Explore QiO's Insights and Innovations

Your servers burn up to 30% of their power budget doing nothing. Whilst your team chases renewable contracts and cooling upgrades, idle CPUs are quietly consuming capacity you need for growth. CPU power-state optimisation delivers immediate savings from existing hardware: no new builds, no refresh cycles, no waiting.

Data centre managers face a familiar paradox: sustainability roadmaps are lengthening whilst power constraints are tightening. Your board wants aggressive carbon targets. Finance wants lower OpEx. Meanwhile, AI workloads and digital transformation are queuing up for capacity you don't have.

The conventional approach -renewables contracts, PUE optimisation, next-generation cooling - remains essential. But these infrastructure investments are capital-intensive, take years to deploy, and often deliver diminishing returns. A more immediate decarbonisation lever sits hidden in plain sight: the CPUs already in your racks.

The problem hiding inside your servers

Here's what most capacity plans miss: a server sitting idle still burns 40–70% of its peak power. [1] Research confirms that even at single-digit utilisation, conventional servers draw roughly half their maximum load. More recent designs have improved at the extreme low end: true idle consumption has fallen from 51% of peak in 2014 to around 36% today. [2] But “better” is not the same as “fixed" because the power curve remains stubbornly non-linear at low utilisation.

Even with modern servers, a drop to ~10% utilisation doesn’t translate into anything like a 90% reduction in power. Consumption often remains closer to ~50% of peak, meaning you’re still paying for electricity and emitting carbon for processing capacity you’re barely using. [2,3] Across a typical enterprise or colo fleet averaging 20–30% CPU utilisation, the idle overhead compounds into a substantial and largely hidden waste stream.

For data centre leaders juggling reliability, cost and carbon mandates, this idle draw functions as a growth tax. It consumes power headroom needed for resilience buffers, hardware refresh cycles and new workload onboarding. In colo environments with contracted power caps, it directly limits expansion. Even worse, because it doesn't show up as a utilisation problem, it rarely triggers action until the electricity bill forces a reckoning.

Why traditional efficiency measures can't solve this

While PUE improvements target everything around the server - chillers, airflow, lighting - they can't address inefficiency inside the chip. A facility with a stellar PUE of 1.1 still wastes watts if its CPUs are drawing 60% power at 5% load. Renewables reduce the carbon intensity of your energy, but they don't reduce the amount you're consuming. In markets with constrained grid capacity or time-of-use pricing, that distinction matters.

Virtualisation and consolidation help by redistributing workloads, but you're still running servers that burn energy disproportionate to their output.[4] Without addressing the power-to-utilisation mismatch at source, these strategies simply redistribute the problem rather than eliminate it.

The at-source solution: CPU power-state optimisation

Modern x86 processors support fine-grained power management through P-states and C-states.[5,6] When configured correctly, these mechanisms allow CPUs to scale power consumption much more closely with actual demand.

The impact is significant. Intel and QiO jointly validated results showing up to 52% idle power reduction and approximately 24% savings under representative load.[7] Breakthrough research from Huawei and ETH Zurich introduced enhanced deep C-states that cut idle power to just 5-7% of active draw, reducing energy consumption for Memcached workloads by up to 71%, with less than 1% performance degradation.[8,9]

In practice, realistic fleet-wide savings fall in the 20–30% range for overall data centre CPU energy consumption. That's achievable with existing hardware, no hardware replacement required.

Immediate deployment, zero disruption

Changes can be deployed via firmware and infrastructure management systems without racking new hardware, avoiding lengthy procurement cycles. [10,11] In live operations, power-state optimisation is a balancing act. Deeper savings can introduce small increases in wake latency as CPUs exit low-power states, typically measured in microseconds, so configurations should be set to match the latency tolerance of each workload and SLA. You’re tuning assets already deployed, not financing new builds. And because lower server power reduces heat output, it cuts cooling demand too, improving PUE and extending UPS runtime, creating a compounding efficiency gain across the infrastructure stack.

For data centre leaders under pressure to deliver more with less, this is the rare initiative that cuts costs, frees capacity and accelerates decarbonisation simultaneously. It's demand reduction at source, not offsetting down the line.  

Start with what's already plugged in

The narrative that renewables and PUE solve everything no longer holds in an era of constrained power and accelerating demand. CPU power-state optimisation is emerging as a critical efficiency enabler and, if neglected, a hidden cost multiplier.

The path to net-zero shouldn't begin with what you'll build in 2027. It should start with optimising what's already plugged in. CPU power-state management isn't a substitute for renewables or infrastructure modernisation; it's the foundation that makes those investments more effective by shrinking the baseline load they need to serve.

For data centre operators, aligning server-level power management with capacity and carbon objectives is no longer optional. It's a prerequisite for unlocking the full return on infrastructure investments and ensuring long-term operational, financial and environmental sustainability.

In a world where power is the new square footage, idle servers are an unaffordable luxury. The largest untapped lever in data centre decarbonisation isn't outside your facility; it's inside every rack.

