News

As AI adoption accelerates, some workers will naturally feel apprehensive. Harvard Business Review found trust in employer-provided generative AI fell by 31% between May and July 2025, despite rising top-down pressure to leverage the technology and boost productivity. This shows a clear gap between leaders’ priorities, and those of their teams.

AI tools have the ability to lighten workloads across business units and drive measurable results: for instance, teams using AI as a core part of their sales functions were 65% more likely to increase win rates.

Yet narratives that frame AI squarely in terms of cost-cutting and headcount reduction are holding some employees back from embracing this technology despite the obvious upside.

The conversation has shifted from speculation to execution, yet many organizations still struggle to move beyond isolated pilots. Closing that gap is a job leadership must prioritize to realize the technology’s potential.

AI initiatives do not fail because the tools themselves are weak or ineffective, but instead because employees do not trust how the technology is being introduced, what new tools mean for their role, or whether they will still have a place once they are embedded.

The real challenge is helping teams overcome their AI apprehension by building their fluency, introducing clear guardrails, carving out time for practical training, and spotlighting use cases that make the solution feel controllable and purposeful, rather than opaque and threatening.

AI is not going to take an employee’s job, but another human who uses it efficiently, and to their own advantage, will.

It will give rise to entirely new roles (prompt engineers, AI ethics officers, AI maintenance specialists) just as digital transformation created functions that barely existed a decade ago. When I worked at a global social media company a decade ago, "social selling" felt abstract. Now it is mainstream, and we are on the same trajectory with AI, only faster.

When leaders talk about AI purely in terms of doing more with less, employees hear threat, not opportunity. People are far less likely to trust AI if they do not trust leadership's intentions behind it. Reassurance cannot come from policy alone, it has to come from behavior.

That starts with a considered view towards building fluency among teams. This is where leadership is responsible for showing employees how AI works in their own roles. Where employees don’t have a clear understanding of how and where they can integrate AI into their workflows, it’s only natural that they will fill this gap with doubt.

Adoption sticks when leaders provide clear guardrails, responsible training and practical examples so AI feels like a technology that amplifies their skills, rather than a threat.

AI adoption cannot succeed if users' roles don’t evolve at the same pace. If people are operating differently to maximize the gains of AI, their individual job scopes can’t be static.

The most effective leaders are proactively redesigning responsibilities to remove low-value drudgery and focus their teams on interpretation, creativity and judgement.

This looks different depending on the business function. In a customer success team, the role must evolve beyond reactive post-sales support, because that wastes the additional business intelligence that AI can deliver. Leaders must recognize that CSMs can do so much more when aided by AI, and reshape their roles so they act as strategic revenue architects.

Using AI to build expertise on specific buyer personas, they can ask sharper, more consultative questions to anticipate their needs and solve problems before a customer realizes they exist. That is a genuine restructuring of the role, not a superficial rebrand.

On the sales side, AI is transforming how managers coach. Rather than sitting beside a rep on a call and writing notes (with all the unconscious bias that can bring), managers can now review an AI-generated brief, apply data-driven scorecards and identify precisely where individuals need support.

They can surface the sales calls where challenging situations were handled best and use real examples as training material. They can see which teams are winning and why, and where processes are breaking down. Coaching stops being as instinctual and subjective and becomes more rooted in data and real outcomes.

While employees might be apprehensive about the implications of AI on their roles, leaders are ultimately being judged on the results their teams drive and moving forward, AI will be a significant multiplier. It’s no longer a question of whether AI can deliver, so the challenge is taking workers on the journey towards understanding how AI will amplify their skills, not replace them.

Encouraging AI adoption in a way that drives real outcomes is one of leaders’ most urgent priorities. Employees do not move from apprehension to agency simply because they are told AI matters. They do it when leaders make the technology understandable, useful and clearly aligned to supporting their roles.

That means demonstrating trust and reshaping how people work, not just demanding that people urgently learn how new tools work.

We've reviewed, rated, and ranked the best HR software.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

The iPhone 7 isn’t remembered as being a particularly important iPhone release, but it did represent several ‘firsts’ for Apple. It was the first iPhone to lose the traditional headphone jack, the first to boast IP67 water resistance, and the first to swap the mechanical Home button for a pressure-sensitive equivalent.

The iPhone 7 was also the first standard-sized iPhone to feature Optical Image Stabilization (OIS), which uses physical gyroscopes to counteract shake-induced blur, and has been a feature of every iPhone released since (that’s 35 models and counting).

