The operations stories enterprise IT leaders actually read.
Volume 341, Issue 3 · 20 articles from 2 issues
Vulnerability scanning answers what is broken inside, not where an artifact came from or whether it changed.
Model weights, training datasets, inference runtimes, and agent tooling increasingly ship as OCI artifacts, and standard vulnerability scanning cannot detect malicious payloads in serialized weights or verify provenance.
Researchers found malicious PyTorch models on Hugging Face in 2024 that used pickle's __reduce__ hook to execute code on torch.load(), and nullifAI models later evaded static analysis by compressing payloads with 7z.
Scanning answers what is broken inside an image, not where it came from or whether it was modified. Without signing, tampering anywhere in the pipeline succeeds silently; with it, an attacker must compromise a scoped signer identity and leave an auditable event. Registry-managed signing plus Kyverno admission control enforces this without teams managing keys.
A practical guide to metrics, logs and traces as the foundation of data centre reliability.
Modern data centres rely on observability, understanding system state through external outputs, to maintain service health. The article explores three pillars: metrics for broad health signals, logs for detailed events, and traces for request flows across services.
An incident walkthrough demonstrates how these tools enable rapid diagnosis. Key practices include SLOs for reliability targets, automation for self-healing and intentional telemetry design.
Starting with Prometheus metrics, then structured logging, then tracing gives organizations a practical implementation path.
Enterprises are consolidating infrastructure as AI-driven demand resets memory and flash pricing.
An Omdia study of 400 IT professionals found 98% of organizations are consolidating infrastructure layers to manage rising component costs.
Memory and flash prices surged 90-95% and 53-58% respectively in early 2026, driven by hyperscaler AI infrastructure purchases.
Software-defined storage ranked first among planned investments for non-users. Seventy-four percent report delayed on-premises AI deployment and integrated data protection strategies.
AI data centers are becoming specialized facilities with larger campuses, advanced cooling, and massive power needs.
AI data centers are shifting from traditional multi-purpose facilities to purpose-built AI factories focused on training and inference workloads.
Global data center power demand is expected to rise 165% by 2030, with electricity consumption reaching roughly 945 terawatt-hours annually.
Modern AI campuses now operate between 100 and 750 megawatts compared with older facilities at 10 to 30 megawatts, requiring liquid cooling and high-density racks exceeding 100 kilowatts. Companies are investing in high-speed networking, alternative energy including solar and battery storage, and AI-powered facility management software to support this growth.
Pennsylvania made AI data centre permits conditional on a signed contract, local consent and public disclosure.
AI News examines Pennsylvania's new approach to AI data centre regulation, which operates through permitting rather than legislation. Permits are now conditional on a signed contract, demonstrated local consent and public disclosure of key project details.
Because the mechanism sits within existing permitting authority, no new law was required to put it in place.
The article argues this makes the approach readily copyable by other states facing community opposition to large AI builds. It frames the model as a template for regulating data centre development through conditions rather than statute.
Insider threats cross physical and digital boundaries, so cyber and physical security teams must integrate.
Research cited shows 68% of surveyed organizations encountered between 21 and more than 40 insider threat incidents in 2026, up from 57% in 2024. Because insiders already hold authorized access, incidents are harder to detect and more damaging.
The article recommends a cross-functional insider threat team spanning cybersecurity, physical security, HR, legal, IT operations and executive leadership. Integrating badge access records, surveillance footage and authentication logs lets teams correlate signals, such as a user downloading sensitive files while entering a restricted area.
Clear policies covering physical and digital conduct form the program's foundation.
IT and OT are converging under IoT and AI, forcing integration of data, networks and security.
IT manages data processing for business applications while OT controls industrial devices and processes.
Though historically separate, IT/OT convergence driven by IoT and AI creates opportunities for cost savings, improved security and better data sharing.
Organizations face challenges including device inventory management, firmware updates and encryption. Merging disparate data formats requires careful integration planning.
Critical infrastructure needs specialized, sovereign AI rather than general-purpose platforms.
The article examines why general-purpose AI systems may not suit critical infrastructure and services, emphasizing the importance of maintaining sovereignty over data and systems in those sectors.
It explores the considerations involved in deploying specialized AI solutions rather than generic platforms for sensitive applications.
Classify changes by reversibility, then automate evidence collection to justify lighter approvals.
Joseph Michela argues that change management can be streamlined for reversible changes by automating the evidence-verification step rather than the approval itself.
He proposes three reversibility categories: bi-directional changes that can be undone through the same API, mono-directional changes reversible only by another method, and destructive changes that cause permanent data loss.
For the reversible categories, teams should gather machine-verifiable artifacts such as automation scripts, audit records, metrics, and partner-system documentation to justify a lighter approval path.
Making this work requires collaborating with the change management team to identify which evidence format produces the largest efficiency gain while preserving compliance and auditability.
Data Center Knowledge rounds up August 2026 hardware developments across energy, minerals, networking, CPUs and racks.
Data Center Knowledge publishes its monthly hardware roundup for August 2026. The month's developments show AI infrastructure attention shifting toward energy and minerals alongside traditional compute concerns.
Networking and CPU announcements feature prominently as clusters scale. Rack-scale systems continue to consolidate what were previously separate procurement decisions.
The roundup also covers power-ready capacity, reflecting how availability of energized space now gates deployment more than hardware supply.
Traditional cloud FinOps breaks on AI because tokens are not taggable resources like VMs.
Gartner projects 47% growth in AI spending in 2026, but enterprises cannot control it without real-time visibility into what actually drives cost.
Conventional cloud cost management fails here because token calls are not taggable resources the way virtual machines are, and autonomous agents can burn through budget with no oversight trigger.