References

[1]: Meisner, D. et al. PowerNap: Eliminating Server Idle Power, Stanford University.
[2]: (2023) Power Proportionality and Idle Power Consumption of Servers, David Kopp Notes.
[3]: (2024) What is the Relationship Between Server Utilization Rates and Energy Consumption?, Sustainability Directory.
[4]: Sharma, S. (2016) Trends in Server Efficiency and Power Usage in Data Centers, SPEC.
[5]: (2024) Understanding the Concept of CPU Power States, Livewire Development.
[6]: Processor P-states and C-states, Thomas-Krenn.
[7]: (2023) Power Management: Leveraging AI for Smarter Data Center Power Efficiency, Intel Network Builders.
[8]: (2023) AgileWatts: Sustainable Server Design for the Modern Data Center Era, arXiv (Huawei/ETH Zurich).
[9]: (2022) AgileWatts: Sustainable Server Design for the Modern Data Center Era, IEEE Xplore.
[10]: (2023) Enhanced Power Management for Low Latency Workloads Technology Guide, Intel Network Builders.
[11]: (2024) Power Management and Energy Efficiency in Data Centers, Cisco Live.

ServerOptix cuts server energy consumption by 25% with zero impact on performance. Every month you don’t act on it, you’re paying for power your servers don’t need.

Speaking to the BBC’s Today programme, Raspberry Pi founder Eben Upton has named high energy prices as the single biggest threat to UK manufacturing [1]. Data centres face the same threat in sharper form. UK industrial electricity is more than four times the US level [2], and no operating model is more exposed to that input than a data centre. For UK operators, this isn’t news. What’s new is hearing it said out loud by someone outside the sector.

OpenAI’s recent £31bn Stargate pause [3] is what the squeeze looks like at the top end of the market. With UK demand set to quadruple by 2030 [4] and grid connections taking up to a decade [5], the next efficiency gain has to come from inside the existing estate. Building out won’t be an option for most. Optimising what’s already there will.

The next big saving is in the kit itself. The industry has spent a decade on everything around the servers, from renewables to PPAs to cooling. While the servers themselves have barely been touched.

Servers draw power decoupled from the work they do. Most pull 50 to 70% of peak even at idle, while average CPU utilisation sits at 15 to 30%. That gap is energy you pay for and get nothing back. It’s the energy equivalent of a hotel keeping the lights and air-con on in every room irrespective of whether they’re occupied.

What the evidence shows

ServerOptix closes that gap at the CPU, autonomously, without touching performance. Installation takes hours, runs on existing hardware, and sits in a category the industry hasn’t yet named: autonomous server-level energy optimisation. Three independent studies have measured what it does.

Qnetix recorded a sustained 35% reduction in server power on a live estate, around 62W per server, with no perceptible performance impact, all independently confirmed at rack-level PDU [6]. Across a 2,000-server estate that’s roughly £260,000 in annual energy savings and around 192 tonnes of CO₂e avoided. The two-week test ran across 50 servers on live applications, workloads and SLAs remained unchanged.

The result holds across both major chipset architectures. The estate included AMD EPYC and Intel Xeon processors on multiple Dell PowerEdge models, so this isn’t a single-vendor effect.

What the lab tests confirm

Two lab studies back up the Qnetix result. World Wide Technology's Advanced Technology Centre recorded 19 to 29% power reductions on Dell servers, with exhaust temperatures down 9.2%, all independently confirmed at the PDU [7]. Intel separately validated reductions of approximately 25% on Xeon Gold dual-socket servers under representative load [8].

All three studies converge: measurable, sustained energy reduction with no performance impact. While the headline 25% is the conservative average across this body of research, the Qnetix study shows what’s achievable at the upper end.

What this means for your facility

The savings don’t stop at the meter. Lower power draw at the rack cuts cooling demand and frees electrical headroom for growth, without expanding the envelope. In a market where a new grid connection takes up to a decade, that headroom is its own asset. For operators, that headroom is what makes the next workload land without going back to the planners.

Three independent studies, two chipset architectures, one conclusion: your servers are drawing energy they don’t need. There’s now a proven way to stop paying for it.

References

1. (2026) Eben Upton: high energy costs the biggest threat to UK manufacturing, Today programme, BBC Radio 4.

2. (2026) The rising challenge of powering data centres, Oxford Economics.

3. (2026) OpenAI pauses Stargate UK data centre over energy costs and regulation, Reuters via Yahoo Finance.

4. (2026) Data centres: planning policy, sustainability, and resilience, House of Commons Library.

5. (2026) 50 GW of datacentre demand queues up for UK grid access, The Register.

6. (2026) Validating scalable server energy optimisation in real-world data centre conditions, by Qnetix, published by QiO Technologies.

7. (2023) Using AI to reduce energy consumption, cost and carbon emissions in data centres, World Wide Technology Advanced Technology Centre.

8. (2022) Power management: leveraging AI for smarter data centre power efficiency, Intel Network Builders.

‍

Fitting the system beats rebuilding it every time. While autonomous server energy optimisation delivers 19 to 29% power savings without new hardware, grid upgrades or major retrofits, the word “autonomous” prompts four key questions DC leaders should be ready to answer.

Many approaches to cutting data centre energy costs involve spending money to save money. New cooling systems. Hardware refreshes. Renewable energy contracts. They all work, but they all take time, capital and sign-off from people who’d rather not spend money of that magnitude. And with projects like these, the return on investment is typically measured in years, not months.

Against this backdrop it’s easy to see why autonomous server energy optimisation is fast attracting interest as an alternative proposition. It fits into your existing infrastructure. No new hardware. No construction. No waiting for grid connections. And critically it uses what you already have, to deliver proven energy savings of 19 to 29% server power reduction under varying loads [1] and up to 25% savings under representative workload [2]. And because solutions like these typically run on a SaaS model, the ROI has to deliver within months, not years. Because if it doesn’t deliver, you don’t renew.