Why am I writing about the iPhone 7? Because it arrived almost a decade ago — on September 16, 2016, to be precise — and because I came across TechRadar’s iPhone 7 review sample during a recent clearout of our office cupboard.

I’m currently using the iPhone Air — released on September 19, 2025 — as my daily phone, which is a similarly thin and lightweight iPhone with only one rear camera, and so I thought it would be fun (and nostalgic!) to compare the camera capabilities of these two devices to see how far Apple’s camera hardware has come in 10 years. The answer, as you can imagine, is 'very far'.

Before I jump into the side-by-side comparisons, here’s a table detailing the key camera specs of the iPhone 7 and iPhone Air:

iPhone 7

iPhone Air

Rear camera:

12MP, f/1.8, 28mm

48MP, f/1.6, 26mm

Front-facing camera:

7MP, f/2.2, 32mm

18MP, f/1.9, 20mm

It's worth noting that the iPhone Air defaults to shooting in 24MP, rather than 48MP, via a 'Fusion' process that merges 12 high-dynamic-range pixels and 12 low-dynamic-range pixels into a single 24MP shot. You can choose to shoot in 48MP on new iPhones like the iPhone Air, but I stuck to the default option for this comparison.

Right, onto the side-by-side photos. I took both phones on a walk around the neighborhood, comparing their wide-shooting capabilities, ability to capture color, digital zoom capabilities (neither device has a dedicated telephoto zoom), and low-light shooting capabilities.

First shot: a tree on the sidewalk. This isn't a particularly demanding scenario, but at first glance, both phones appear to have captured a similarly detailed image. Apple's approach to color science doesn't appear to have changed all that much in 10 years, either (at least in this example — more on color science later).

If you zoom in, though, the iPhone Air's shot is clearly superior. Check out the detail on the property nameplate to the left of the tree, for instance, or the paving stones in the foreground. The iPhone 7 smears over details it can't capture, while the iPhone Air's sensor picks up lots more information. These are subtle differences, but the iPhone Air's shot is the better of the two.

This is another example of subtle differences. The two images look similar at first glance, but the iPhone Air captures more brick, metal, and pavement detail than the iPhone 7. The dog is just as cute in both images, mind you.

Pub time! Again, the iPhone 7 does a decent job here, but the detail and color of the main building are more real-looking in the iPhone Air's image. If you zoom in on the main Holly Bush logo or the Hollybush House sign, you'll notice the difference in clarity. The shadows in the latter photo are also more pronounced, which speaks to the iPhone Air's superior dynamic range.

How about a butterfly? The iPhone Air's image is clearly the richer of the two. The leaves, branches, and pattern on the butterfly itself are more detailed in the shot captured with Apple's newer phone, while the iPhone 7's image is softer, almost as if there's a streak of sunblock on the lens. This is the first shot where I think, "Yeah, that photo was shot on an old iPhone."

This is a tricky one. The petal detail on the main flower is slightly better on the iPhone Air shot, and the color of the rear wall is more accurate, too. But the iPhone 7 keeps more foreground detail intact (see the green leaves on the left and the red label at the bottom). We'll call it a draw.

And here we come to the best example of Apple's modern approach to color science in action. On newer iPhones, Apple prioritizes style over realism, occasionally warming things up with more saturation, so shots have an almost yellowish quality. As you can see above, the ice cream is literally a different color in both pictures.

The iPhone 7's approach to this image is too cold — I chose the salted caramel flavor, and the older iPhone saps all the fun and warmth out of that decision. It makes the ice cream look unappetizing and the weather miserable. The iPhone Air's photo looks immeasurably warmer, and while my surroundings weren't quite that yellow, I'd rather exist in this sun-kissed world than in the iPhone 7's gloomy alternative.

Another big win for the iPhone Air here. Neither phone has a dedicated zoom camera, so I had to employ digital zoom in both cases (around 4x), and the iPhone Air captures far, far more detail than the iPhone 7.

So far (the zoom example above notwithstanding), the iPhone 7 has delivered less detailed but still largely usable images versus the iPhone Air. That stops when you switch to low-light scenarios.

In this example, the iPhone 7 just can't handle the bright light of my IKEA donut lamp — you can't even tell that it's a donut at all. The iPhone Air, meanwhile, captures the charming shape of the lamp itself as well as details in the shadows it creates. Look at the pattern on the cupboard door — it's simply not visible in the photo captured by the iPhone 7.

