The operations stories enterprise IT leaders actually read.
Volume 343, Issue 3 · 14 articles
Rising chip power is breaking cooling, power delivery and interconnect at once; xFusion argues they must be co-optimized. Partner content.
Accelerator racks have climbed from about 20 kW in 2023 to 240 kW in 2026, with chips past 2,000 watts and megawatt-class racks in validation.
Recapping an xFusion keynote at the 2026 OCP Global Summit, the article argues this step change breaks cooling, power delivery and electrical interconnect simultaneously, and solving two while ignoring the third still strands infrastructure.
Air cooling is finished for high-density AI racks; xFusion's diamond-copper cold plates claim about twice copper's conductivity and more than 4,000 watts of single-chip cooling capacity.
This is labeled partner content.
How AI threat detection in data centers works: unified telemetry plus supervised, unsupervised and sequence models, and what to ask vendors.
Breach costs keep rising, with IBM's 2026 report putting the global average at $4.99 million, $6 million for AI-enabled attacks, while organizations using security AI and automation saved nearly $2 million.
Perimeter tools miss lateral movement inside facilities with fast-changing virtual workloads, shared tenants and connected building systems.
AI detection starts with a pipeline feeding hypervisor, identity, network flow, out-of-band management and physical access telemetry into a platform such as a SIEM.
Effective platforms combine supervised models for known attacks, unsupervised models for anomalies and sequence models for event order, and the article covers vendor evaluation.
A reference list of 100 leading data centre operators, cloud providers and telcos, with headquarters, size, revenue and footprint.
Analytics Insight lists the top 100 global data centre companies of 2026, from established operators such as Equinix and Digital Realty and hyperscalers including AWS, Microsoft, Google and Meta to fast-growing regional players in Asia, Europe, the Middle East and Australia.
Entries cover headquarters, founding year, company type, employee count, annual revenue and reported data centre footprint.
The article says growth will depend on energy availability, sustainability, investment capacity, cybersecurity and support for demanding AI workloads, and cautions that figures are approximate and revenues may not reflect data centre operations alone.
Splunk survey: 57% of organizations spend more on observability than on the infrastructure being monitored, and AI agents add more data.
Cisco Splunk's State of Observability 2026 report, based on a survey of 2,050 IT operations and engineering professionals, found that 57% of organizations have observability costs that exceed the cost of the infrastructure they monitor.
Splunk says a production AI agent emits four to five times more telemetry than a traditional application, covering workflows, tool calls, context and model interactions.
Noisy logs (31%) and high-cardinality metrics (27%) remain the biggest cost drivers, ahead of AI infrastructure.
The report found that 38% of organizations have cut telemetry coverage or retention to control costs, which can create blind spots for later troubleshooting.
October Windows Server updates destabilized on-premises Exchange, and Microsoft offers a registry fix that raises Endpoint Mapper limits.
After installing October 2026 Windows Server updates, organizations running on-premises Exchange Server reported Outlook on the web failures, growing mail queues, inconsistent access and database availability group problems.
Microsoft traced the issue to a limit on simultaneous Endpoint Mapper requests, set by default at 512, which heavier Exchange workloads can exceed.
It has published a registry-based mitigation that raises capacity while it works on a permanent fix.
The bug affects Exchange on Windows Server 2016, 2019, 2022 and 2025.
Google, Amazon and Meta are funding uprates at existing nuclear plants, adding capacity without building new reactors.
Google and Amazon have signed agreements with Constellation Energy that support about 1,080MW of additional nuclear generation from existing reactors, and Meta is backing uprates at plants in Ohio, Pennsylvania and Illinois.
Under Google's 20-year deal, Constellation will invest more than $4.3 billion upgrading 11 reactor units, with first output in 2028.
The NRC expects 31 uprate applications through 2032, worth about 2.4GW.
Uprates avoid building new reactors but still involve multiyear timelines for equipment, outages and approvals, and FERC's rejection of parts of PJM's large-load framework shows the gap between financing generation and delivering power to individual sites.
Power compute effectiveness (PCE) measures how much provisioned power actually reaches compute, a complement to PUE.
A proposed metric, power compute effectiveness (PCE), measures what share of a facility's provisioned power is structurally available for compute and how much is tied up in cooling, electrical losses and other support systems.
It is defined in an Open Compute Project paper from Airsys and is positioned as a complement to PUE, not a replacement.
Where PUE reflects operational efficiency once a site is running, PCE looks upstream and asks how much server capacity really fits within a fixed power envelope.