While that’s an attractive proposition the word “autonomous” can make leaders uneasy for understandable reasons. If you are preparing to brief your board on autonomous server energy optimisation here are four questions the board are likely to ask.

4 Questions a Board Will Ask About Autonomous Energy Optimisation

1. “What if it causes an outage?”

Your SRE and reliability teams will raise this first, and rightly so. The benchmark to look for is a hard, automatic threshold. A well-designed autonomous optimisation tool won’t wait for a human to intervene. If CPU utilisation breaches a set limit, the optimisation system shuts itself off and reverts the server to its default settings. No delay. No escalation required. That’s the difference between genuine autonomy and tools that just make recommendations. The guardrail isn’t bolted on afterwards. It’s how the technology works.

2. “How do we know what it’s doing, and can we audit it?”

Ops, engineering and compliance all want the same thing – visibility. Any solution worth considering should log every action taken while enabled. Look for a clear record of what was changed and when, accessible to your team without needing to ask the vendor for it. That provides compliance with an evidence trail without the need to maintain a separate register.

3. “Who controls it?”

In any well-run data centre, server-level access sits with the IT team and SREs, governed by role-based access control under ISO 27001. What matters is that those with access can pause or disable the optimisation at any point, and the system reverts cleanly when they do. That operational confidence is what keeps teams running autonomous tools rather than switching them off after the first week.

4. “Who owns the outcomes?”

Finance and sustainability will want to know who reports the savings and who is accountable if something goes wrong. The answer is a lightweight governance framework, not a six-month programme. Three tools should do the job. An ownership matrix that maps decision domains with no ambiguity on accountability. Second, an incident review loop for regular assessment of outcomes and near-misses. And third, a review of the system’s own action logs as the evidence base.

The bigger picture

None of these questions should kill the conversation. They all have evidence-backed answers [1][2][4], and that’s what makes autonomous optimisation worth serious consideration. When evaluating solutions, the things to insist on are built-in thresholds, automatic rollback and full action logging as standard.

In the current geopolitical climate, data centre energy costs won’t be going down for some time, and neither will grid constraints ease. The organisations that act first on fitting rather than rebuilding will be the ones with headroom when it matters most.

If you’d like to see what autonomous energy optimisation could deliver across your estate, we’re happy to walk you through the evidence and explore whether it’s a fit for your environment.

References

1. (2023) Using AI to Reduce Energy Consumption, Cost and Carbon Emissions in Data Centres, WWT.

2. (2022) Power Management: Leveraging AI for Smarter Data Centre Power Efficiency, Intel / Network Builders.

3. (2017) Know When to Use Open- or Closed-Loop Control, Control Engineering.

4. (2026) Data Centre Energy Optimisation at Scale: Why Manual Tuning Can’t Keep Up, QiO Technologies.

‍

The UK Cyber Security and Resilience Bill will make cyber compliance mandatory for data centres. At the same time, energy costs are pushing operators towards optimisation tools that rely on cloud connections and external APIs. That’s a problem, because every new external dependency widens the attack surface you’ll soon be assessed against. There’s a way to cut server power without creating a compliance liability.

The rules are changing

The UK Cyber Security and Resilience Bill is currently making its way through parliament and will bring data centres into scope as essential services for the first time. Standalone facilities above 1 MW and enterprise sites above 10 MW will face mandatory cyber security and operational resilience requirements, with Ofcom as the regulator. Penalties for non-compliance run up to £17 million or 4% of global turnover, plus daily fines of up to £100,000. If you operate in the EU, NIS2 and DORA are already in force. This isn’t coming over the horizon. It’s here.

The energy problem hasn’t gone away

At the same time, energy consumption is rising, regulation is tightening [1][3], and costs are soaring. It makes sense that data centre leaders are turning to tools to optimise energy management. Unfortunately, most of those tools need a cloud connection, an external API, or a hosted pipeline to work. In an environment where 74% of large UK businesses were hit by a cyber breach last year and supply-chain compromises have doubled [2], each new external dependency is something you’ll need to account for under the incoming regime.

SolarWinds showed us what happens when infrastructure tools with deep network access phone home. Privileged access to everything, and its outbound connection became the entry point for one of the worst supply-chain attacks on record.

Optimisation without the exposure

But what if the optimisation tool didn’t need an outbound connection at all? Design it to run entirely on-prem, by default, and you remove the attack surface rather than trying to manage it. Everything stays inside the data centre boundary: the telemetry collection, the AI that decides what to do, and the control actions themselves. The system monitors workload levels only, not what that workload is. Nothing leaves your network unless you choose it to.

For example, each instance of ServerOptix includes a local dashboard on the server where it’s deployed, so you can monitor savings without any external connection. If you’re running multiple instances, whether across servers in a single facility or across multiple data centres, an optional cloud-based dashboard lets you aggregate the data in one view rather than checking each instance individually. Either way, the control loop never depends on a cloud connection.

It works in air-gapped environments too. No internet connection required. And rather than a dashboard that tells you what to do and hopes someone acts on it, a closed-loop system senses workload, adjusts CPU power states, and rolls back if performance dips.

Real savings, independently tested

WWT independently measured 19 to 29% power reductions at the PDU, not from onboard telemetry, across Dell R650 and R750 servers. Exhaust temperatures dropped by over 9% [4]. Intel proof-of-concept testing showed up to 25% savings under representative workload and roughly 52% at idle [5]. Live customer deployments have recorded average savings above 29%.