Again, the difference is night and day here (almost literally). The candle in the iPhone 7 photo is extremely blown out — no pun intended — while the iPhone Air more accurately recreates what I was seeing with my eyes (read: light).

If this were a competition to see which phone could best recreate the neo-noir visual style, the iPhone 7 would win (the first photo is very Lynchian). Alas, it's not, and so here we have a great example of just how far the iPhone's night photography skills have come. The iPhone Air photo is more detailed, has better dynamic range, and has more accurate colors — try and read the number plates in the iPhone 7 photo, and tell me I'm wrong.

Oh look, it's me! I'm actually quite impressed with the detail served up by the iPhone 7 here — the hairs on my head and face are equally visible in both examples — but the iPhone Air more accurately captures the whites of my eyes and the details of my complexion (read: the blemishes. Sigh.)

In fact, now I'm looking at the iPhone 7 photo again — specifically, my black eyes — I look a bit like a Great White Shark who's enjoying his last hours as a human. Maybe that's what Apple was going for in 2016?

Surprise! The iPhone Air captured a better photo than the iPhone 7 in almost every example. That was to be expected, and, if you're spending close to four figures on one of the best iPhones in 2026, hoped for.

Apple's latest single-camera iPhone delivers superior details, colors, dynamic range, and digital zoom clarity than its predecessor 10 generations removed, and if that wasn't the case, you'd be worried for the company's future.

But in writing this comparison, I was pleasantly surprised by how well the iPhone 7 held up against the iPhone Air. At first glance, it delivered comparable photos on several occasions, only revealing itself to be a 10-year-old device when I zoomed in and dug into the details (or lack thereof). The low-light examples were a different story, but again, that's to be expected.

So, yes, iPhone photography has come a long way since 2016, but if you find yourself forced to use a banged-up old iPhone 7 for a few days (for whatever reason), it won't be totally incapable of capturing usable photos.

When I first caught a glimpse of The Adventures of Elliot: The Millennium Tales, it sparked a great sense of anticipation within me. After all, Square Enix and Claytechworks were collaborating to bring a brand new HD-2D RPG to the table, which appeared to combine sprinklings of classic Zelda titles with the visual, sonic, and environmental grandeur from series such as Mana and Dragon Quest.

Platform reviewed: PS5
Available on: PS5, Nintendo Switch 2, Xbox Series X and Series S, PC
Release date: June 18, 2026

I’m a huge fan of the HD-2D graphical style, and massively enjoyed recent releases such as Final Fantasy: The Ivalice Chronicles and Dragon Quest 1 & 2 HD-2D Remake, but I still wasn’t quite sure if I’d love The Adventures of Elliot. Was the action combat going to be polished and engaging enough? Was the world going to deliver the spectacle and appeal conjured up with other series? Would the narrative have me hooked?

Well, after playing the game for more than 25 hours now, I have an answer to all of those questions. Here’s what I made of The Adventures of Elliot: The Millennium Tales.

Let’s start by addressing my curiosity surrounding combat in The Adventures of Elliot — just how good is it? Well, I’m pleased to report that I had a lot of fun with the action in this game. It clearly pulls on 2D Zelda, with real-time combat that challenges you to use a variety of weapons to overcome your foes.

You can equip two weapons at once, and while I typically used my sword for close range attacks and bow for projectiles, I found genuine utility in a lot of equipment, be that bombs, a hammer, spear, and more. Combat feels fluid, responsive, and well balanced. You can also defend or parry with a shield, and using this to avoid big damage can be crucial.

There’s a pretty fast, high-octane feel to battles, big or small, and by defeating multiple foes in a row, you can build up a streak to obtain better drops. This adds a layer of fun to combat, and gives you a genuine reason to seek out and destroy random creatures in your vicinity — I had a lot of fun pushing myself to get that streak as high as possible.

Combat is absolutely at its best during boss fights, though. These can offer genuine challenges, and often require you to switch up attacks, defend with care, and employ a variety of weapons to get the win. The satisfaction I got when evading a robotic titan’s assault and slashing it to smithereens with my blade was nothing short of exhilarating.

The highlight of this game is without question its majestic boss battles. Whether I was taking on Minister Kaifried or a bunch of deadly robotic guardians, I enjoyed making use of my full arsenal of weapons in order to emerge victorious.

