From high-end AI hardware to software-supply-chain defence
Three Selectel articles connect AI infrastructure with operational security. In its report on HP’s ZGX Fury G1n, Selectel relays HP’s claims that a workstation built around NVIDIA’s GB300 Grace Blackwell Ultra can run models with up to one trillion parameters locally and fine-tune models of up to 100 billion parameters. The article lists 748 GB of combined memory, up to 20 PFLOPS of FP4 compute and a US starting price of $132,000; it also promotes Selectel’s own GPU-server and AI-router services, but does not establish that the HP system is available in Russia. A second article, citing data attributed to security company Aikido, describes an alleged GitHub campaign involving 4,837 compromised accounts and pull requests targeting more than 61,000 repositories, with malicious workflows intended to steal credentials. The third article documents a Selectel experiment in which a Qwen3-4B model, Python and FUSE generated a synthetic filesystem on demand. According to the author’s five test runs, an AI agent acting as an attacker accepted the fabricated server as genuine each time and extracted 18 invented secrets in total. These are provider-published reports and experiments rather than independently validated benchmarks or incident findings.
A hybrid AI strategy is emerging around locality, access and containment
Taken together, the articles suggest that Selectel is positioning AI as an infrastructure continuum rather than a single cloud service. At one end is local execution on workstations or modest consumer hardware; at the other are promoted dedicated GPU servers and an API-based router for accessing multiple models. The recurring strategic theme is control over where models and data run. Local inference is presented as technically practical, while hosted GPUs and routed model access reduce the integration and capital burden. Security is becoming part of the same proposition: the GitHub report focuses on credentials moving through CI/CD systems, while the FUSE experiment explores how generative AI can create adaptive deception environments and how agents can be misled by synthetic evidence. For Russian organisations, the emphasis on local deployment, Russian-language scenarios, Cyrillic filenames and self-managed infrastructure is regionally relevant. It points towards AI adoption shaped by data control, infrastructure autonomy and compatibility with existing Linux tooling, although the articles do not define a formal Selectel product roadmap.
Procurement, consistency and verification remain material risks
The strongest limitations appear where experimental promise meets production requirements. HP’s specifications and model-size limits are manufacturer claims repeated by Selectel, with no independent performance, energy-efficiency or workload benchmarks. Its quoted prices and sales reference the United States, while the article supplies no evidence about Russian availability, lawful procurement, vendor support, firmware access, warranties or replacement parts. Those omissions matter in Russia’s sanctions context and should be resolved through legal, supply-chain and lifecycle due diligence rather than inferred from technical specifications. The GitHub article provides actionable indicators, but its scale figures are attributed to a third party and are not corroborated within the supplied material. The synthetic-filesystem experiment is transparent about deeper technical weaknesses: files initially had implausible sizes and ownership, guessed paths always existed, credentials lacked realism, and independently generated configurations contradicted one another. The attacker was also the same small model family used by the trap, so the result does not demonstrate effectiveness against skilled humans or more capable agents. Enterprises should therefore treat adaptive honeypots as research or an additional detection layer, not as a replacement for repository controls, secret rotation, monitoring or established deception systems.
Decision guidance from the local AI workstation report, GitHub security analysis and FUSE experiment
For IT decision-makers in Eastern Europe—and especially organisations operating in Russia—the practical conclusion is to evaluate Selectel around three separate questions: whether its locally obtainable GPU infrastructure satisfies workload and lifecycle requirements, whether its model-access services meet data-governance constraints, and whether its security guidance can be integrated into existing controls. The HP ZGX Fury G1n article illustrates demand for powerful local inference but should not be treated as evidence of regional supply. The GitHub attack report reinforces the need to audit workflows, contributor permissions, package-publishing accounts and exposed secrets. The LLM-backed filesystem experiment offers the most distinctive direction: AI can generate both defensive deception and convincing false evidence for autonomous agents. Selectel’s apparent priority is therefore not merely more AI compute, but a locally relevant operating model spanning private execution, hosted acceleration, model access and agent-aware security. Buyers should validate each layer independently, with particular attention to sanctions-related procurement, support continuity, data location, model provenance and measurable production performance.