The author proposes moving cost attribution down from the application layer to the infrastructure layer using eBPF kernel-level sensors, which track continuously without heavy instrumentation.
That shifts governance from periodic top-down audits to real-time practitioner accountability, letting organizations treat AI as a managed investment rather than an unpredictable expense.
Vacancy sits near 1% but few facilities support 140-160 kW racks, so AI-ready capacity is scarcer than headline numbers suggest.
Data Center Knowledge argues that headline vacancy figures overstate the capacity actually available for AI workloads.
Vacancy sits at roughly 1%, but very few existing data centers support racks drawing 140 to 160 kW. Liquid cooling availability narrows the usable set further, since air-cooled halls cannot take those densities.
Credit screening by operators narrows options again for tenants without strong balance sheets. The result is that AI-ready capacity is considerably scarcer than the market-level vacancy statistic implies.
Enterprises face 9-18 month lead times and steep price increases for IT gear as hyperscalers absorb memory capacity.
Memory shortages driven by hyperscaler demand are causing dramatic price increases and extended lead times across servers, storage, and networking equipment through 2027.
Industry experts recommend optimizing existing infrastructure utilization, engaging finance teams with detailed forecasts, and working closely with vendors and VARs to map workload requirements. Organizations should consider alternative suppliers, cloud solutions, and colocation providers while avoiding single-vendor dependencies, particularly for AI infrastructure.
Strategic approaches include using public cloud for proof-of-concept work and choosing performance-tiered solutions rather than pursuing top-tier equipment exclusively. Planning and ordering immediately is critical as delivery times stretch past 12 months.
A routable on-prem Kubernetes design using BGP, ECMP and BFD reduces NAT and overlay complexity and speeds failure recovery.
Author Sergey Speranskiy describes building an on-premises Kubernetes cluster that participates in real data center routing rather than hiding behind NAT and overlays.
The design uses BGP to advertise pod and service networks, ECMP for multipath load distribution, and BFD for fast failure detection. Removing the overlay reduces complexity and eliminates a layer of troubleshooting difficulty.
The article covers the supporting components including the BIRD routing daemon, kube-router, IPVS and pod CIDR planning. It also covers control plane resilience and the observability needed to operate the result.
Over 66 GW is under construction yet vacancy has stayed at 1% for a third straight year.
More than 66 gigawatts of data center capacity is under construction in North America, and vacancy has still held at 1% for the third consecutive year.
JLL attributes the gap to structural demand from growing digital services and AI adoption that keeps outpacing supply.
Power constraints and land scarcity in established hubs are pushing developers into frontier markets such as West Texas, Ohio, and Louisiana.
The sector faces grid connection waits exceeding four years, rising local opposition, and the need to secure electricity, land, financing, and community support simultaneously.
A survey finds ITSM teams save time on some tasks while work piles up elsewhere, so net workload rises.
ITPro reports on new survey research into AI in IT service management. The research finds AI in ITSM departments may be delivering a reasonable return on investment but is not yet making life easier for IT professionals.
Teams save time on some tasks while work accumulates in other areas.
The net effect is a larger workload rather than a smaller one. The article examines where the displaced work goes and why the savings do not net out.
Two-thirds of companies have moved AI workloads off public cloud to on-prem or private cloud.
A Cloudera survey of 1,500 enterprise leaders finds legacy infrastructure is actively blocking AI deployment, with nearly three-quarters of enterprises needing to modernize before they can support AI demands.
Many have delayed or cancelled initiatives over data governance and regulatory obstacles. The report's framing is that architectures built for traditional analytics were not designed for the scale, governance, and flexibility AI now requires.
Two-thirds of companies have already shifted AI workloads from public cloud to on-premises or private cloud, and 84% report higher infrastructure costs from AI, driving hybrid architectures that trade off flexibility, cost, and governance.
An MGX-compatible 800-VDC power rack enters production in H2 2026, with AC facility power still in place.
NVIDIA has moved its 800-VDC power architecture into production planning, with an MGX-compatible 800-VDC power rack entering production in the second half of 2026.
The approach is deliberately hybrid: high-voltage DC goes to AI compute racks while the broader facility keeps running on existing AC distribution.
Three deployment options are planned, starting with rack-adjacent power racks in Q3 2026, centralized power centers targeted for Q3 2027, and data hall-level power blocks using solid-state transformers around 2029.
Maintaining compatibility with existing facility infrastructure is what makes the path viable, and the design targets up to 1 MW per rack in future native 800-VDC configurations.
AI currently reaches just 45% of enterprise data, and only half of firms trust their agents' decisions.
A Google Cloud and MIT report surveying 300 IT executives concludes that enterprises need strong data foundations before AI agents can scale. Nearly all surveyed companies plan to use agents and two-thirds intend to deploy within two years, yet only 10% currently have widespread agentic deployment.
AI reaches an average of just 45% of enterprise data, and only half of organizations trust their agents' decisions, a gap the report ties directly to data readiness.
Legacy system modernization, workforce training, and governance updates could push full agentic implementation out another three to four years.
OCP push for 800-volt DC power aims to cut copper use 50-80% in high-density AI data centers.
Google, Microsoft, and NVIDIA are working through the Open Compute Project to make 800-volt direct current the standard for powering high-density AI data centers.
The goal is to handle surging GPU power draw while reducing infrastructure complexity and operating cost relative to traditional AC distribution.
Cited benefits include a 50% to 80% reduction in copper usage, an 8% to 12% cut in annual energy-related OpEx, and $4 million to $8 million in CapEx savings per 10 MW built. AWS is notably absent from the effort and is reportedly pursuing its own internal infrastructure redesign instead.
More from Operations · All IT section newsletters · This week