What’s more, you can customize weapons with something called Magicite, which imparts specific abilities to help you wipe out the opposition with greater ease. This is executed very well, and helps you to raise attack power, unlock elemental attacks, and extend the reach of your attacks, for instance.

By spending more Tul (the in-universe currency), you can use more Magicite, and this helps you scale in terms of power as the game unfolds, making progression feel natural and well-paced.

Other gameplay elements are solid too. Platforming isn’t a massive part of the game, but feels precise and smooth. Your companion for most of the journey, Faie, also has abilities such as dashing and warping, which make traversing environments and taking down enemies even more varied and seamless, and you can unlock more of — and improve on — said abilities as the game progresses.

So, the gameplay in The Adventures of Elliot is a hit, in my book. The brilliant bosses and close contests against frogs, robots, slugs, and more kept me coming back for more. But unfortunately, some things made me feel reluctant to indulge in long, uninterrupted play sessions — namely, the game’s narrative and dialogue.

Simply put, the story in The Adventures of Elliot lacks the spark that I was looking for. It often feels flat, lacking moments of surprise and suspense, and its largely predictable plot points paired with sluggish and dull dialogue meant that I was tempted, at times, to skip through a few scenes — something I never do with story-driven RPGs.

On top of this, the cast of characters is surprisingly weak for a Square Enix game. The protagonist, Elliot, feels somewhat hollow, and spends much of the game telling people to follow their heart, chase their dreams, and to believe in themselves. To be blunt, it feels a bit sappy, and despite his striking appearance, he’s actually quite an uninteresting lead.

A lot of the other characters are written in a slightly wooden way, too. Elliot will meet them, they’ll reveal something that troubles them — be that isolation, missing a loved one, or seeking connection with others — the hero will do something to assist them, and then you move on. As a result, characters often lack nuance or depth, and it feels hard to care about the various individuals involved.

Like a lot of other players have pointed out online, your companion, Faie, is also rather irritating. She speaks up…a lot…and her hand-holdy, pointless interjections can feel grating. You can mute your companion, thankfully, which is a good thing given that the fairy’s high-pitched tone is still haunting me.

Much of the game is centered around time travel, another element that could’ve been handled more effectively in my view. A lot of the environments look identical across different eras, and enemy variety can be pretty limited across time as well.

I did like the discoveries you could make across different ages, though, and hunting for new weapons in the various dungeons, and general exploration, was pretty enjoyable. My critique here, however, is that the puzzles within various areas are very easy, and require little effort to overcome. Therefore, anyone seeking out the ingenious design of classic Zelda dungeons may be left wanting more.

The Adventures of Elliot still nails a lot of the fundamentals, with a beautiful soundtrack, gorgeous HD-2D visuals, and a neat UI. But when I look at the full package, I’m left feeling conflicted.

While the combat is slick and enticing, the underwhelming story and lack of variation in environments and enemies slightly disappointed me.

Although I still had a decent time with The Adventures of Elliot, and I enjoyed its delicious HD-2D graphics and high-octane battles, it’s clear that the new IP has gone through a few growing pains. And unfortunately, its forgettable characters and lacking dialogue bring the overall experience down a touch, meaning it doesn’t quite hit the highest of heights.

You love classic Zelda combat
If you’re a sucker for combat in 2D Zelda games, then this title will surely hit the spot for you. The pace of battle and numerous weapon types keep combat feeling varied and exciting throughout the game’s runtime.

You’re a fan of the HD-2D visual style
If, like me, you’ve enjoyed the HD-2D visual style before, you'll almost certainly love it again here. The game is full of beautiful backdrops and environments, and the expressive 16-bit style sprites really pop.

You’re expecting a gripping story
The biggest weakness of this title is its underwhelming story, with dull dialogue and an uninspired cast of characters holding the overall experience back from greatness.

You want tough puzzles
Although in-game dungeons hold some highly entertaining boss fights, reaching them can often feel like a formality. That’s largely because puzzles are very straightforward, with little challenge involved.

There are a number of ways to customize the experience in The Adventures of Elliot: The Millennium Tales. There are a handful of text languages, and you can swap between English or Japanese voices. You can alter text display speed, and set dialogue to auto if you want to watch scenes unfurl naturally.

There are a range of difficulty modes too, and you can remap controls to your liking for a more custom experience. Unfortunately, there’s no colorblind mode, or similar.

I spent more than 25 hours playing through the main story and side quests in The Adventures of Elliot: The Millennium Tales. I played on Normal difficulty in this instance.