At 2,000-server scale, that’s around 1.146 GWh off your energy bill. The associated carbon savings depend on local grid carbon intensity: in the US, that equates to over 410 tonnes of CO2 saved annually; in the UK, approximately 202 tonnes; and in a low-carbon grid like France, around 30 tonnes. Depending on local energy prices, typical payback periods fall within eight to twelve months.

Built for the compliance environment that’s hoving into view

The energy savings matter, but they only count if the tool can get deployed. That means passing muster with your security and compliance teams. On that front: no new external firewall rules required (depending on your network configuration, internal rules may need updating). No new external access paths unless you opt into optional dashboarding. It sits inside your current network segmentation and is designed to work within existing zero-trust architectures, not against them. All telemetry stays within your monitoring boundary by default. It can contribute to EU EED reporting by providing data on IT-specific power consumption across the servers where it’s deployed, though full reporting coverage would require deployment across all machines in scope [3]. And there’s no external vendor in the data path, which addresses supply-chain risk requirements under NIS2 and the UK Cyber Security and Resilience Bill.

Energy savings that don’t come with a compliance headache

You shouldn’t have to choose between cutting energy costs and staying on the right side of incoming regulation. A fully on-prem, air-gap-ready, closed-loop system lets you drive continuous power reduction without widening your attack surface. With the Cyber Security and Resilience Bill on its way, the smartest time to reduce your energy costs is before the compliance burden lands on top of them.

References

[1] IEA (2025) Energy and AI, International Energy Agency. https://www.iea.org/reports/energy-and-ai

[2] DSIT/Home Office (2025) Cyber Security Breaches Survey 2025, UK Government. https://www.gov.uk/government/statistics/cyber-security-breaches-survey-2025

[3] European Commission (2025) Energy performance of data centres, EU Energy Efficiency Directive. https://energy.ec.europa.eu/topics/energy-efficiency/energy-efficiency-targets-directive-and-rules/energy-efficiency-directive/energy-performance-data-centres_en

[4] WWT (2023) Using AI to reduce energy consumption, cost and carbon emissions in data centres, World Wide Technology. https://www.wwt.com/article/using-ai-to-reduce-energy-consumption-cost-and-carbon-emissions-in-data-centers

[5] Intel/QiO (2022) Power management: leveraging AI for smarter data centre power efficiency (solution brief), Intel Network Builders. https://builders.intel.com/docs/networkbuilders/power-management-leveraging-ai-for-smarter-data-center-power-efficiency-solution-brief-1664790296.pdf

[6] European Commission (2025) Assessment of next steps to promote the energy performance and sustainability of data centres in EU, including the establishment of an EU-wide rating scheme, October 2025.

‍

Every stream you watch, every search you run, every message you send passes through a data centre. These buildings keep modern life ticking. They also use a lot of electricity doing it. So, here’s a question worth asking on International Data Centre Day: where does the next big energy saving come from? And what part does AI play?

The score everyone chases - shrinking the 1

The data centre industry measures efficiency using something called PUE – Power Usage Effectiveness. The maths is simple enough. Take all the power a data centre uses (servers, cooling, lighting, networking, the lot) and divide it by the power used by the IT equipment alone. A perfect score is 1.0. You’ll never get there, but that’s the goal.

Twenty years ago, the average PUE sat at roughly 2.5. Today it’s closer to 1.4. That’s a big improvement and the industry has worked hard to get there.

How did they do it?

By going after 'everything else.' The cooling, the power systems, the lighting. Operators redesigned airflow to stop hot and cold air mixing. They swapped old UPS systems for high-efficiency ones to cut conversion losses. They put lighting on motion sensors so empty halls stopped burning power for nobody. Each change chipped away at the overhead and nudged that PUE score closer to 1.

But here’s the thing

Most of the easy wins have gone. The low-hanging fruit on the “everything else” side of the equation has been well and truly picked. Cooling and infrastructure have been optimised about as far as they can go with conventional approaches. So where do you look next?

The fruit nobody thought to pick

Here’s what almost nobody is looking at: the IT power itself. The servers. The kit doing the actual useful work. It’s the “1” in the PUE equation and it’s been treated as untouchable for decades. The thinking? That’s the sacred bit. That’s revenue. That’s the workload. Don’t touch it.

Which means there’s an enormous opportunity hiding in plain sight.

What if AI could touch the untouchable and cut energy bills?

AI-enabled technology now makes it possible to optimise how much energy servers use, without affecting performance. QiO’s ServerOptix does exactly this. It uses AI to analyse server power consumption in real time, matching energy use to actual demand rather than letting servers run flat out whether they need to or not.

Does it work? Organisations using ServerOptix are seeing around 25% savings on IT power.^[1][2] And because servers running more efficiently generate less heat, there’s a knock-on reduction of roughly 15% on cooling costs too.^[1]It’s quick to deploy and the savings start straight away.

It’s time to shrink the bills, not the 1

PUE as a metric has its problems and that’s well understood. Reducing IT power - the “1” - can actually make your PUE score look worse. A smaller denominator means the ratio goes up, even though you’re genuinely using less electricity. On paper, it looks like you’ve gone backwards.

Other metrics exist. But the mindset of chasing the 1 is so deeply engrained that it still shapes how the industry thinks about saving energy. Which means there’s a massive opportunity that most people simply haven’t come across yet: optimising the IT power itself.

