requires it.
Source: Slash Digital, 26 Sep 2026
Since 2 August 2026 the EU Artificial Intelligence Act has applied generally. For most organisations the most important obligations are the transparency rules in Article 50: users must know they are interacting with an AI system, and synthetic content (text, images, audio) must be labelled. Fines go up to 15 million euros or 3 percent of global turnover. Obligations for stand-alone high-risk systems have been postponed to December 2027.
What this means for you: if you deploy a chatbot for citizens or customers, or generate texts and images that go out to the public, you have had to label them since August. Internal tools used only by employees are not affected. For public institutions preparing AI solutions involved in decisions about people (benefits, permits, recruitment), December 2027 is the date from which strict requirements on documentation and oversight apply – and projects should be designed accordingly now.
Write to us through the contact form – we will tell you which parts of this review concern you and which can wait.
Book a consultation ›