The Techstrong.IT stories enterprise IT leaders actually read.
Volume 343, Issue 2 · 9 articles from 3 issues
McKinsey projects data center electricity use could grow 24% a year through 2030, with power supply falling short of AI demand.
Techstrong IT reports that McKinsey's Global Energy Perspective 2026 projects data center electricity consumption could grow 24% a year through 2030 and reach 10% to 15% of global electricity, the fastest-growing new load in developed markets.
McKinsey's near-term scenarios show power supply falling short of IT compute demand through 2030, with transmission lines, transformers, grid approvals and storage gating capacity as much as new generation.
Operators are turning to on-site natural gas, fuel cells and batteries while they wait for utility connections.
Google Cloud's Sept. 1 outage from a router maintenance error shows why apps must be designed to survive partial-zone failures.
On September 1, 2026, parts of Google Cloud's us-central1-b and us-central1-f zones suffered network degradation for 4 hours and 11 minutes after a procedural error during routine router maintenance disconnected redundant fiber links.
Nearly 20 products, from BigQuery and Cloud SQL to GKE and Cloud Run, were affected because network fabric sits beneath everything, including managed services.
The author notes that partial-zone incidents are normal at hyperscale and that whether they become application outages is a design choice on the customer's side of shared responsibility.
ManageEngine CEO Rajesh Ganesan says IT teams need to fix fundamentals and context before deploying AI agents, or risk real damage.
At the ManageEngine User Conference, CEO Rajesh Ganesan told IT professionals that more effort must go into readiness before applying AI to IT operations, naming six priorities: infrastructure reliability, data quality and sovereignty, operational resilience, power and compute capacity, governance and compliance, and security.
IT teams already have plenty of models to choose from but lack the operational intelligence, shared data fabrics and orchestration needed to automate workflows.
Without identity controls and guardrails, autonomous agents could reach sensitive data or take destructive actions such as deleting databases.
D-Wave opens beta access to a 21-qubit classical simulator of its dual-rail gate-model quantum architecture.
Techstrong IT reports D-Wave has opened beta access to the gate-model quantum simulator it announced in June, available through its Leap cloud service and Ocean SDK.
The classical Monte Carlo simulator, built on Qiskit's Aer statevector software, models D-Wave's dual-rail superconducting architecture designed to make errors easier to detect, supporting up to 21 qubits plus detectable errors and real-time control.
Beta users including BBVA, FirstQFM, Florida Atlantic University and the Julich Supercomputing Centre are exploring financial applications, quantum machine learning and error-aware computing before D-Wave's gate-model hardware arrives.
Multi-cloud data friction, not models, is sinking AI projects; CIOs should simplify the data layer rather than the clouds.
Jim Gannon argues that most CIOs have a cloud strategy but not one for keeping it simple.
Expanding estates across providers and on-premises systems add the overhead of copying data, separate tools and policies, and reconciling governance, with each silo adding risk and cost.
Citing S&P Global's finding that 42% of companies abandoned most of their AI initiatives in 2025, up from 17%, he says failures trace to data infrastructure rather than models.
Simplicity doesn't mean fewer clouds but consistent data services, governance and mobility across them, so data can be used wherever AI services run.
Consistent resource naming across AWS, Azure and Google Cloud lets teams tie scattered alarms to a single hybrid failure.
Alastair Cooke, reflecting on Selector's Cloud Field Day 26 presentation, notes that physical location gave on-premises naming its structure, but cloud resource IDs say nothing about account, region or owner, and each provider and Kubernetes flavor organizes resources differently.
A dropped BGP session between data center and cloud can raise an SNMP trap for one team and an app-unreachable alert for another, with nothing linking the two.
Selector normalizes every resource name by cloud, account, region and type plus the native ID, and applies the same scheme to alarms from CloudWatch and Google Cloud, so correlation becomes a simple prefix match.
Note: published in Techstrong's sponsored section.
AI agents that reverse engineer legacy middleware metadata could cut migration costs and weaken platform lock-in.
Mike Vizard reports that SnapLogic has enhanced its SnapLogic Intelligent Modernizer (SLIM) engine with AI agents that reverse engineer metadata from poorly documented legacy middleware.
The agents generate integration pipelines without needing a purpose-built connector for each source.
SLIM assesses how integrations relate to recommend a starting point, runs operations concurrently and lets teams add context and rerun agents.
CTO Jeremiah Stone says generative AI is cutting migration costs enough to loosen vendor lock-in, which matters because legacy maintenance fees consume budget that could fund innovation.
90% want autonomous IT operations, yet 77% won't let AI act without human approval and only 17% have unified visibility.
A Coleman Parkes survey of 1,200 decision-makers for Riverbed found that 90% of organizations want AI agents to drive autonomous IT operations and expect human-supervised or highly autonomous ops within two years.
However, 77% hesitate to let AI decide without human approval, and only 24% trust AI recommendations with limited review.
Readiness is low, with just 17% having unified visibility across networks, apps, endpoints and cloud, 19% of IT operations automated and only 8% applying AI to operations enterprise-wide.
The top obstacles are security and compliance (55%), disruption risk (45%) and lack of trust (39%).
GPU virtualization overhead is usually under 2%, and VM lifecycle tools make it a better fit than bare metal for most enterprise AI.
Alastair Cooke recaps VMware by Broadcom's Cloud Field Day 26 case that the bare-metal argument doesn't hold for most enterprise AI, since GPU attachment in VMware Cloud Foundation is essentially pass-through and reported overhead is usually under 2%.
The bigger cost of bare metal is operational, including servers tied to one OS build, patching downtime and rebuilds to move workloads, which templates, snapshots and vMotion avoid.
NVIDIA vGPU, with its licensing, enables vMotion for 'pet' workloads like researcher workstations, while pass-through suits Kubernetes model serving and bare metal remains best for training large models from scratch.
Note that this article was published in Techstrong's sponsored section.
More from Techstrong.IT · All IT section newsletters · This week