Would you rather protect a PUE score on a slide, or take tens of thousands of pounds off your electricity bill? It’s an opportunity that’s been overlooked for too long. AI now makes variable energy optimisation possible, in real time, without compromising performance. There’s nothing stopping you from grabbing it.

Ready to stop chasing the 1 and start cutting the bills? Talk to QiO.

References

[1] (2024) ServerOptix lab evaluation, World Wide Technology (WWT)

[2] (2022) Leveraging AI for smarter data centre power efficiency, Intel

‍

Fear that automating optimisation will take away control from DC managers is understandable. But with the proper guardrails in place, you’ll experience a better kind of control. One you can trust.  

I find the psychology behind the adoption of autonomous optimisation energy fascinating. Most data centre managers are waste-aware. Have dashboards and reports telling them where efficiency can improve. As well as evidence that automating optimisation (AO) increases efficiency. Yet fear that automation means handing over control.

I understand why letting a system make real-time decisions feels risky. Especially when your reputation and bonus depend on reliability and uptime. But with the right software in place, instead of losing control, users get a better kind of control and with that peace of mind. Let’s look at the facts in more detail

Visibility isn't control

Many DC managers are drowning in waste data. Your dashboards light up like Vegas with inefficiencies. But seeing a problem and fixing it are two different things.

Visibility doesn't stop drift. Manual tuning delivers a burst of savings, then tickets pile up and you're back where you started. Workloads shift by the minute, estates get denser, and last quarter's settings become today's liability.

The fact remains that the cheapest, cleanest unit of energy will always be the one you don't spend.

Power as a reliability discipline

Think about reliability. It's a feedback loop: sense, decide, act, verify. That's how we should manage energy. Too many organisations rely on human intervention to make adjustments leaving them stuck in “measure and recommend mode”. Contrast this approach with closed-loop control which adjusts CPU power settings in real time while protecting your service targets.

Research shows that a closed-loop optimiser creates repeatable outcomes, enforced within guardrails because it pulls in workload signals, power telemetry, and service indicators like P95 latency. Applies voltage and frequency scaling, power caps, and sleep states continuously, with verification.^[2] Enabling you to reduce power whilst staying inside latency constraints.^[3] You define what it can do, how fast it moves, and when it rolls back.^[2]

Evidence you can defend

Let’s take a closer look at the evidence. House-hold name, WWT reported power reductions of 19 to 29% across fixed and variable loads, validated via PDU readings.^[1] Intel and QiO's testing shows up to 52.61% lower power for idle servers and 24.78% lower under real-world load. ^[2]. Measurements leaders can trust rather than taking a leap of faith on projections.

Headroom is the real prize

Today, power is the leading capacity constraint. Peaks force worst-case design, stranding paid-for capacity. Closed-loop control smooths demand so you run closer to average, safely. In practice, for a 2,000-server estate at 26p/kWh, a 25% reduction enables roughly 667 additional servers within the same power envelope.^[6]

And of course, lower server power means lower heat which requires less spend on cooling systems. Our own research shows customers enjoy around 15% indirect savings on top of 25% direct savings.^[6]

Governance first, then scale

Here’s where trust becomes critical. Organisational trust in AO is earned or lost in the detail. So what does AO mean for guardrails?

It's setting that safe operating envelope. You're still in control, just differently. AO can’t break anything because the permissions you grant it mean it only acts within those parameters. Research reinforces that workload-aware controls protect latency-sensitive services.^[4] Ensure you explicitly define those guardrails including latency and error budgets, power caps, rate limits, and automated rollback triggers.

If needs be, start small on low-risk workloads. Prove compliance and stability. Then expand when it's boring. Your dashboards show the gap. Closed-loop control closes it, safely, every day.^[4]

Focus on getting the right guardrails in place to build trust in AO because discussion at the India AI impact summit makes one thing clear, the question is no longer whether a DC can afford to automate but whether it can afford not to.

References

1. (2023) Using AI to Reduce Energy Consumption, Cost and Carbon Emissions in Data Centres, World Wide Technology.
2. (2022) Power Management – Leveraging AI for Smarter Data Centre Power Efficiency (Solution Brief), Intel.
3. (2024) Leveraging Core and Uncore Frequency Scaling for Power-Efficient Serverless Workflows, arXiv.
4. (2024) PADS: Power Budgeting with Diagonal Scaling for Performance-Aware Cloud Workloads (IGSC 2024 record), dblp.
5. (2024) Simulator-based Reinforcement Learning for Data Centre Cooling Optimization, Engineering at Meta.
6. (2026) ServerOptix by QiO, QiO Technologies.

‍

Modern data centres are drowning in energy data but starving for decisions. When telemetry updates every few seconds and conditions shift minute by minute, dashboards and recommendations can't keep up across a multi-site estate. For sustained cost and carbon reductions without gambling on reliability, you need to shift from visibility to governed, closed-loop action.

The challenge facing data centres isn't gathering energy data anymore. It's making decisions at the speed the data demands.

Telemetry arrives every few seconds: server power draw, utilisation, temperatures, tariff shifts, grid carbon intensity, maintenance states. When you're sampling that frequently, you're dealing with control, not reporting. If humans are still the control loop, the data's arriving faster than we can safely act on it.