For the most part, I played the game on my PS5, which is connected up to my Sky Glass Gen 2 TV and Marshall Heston 120 soundbar. However, I occasionally dipped into the title on my PS Portal, and used the Sennheiser CX 80U to enjoy in-game audio while on the go.

More generally, I’ve reviewed a wide range of games here at TechRadar, though my main focus has been on RPGs, including Square Enix titles like Final Fantasy: The Ivalice Chronicles and Dragon Quest 1 & 2 HD-2D Remake.

First reviewed: July 2026

Microsoft has been battling to truly establish itself in the device market for years now, with its Surface suite covering everything from foldable smartphones to 2-in-1s up to more traditional laptops and desktops - all the way up to the enormous Surface Hub (RIP).

But the company has seemingly always fallen short - whether it's battery life, falling short on power, or the unavoidable lock-in with the Microsoft 365 experience.

However its latest collection of Surface for Business releases, framed squarely at work and enterprise users, looked to address all of that, and having been using one for the last few weeks now, I can safely say, Microsoft may finally have cracked the formula for a great working laptop at last.

Microsoft has positioned the Surface Laptop for Business squarely at enterprise customers rather than consumers, and it's a substantial refresh over the previous generation, with the biggest improvements around AI, security, manageability, battery life and repairability.

The device is light and portable, weighing in at just over 1.35kg, and its slim build (just 0.69in in width) means it slipped easily into a rucksack or carry-on bag.

It's a stylish device to look at as well - the polished black anodized aluminium build is far more striking than other identikit dull business laptops around today, with a well-designed keyboard and touchpad that offer more than enough space.

But where Microsoft is looking to take a real step forward with the Surface Laptop for Business is in hardware, where the company has equipped the device with Intel Core Ultra (Series 3) processor, up to 64 GB LPDDR5X RAM and up to 1 TB removable Gen4 SSD.

Crucially though, as with many modern devices, it also features AI-specific hardware, with an Intel AI Boost NPU delivering 50 TOPS of AI performance.

Along with offering local AI processing on the device, the NPU looks to perform a number of other tasks, from improved battery life to greater Windows performance. This is a decent level of performance, but if you're looking for tasks such as CAD, 3D rendering or AI model training, you may want something a bit more powerful, with a dedicated Nvidia GPU.

This looks to boost productivity and efficiency across the board - and I can say this was definitely true when I was working on the go.

Sadly I wasn't doing particularly AI-heavy work to really test it out, but I was able to take the device with me on a week-long overseas work trip, using it in conference keynotes, remote interviews, and site visits, and it absolutely ticked all my boxes when it came to responsiveness, usefulness and battery life.

At home however, it wasn't quite the same story.

It should also have meant the device was well-placed to be the centerpiece of my home office set-up, but unfortunately I frequently found issues when trying to connect a range of devices, from monitors to Bluetooth keyboards - this may have been a driver-led issue, but it was frustrating for quite some time.

The lack of ports may be an issue for some users - as there is just one USB-A connection, and two USB-C ports, which might be an issue for some creators. I use a docking station for my set-up, so for my usage I was largely OK - however as mentioned, even this wasn't always responsive - and it's a shame Microsoft has ditched the USB-A port on the charger block cable as well, as this has definitely saved me in the past with previous Surface devices.

Microsoft has also introduced a more advanced haptic touchpad for the Surface Laptop for Business, promising a more consistent click feel and improved gesture support, as well as customizable feedback. Although slightly smaller than my usual work device (a HP EliteBook) I found the touchpad incredibly responsive and interactive, making navigation between different apps and windows a breeze.

My device was equipped with the new integrated privacy display - a feature which gained a lot of attention when it was included in the latest Samsung Galaxy flagship smartphone earlier this year.

Toggled on via the alternate F1 button function, the privacy display instantly makes the screen difficult to read from side angles - no need for an external magnetic privacy filter any more.

This is obviously pretty useful for those workers accessing private or proprietary information, stopping snoopers or spies from catching a glimpse, but it also offers an anti-glare coating which should be a hit with everyone.

Will it be the ideal device for everyone? Probably not, as being a Surface device means it is closely tied-in with the Microsoft ecosystem - so if you use Microsoft 365 at work (which I don't) you'll have a much smoother set-up and overall experience.

The price will also be a sticking point for some shoppers, as my model starts from £1,599 - with the top spec hitting £2,499 - probably out of reach for most start-ups and SMBs.