AFCOM's Bill Kleyman sees it as especially useful for retrofits and expansions where more grid power is not available.
Microsoft's preview odsync CLI reports OneDrive sync health, policies and compliance per device, with JSON output for automation.
Microsoft is previewing a OneDrive Sync Client command-line interface, invoked as odsync, to make troubleshooting sync problems across large device fleets easier.
It reports configured accounts, client and CLI versions, folder backup settings, applied policies, sync and account status, device compliance, and whether and how the client is installed.
Results can be filtered by account, include verbose diagnostics, and be exported as JSON for scripting.
Petri notes that organizations may need to rethink their device management and remediation automation to take advantage of it.
Fast GPU refresh cycles make lifecycle planning for redeployment, resale and recovery a capital and facilities issue.
In a DCF (Data Center Frontier) Show podcast write-up, Linda Li of IT asset recovery firm Re-Teck argues that faster AI accelerator refresh cycles turn GPU lifecycle management into a question of capital preservation, facility design and governance, not just IT asset disposition.
Hardware that no longer suits a leading-edge cluster can be redeployed, resold, harvested for parts or recycled, but the outcome depends on decisions made long before it leaves the rack.
She recommends facilities with scalable power, modular racks and flexible cooling, and choosing ITAD partners during procurement.
She also points to telecom contracts that require suppliers to take back the equipment they replace.
Inference workloads differ from training in hardware, latency, location and rack density, so facility design and capacity plans must adapt.
Most AI infrastructure build-out has targeted large training clusters, but production inference brings different requirements.
Inference must answer requests with low latency, and needs range from CPU-only for small models to GPU clusters for large generative models, depending on model size, memory and concurrency.
Uptime Institute's 2026 survey shows more operators reporting peak rack densities of 30 kW or more, and an NVIDIA GB200 NVL72 rack draws about 120 kW.
The article advises sizing from expected traffic and response targets, analyzing power at rack, row and facility level, and planning for peak loads, headroom and expansion.
A Grafana Labs/ETR survey finds most firms spent more on observability as AI adds systems to watch, and favor consolidation.
Grafana Labs and ETR surveyed 300 IT and engineering leaders for the State of Observability 2026 report, which comes from an observability vendor.
More than three-quarters reported higher observability spend, driven by new AI features (61%), wider coverage as AI-accelerated development creates more systems (60%), business growth and vendor price rises.
79% rate tool consolidation very or extremely important, and 86% call their consolidation efforts successful.
52% are comfortable letting AI carry out full remediation, with trust depending on accuracy and explainability.
Dell'Oro's OCP recap: optics moving into compute, cold plates eclipsing CDUs, 800 Vdc sidecars and a path to 1MW racks.
Dell'Oro's recap of a sold-out OCP Global Summit, with about 15,000 attendees, finds the AI buildout running at full throttle despite credit market jitters.
On the IT side, optics stole the show: Nvidia and Cisco introduced the USFP module, and co-packaged optics moved toward the compute domain, letting Nvidia's NVL576 span eight racks with copper inside each rack and optics between them.
For physical infrastructure, the cold plate took the spotlight from CDUs, though multi-megawatt CDUs are now standard.
The 800 Vdc power sidecar arrived in force, solid-state transformers drew crowds, and OCP sees a path to 1MW racks.
Evaporative cooling uses a lot of water, but lower upfront cost and better efficiency keep it in use, especially in dry climates.
Water use now rivals power as a community flashpoint for new data centers.
One estimate cited in the article puts a 1MW facility using evaporative cooling at 20,000 to 50,000 gallons of water a day, or 20 million to 50 million gallons for a 1GW site.
Operators still choose evaporative systems because they reject heat more efficiently than dry cooling and cost less to build, with one HVAC executive estimating that open-loop systems cost 60% to 70% less upfront than comparable closed-loop designs.
The article weighs these tradeoffs against water availability, climate and community acceptance.
Infosys analysis: efficiency and flexible AI workloads buy near-term capacity but will not remove the need for new grid infrastructure.
An Infosys analysis of AI energy demand through 2030 concludes that efficiency improvements, advanced cooling and shifting flexible workloads can create near-term headroom for AI, but are more likely to delay than eliminate the need for new generation and transmission.
Infosys recommends a dual-track strategy: get the most from existing capacity while speeding up grid investment.
Training, retraining and batch jobs can often be scheduled around grid conditions, while real-time inference cannot, so operators cannot treat their whole AI load as curtailable.
Wider flexibility will need SLAs that match workload requirements, plus demand-response incentives and pricing.
More from Operations · All IT section newsletters · This week