When telemetry updates every 10 seconds, humans become the bottleneck

This is why so many energy programmes hit a wall at visibility. Dashboards, alerts, recommendations are useful, but they're open-loop by design. They surface insight, then wait for a person to execute. Insights don't drive change. Action does. At estate scale, that execution doesn't happen at the pace the environment changes.

Workloads are bursty. A host can jump from 40% to 60% utilisation as multiple services shift demand. Meanwhile, thermal headroom, redundancy states, maintenance windows, pricing signals and carbon intensity all change around it. By the time someone reacts, the state's already moved on.

Add in the operational reality of understaffed teams, shift handovers, and sensible caution about changes that might trigger an incident, and manual tuning becomes lots of small adjustments, lots of checking, but not much sustained improvement.

This breaks down fastest in multi-site estates and colocation environments. Each site has its own tooling, procedures, and tolerance for change. Outcomes vary, and the estate never converges on best practice.

Open-loop vs closed-loop: advice vs outcomes

The real question isn't whether you can see what's happening. It's whether you can act on it safely, consistently and at scale.

Autonomous optimisation doesn't mean hands-off. It means governed automation where systems sense, decide and act within explicit policies that protect reliability. In control engineering terms, open-loop systems inform; closed-loop systems correct continuously based on feedback. In data centres, where disturbances are the norm, open-loop advice can't compensate fast enough. Closed-loop control can, provided it's properly bounded.[3]

Evidence that action beats recommendation

In a lab evaluation by World Wide Technology, closed-loop optimisation reduced average server power draw by 19 to 23% under steady loads and 27 to 29% under varying loads, with immediate effect once activated.[1] WWT validated power readings against PDU data, not just onboard telemetry, separating measurement from vendor claims.

An Intel solution brief showed similar patterns: up to 25% lower power under representative workload and roughly 53% lower power at idle with closed-loop control enabled.[2]

These are controlled tests, not fleet-wide production benchmarks. But they reinforce a simple point: when optimisation executes automatically within guardrails, savings show up quickly and repeatably, especially when conditions keep changing.

Energy savings land across the organisation. Finance sees cost reduction. Sustainability sees carbon reduction. Operations and SRE teams see power headroom, meaning capacity freed within the existing envelope, which can defer upgrades and reduce risk.

If your current approach ends at visibility, you've optimised reporting, not outcomes. The shift to make now is treating energy optimisation as a governed, closed-loop control problem. Ask yourself one practical question, where are humans still expected to make sub-minute optimisation decisions in an environment that changes faster than they can safely respond? That's where autonomy stops being optional.

References

[1] (2023) Using AI to Reduce Energy Consumption, Cost and Carbon Emissions in Data Centres, World Wide Technology.

[2] (2022) Power Management: Leveraging AI for Smarter Data Centre Power Efficiency (Solution Brief), Intel (Network Builders).

[3] (2017) Know when to use open- or closed-loop control, Control Engineering.

‍

Adopting AI-supported Automated Industrial Operations (AIO) should make clear commercial sense, yet many manufacturers still meet understandable human resistance. This article explores the main causes and sets out five guiding principles to introduce and scale AIO while bringing your people with you.

We know that technology delivers only when people engage with it, and AI-supported AIO is no different. Despite its proven ability to lift OEE, cut energy and reduce errors without adding risk, adoption is uneven. The real barrier is not a generic fear of AI but fragile trust that shows up in four recurring concerns from engineers: opacity, control, rigidity and job security. [1][2][4]

First, opacity. People trust what they can follow. If a model cannot explain its decisions or clashes with process knowledge, trust quickly weakens. One visible mistake can outweigh months of quiet success. Clear explanations, simple guardrails and shared understanding of limits are vital to avoid the sense of a “black box”. [6][7][9][8]

Second, control. When systems act on their own, people fear losing control over work that defines their reputation. If operators feel they cannot safely intervene, hesitation is inevitable. Roll-out must keep humans in real decisions, with clear, respected override routes and visible support when they act. Control and reversibility build trust. [10]

Third, rigidity. If staff think the system runs on fixed rules that ignore the quirks of their line, they switch off. “Our plant is different” really means “this system will not listen or learn”. Showing how models are updated, and involving operators in thresholds and exceptions, proves it can adapt. [3]

Fourth, job security. With constant headlines about automation replacing jobs, it is natural for experienced operators to worry. In reality, roles tend to change more than disappear. Early AI needs human boundaries and a “sniff test” on decisions. Done well, AIO is clear about role changes, reskilling and fair sharing of efficiency gains. [5][8]

Together, these concerns call for an engagement plan that shows how staff stay involved, how decisions are governed and control modes change, and that AIO is done with people, not to them, building trust.

Five guiding principles of an effective engagement plan

What works will vary from one organisation to another, but research highlights five principles that best-practice organisations have used to build trust at pace without incurring unnecessary risk.

1. Treat change as deliberately as you treat safety processes. Start with short shadow pilots in which AI recommends and people decide. Reframe roles so that operators handle real-time exceptions and supervisors are accountable for how the AI is configured, monitored and improved. Report safety, quality and energy against current practice and publish the results, including the problems. After any error, use a simple review so everyone understands what happened, what changed and how risk is reduced. This shows that AIO is governed, not experimental. [2][3]
2. Be explicit about control modes and earn the right to move up. Start with human in the loop, where AI advises and people decide, then move to human on the loop only when you have clear stop rules, safe states and proven reliability. Keep human out of the loop for certified cases only, with automatic fallback. Earn each step with stable performance and drilled rollbacks. People trust systems they can stop and restart. [9][10]
3. Make accountability visible. Use a recognised AI management approach, such as ISO/IEC 42001, and align with the EU AI Act where it matters. Industry leaders like IBM show how ethics frameworks can blend governance structures, human oversight and technical guardrails across the lifecycle; you can adapt these rather than start from scratch. [14] Build a simple decision matrix and living model register so ownership, limits and incidents are clear and auditable. This shows people that AI is genuinely governed, not experimental, and that accountability is real. [12][11]
4. Measure what matters to the plant and challenge the numbers. Link model performance to OEE, energy per unit, first-time-right, scrap and rework, and near-misses avoided. Track intervention rate, drift and the false-positive load placed on crews. Publish a short quarterly report on AI controls with trends, not just snapshots. Mark the wins, and record where human judgement prevented a loss, then feed those lessons back into the model. This reinforces the message that human oversight is valued, not bypassed. [13]
5. Invest in people and build a culture that earns trust. Train teams on confidence bands, override triggers and structured feedback, and do not penalise sensible overrides. Adoption rises when people see that preventing a mistake is valued. Policies and governance frameworks matter, but if daily behaviour contradicts them, trust will still erode.In mature organisations, ethical use of AI is part of culture, not just compliance. IBM, for example, couples formal AI governance frameworks with everyday norms that encourage multidisciplinary review, open challenge and continuous feedback on model behaviour. [14] Co-design guardrails with operators so they see their expertise reflected in the system, building better thresholds, faster learning and stronger buy-in. Making skills, judgement and collaboration visible is one of the clearest signals that the organisation is serious about fair, human-centred automation. [3][5]

Five incentives that unlock engagement with AI-supported AIO

If you want that culture of trust to last, you’ll need to reinforce staff engagement. People are more willing to back AI when the benefits feel clear, fair and under their control, and when they know they can step in without being punished if something looks wrong. Remember, incentives don’t have to be financial and might include:

• Piecework booster: spell out that higher, steadier throughput and first-time-right lifts piece-rate pay, making AI feel like immediate personal gain.
• Team gainshare on plant KPIs: quarterly bonus on OEE, energy per unit and scrap, with no mid-period ratchets, signalling fair, transparent rewards.
• No-penalty overrides: formal policy that sensible overrides never hurt pay or ranking, protecting agency and safety.
• Skills pay and certification: pay uplifts for “AI Oversight” levels (confidence bands, drift spotting), showing investment in people, not just machines.
• Clear control modes and stop rules: progress from HITL to HOTL only after incident-free hours and drilled rollbacks, proving control and reversibility.

Ultimately, the challenge isn’t “making AI smarter” but steadily strengthening fragile human trust. If you show people how decisions are made, when and how they can intervene, what will happen to their roles and how accountability really works, AIO shifts from a source of anxiety to a system they willingly engage with.

References

[1] (2024) The challenges preventing AI adoption in manufacturing, The Manufacturer.
[2] (2024) Future Factories Powered by AI, Make UK.
[3] (2024) The use of AI and ML in process plant operation and control, IChemE.
[4] (2024) Generative AI and human decisions in high-tech manufacturing, The Manufacturer.
[5] (2023) Psychological factors underlying attitudes toward AI tools, Nature Human Behaviour.
[6] (2024) Trust in AI: progress, challenges, and future directions, Humanities and Social Sciences Communications.
[7] (2024) Trust, trustworthiness and AI governance, Scientific Reports.
[8] (2024) Trust and reliance on AI: an experimental study on the extent and costs of overreliance on AI, Computers in Human Behavior.
[9] (2024) Effective human oversight of AI-based systems: a signal detection perspective on the detection of inaccurate and unfair outputs, Minds and Machines.
[10] (2024) Institutionalised distrust and human oversight of AI, AI and Society.
[11] (2024) Regulation (EU) 2024/1689 (EU AI Act), Official Journal of the European Union.
[12] (2023) ISO/IEC 42001: Artificial Intelligence Management System, International Organization for Standardization.
[13] (2024) How AI can build a more sustainable future for businesses, edie.
[14] (2025) Trustworthy AI at scale: IBM's AI safety and governance framework, IBM.

‍

Prescriptive decision-making has transformed maintenance management, output optimisation and sustainability from cost-centre activities into a strategic competitive advantage. Yet despite clear data showing that adopting prescriptive systems can significantly increase output and sustainability while reducing downtime and errors, many boards have yet to grasp the full business implications of predictive, let alone prescriptive, capability. This article explores why that gap persists and offers a framework to help leaders assess where they stand on the journey.

For manufacturers operating at the sharp end of energy and resource volatility, data-driven decision-making has moved from an engineering concept to a boardroom differentiator. In energy-intensive industries, where a single unplanned shutdown can erode millions in EBITDA, predictive and prescriptive systems have proven their value.  

Specifically, evidence from multiple sectors shows that moving from reactive to predictive maintenance can deliver up to a 75% reduction in unplanned downtime, a 60% reduction in maintenance costs, and a 40% increase in equipment lifespan, with a direct EBITDA uplift through increased throughput and avoided lost production [1][2]. In one widely cited case, Duke Energy detected a turbine anomaly before failure, avoiding catastrophic damage and saving 34 million US dollars in lost generation and repair costs [3].

These aren’t marginal gains. Rather they represent a structural shift in how industrial assets create value, extending capex cycles, enhancing utilisation, and improving sustainability performance through reduced waste and energy intensity. Instead of seeing maintenance as a cost centre it becomes a competitive advantage. Yet many boards still see predictive analytics as a technical initiative rather than a strategic lever. Full value can’t be realised until leaders connect digital capability directly to financial and sustainability outcomes.

Why predictive remains misunderstood

Most senior executives understand what predictive analytics does but not what it changes. Predictive capability is often treated as an operational add-on rather than a transformation in how risk, capital efficiency and resilience are managed. Boards that evaluate predictive projects purely through engineering KPIs rather than EBITDA, deferred capex or energy intensity inadvertently cap their return. The result is “pilot purgatory”: technically sound initiatives that never scale because they are not anchored in governance or accountability [4].

Predictive and prescriptive systems alter who makes decisions, how quickly and on what basis. That demands leadership alignment on data ownership, organisation-wide trust in the algorithms and performance incentives. Until boards recognise predictive analytics as a new operating model rather than a new IT platform, transformation will remain stalled.

Understanding the maturity curve

For board leaders, understanding where their organisation sits on the decision-making maturity curve isn’t a technical exercise. It’s a matter of governance, resilience and value creation. Reactive operations rely on experience and manual intervention; problems are discovered only after they occur, creating volatility in output, safety and cost. Predictive operations introduce visibility, allowing early intervention, but often remain a capability rather than a systemic advantage.

Prescriptive operations represent a structural shift, where integrated data across assets, supply chains and external variables allows algorithms to recommend or automatically execute optimal actions. Decision-making becomes continuous, contextual and closed loop.  Producing higher asset productivity, deferred capex and reduced energy intensity [5][6]. As a result, industrial reliability becomes a predictable business outcome.

In this context, maturity in industrial decision-making is emerging as a new indicator of competitiveness. In a volatile energy and resource environment, prescriptive capability can’t be about automation for its own sake.  It must be about embedding consistency, optimisation and learning into the fabric of enterprise performance.

The leadership fault lines

Despite the proven upside, many manufacturers remain stuck between predictive insight and prescriptive action. In this case, progress is typically constrained by a lack of alignment between leadership intent, governance and investment logic, rather than technology [7][8]. Three leadership fault lines consistently distinguish businesses that advance from those that stall: culture, governance and investment.

Many businesses still celebrate firefighting heroics. The engineer who saves the shift, the team that restores production overnight. This mindset rewards reaction over prevention. Boards shape culture through story, and when digital transformation is framed as cost control, it stays in the back office. In contrast, when framed as resilience and competitiveness, it becomes a strategic mission.  

Governance and ownership are equally critical. As algorithms recommend or automate decisions, accountability blurs. Most manufacturers lack formal structures for data trust, algorithmic transparency and cross-functional responsibility. Boards must define governance, embed AI explainability within enterprise risk frameworks and make data trust a standing agenda item, not an IT matter [9].

Finally, digital infrastructure and analytics talent are too often funded as short-term projects rather than enduring capabilities. This limits integration and locks businesses into dependency on external partners. Treating data systems as strategic infrastructure and workforce capability as a competitive moat allows returns to be evaluated over asset lifecycles rather than quarterly cycles.  

These three fault lines, culture, governance and investment, are the true barriers to prescriptive maturity. Technology can be purchased but alignment requires leadership.

Progress through discipline, not disruption

In an uncertain economic climate, few manufacturers can justify wide-scale transformation projects, but progress doesn’t have to be revolutionary. The most successful organisations take an incremental, evidence-led approach, focusing on areas where financial return is clearest and operational risk lowest [10]. Improving existing processes often delivers faster value than disruptive reinvention. For industrial firms, that means layering predictive and prescriptive capability onto existing assets rather than replacing them. Incremental transformation doesn’t signal hesitation. It’s disciplined progress that builds confidence while protecting continuity.

Prescriptive decision-making is reshaping industrial performance, turning data into insight and insight into sustained value. For energy-intensive manufacturers, it offers a route to higher productivity, stronger resilience and measurable sustainability gains. We have the technology, but the differentiator is leadership maturity.  By that I mean, the ability of boards to convert predictive intelligence into governance discipline and operational insight into enterprise advantage. This evolution doesn’t require organisations to rip out and replace what they already have but the intent to collect data and critically empower it to make decisions.    

References

[1] (2023) Predictive Maintenance Using AI to Prevent Equipment Failures, AVEVA Blog.
[2] (2023) Predictive Maintenance: The Future of Manufacturing Productivity, McKinsey & Company.
[3] (2021) Case Study: Duke Energy Predictive Analytics Prevents Turbine Failure, IBM.
[4] (2024) Explainable Predictive Maintenance: A Survey on the Intersection of Predictive Maintenance and Explainable AI (XAI), arXiv.
[5] (2024) AI in Predictive and Prescriptive Maintenance, Siemens / Business Insider.
[6] (2024) Global Prescriptive Analytics Market Report 2024–2033, IMARC Group.
[7] (2023) Digital Transformation in Energy-Intensive Manufacturing, PwC.
[8] (2023) AI Governance and Trust: Board Imperatives for Industry, EY.
[9] (2024) The Industrial Data Maturity Index, Capgemini Research Institute.
[10] (2025) Top Business Intelligence Trends in Manufacturing to Watch in 2025, Moldstud.