But overall, I was incredibly impressed by the Surface Laptop for Business, easily the best Microsoft device I've used for work by a mile - and one I hope to get to use again sometime in the future.

The data center industry has found itself in unchartered waters recently. The boom in demand for AI tools and services has led to a 17% increase in electricity consumption. And despite a rapid decline in power consumption per AI task, energy consumption is set to triple for AI-focused facilities.

Recent industry research also reveals the average cost of unplanned downtime has climbed to $9,000 per minute, or $540,000 per hour. For large enterprises, major outages can exceed $5 million per hour when all factors are considered.

Even more sobering: 60% of small and medium businesses that experience catastrophic data loss close within six months.

For data centers, power is non-negotiable, and it’s clear that for those who can take control of their power supply, success in this AI boom can be achieved.

But the challenges are also clear. Europe’s existing, and ageing, grid infrastructure is becoming unreliable under the demand of data centers, renewables, and electric vehicles.

Data centers will not always be the priority, but the need for consistent operations cannot falter. Not if businesses want to stay competitive in a crowd that is growing.

The AI boom is redefining success and failure within the energy sector. And the standout feature of the winners is the access to a reliable power supply. Bloom Energy’s 2026 power report, which looks at developments in the US data center market, indicates a clear trend from areas where the grid is strained to those with ample supply.

Traditional leaders are making way for the new, with data showing that areas like California, Oregon, Iowa, and Nebraska are set to lose up to 50% of their market share due to tighter power availability. Meanwhile Texas’ data center load is set to more than double by 2028. A reliable power supply is not only an advantage, but is king, when considering the construction of new facilities. 

For nearly 20 years, I’ve worked in the sector, on a variety of projects and continents. One I’ve learnt is that the age-old idiom "when America sneezes, the world catches a cold” is true. The lessons of this dramatic shift in the US market should be understood and acted on by leaders in Europe sooner rather than later. 

And the question for European data center operators is obvious – how can you stay take control of your own power availability and stay afloat?

The answer for many has been independence away from grid energy. Rooted in the need to avoid grid capacity challenges and delays due to waiting lists, Bloom’s report indicates  that up to a third of US data centers are expected to be fully off-grid by 2030. And this data on decentralization is growing. The IEA has observed a sharp increase in orders for gas turbines, with the intention to power data centers directly, circumventing a grid connection. 

But this specific solution isn’t without its drawbacks. For example, gas turbines alone struggle to support the large swings in demand that are induced by AI training and modelling.

Increasingly, businesses are finding Battery Energy Storage Systems (BESS) are key to overcoming these sudden waves of demand. BESS can act as a buffer for the load swings AI produces when implemented as a supplementary power source. And the IEA forecasts that up to 25GW of battery storage could be installed in data centers globally by 2030. 

At Aggreko, we’re already deploying this option for a number of our partners across Europe, ranging from 10MW all the way up to loads in excess of 100MW. For example, when a major colocation provider in Dublin found that they would not be able to connect their data center to the grid until several months after the site came online, we were called upon to bridge the gap in the meantime. 

We supplied 12 next-generation gas (NGG) generators  totaling 14MWe, including 10kV of switchgears, transformers, and auxiliary equipment that allowed the site to operate entirely independent of the grid. This was supported by a 1MW BESS, providing resilience against energy swings with the added benefits of improved power quality, emission reduction, and fuel savings. As a result, the site was online right away, compared to the estimated two-year delay for a grid connection. 

The pace of development has been faster than anticipated for most operators, which has led to dramatic shifts in the way they approach power and cooling.

So, secure your energy supply, start decentralizing, and stay on top of new technologies, such as BESS, and you might just find your company winning the AI boom.

We've reviewed, rated, and ranked the best web hosting services.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

Discussions about technological sovereignty in Europe can tend towards doomerism. It is easy to see why. A handful of hyper-scalers in the US and China control the foundational infrastructure of the modern world. The reliance on these technologies from companies and governments in Europe grows with each week that passes.

Near-total dependence on foreign-hosted and trained AI models presents a massive national security risk. If there isn’t a sovereign layer to your infrastructure, you don’t have control over the future. What if the plug is pulled or if security is compromised?

These are the nightmare scenarios being discussed in boardrooms and government departments across the continent.

And yet, Europe has many reasons to feel optimistic in its pursuit of sovereignty. Perhaps the LLM ship has sailed, but it is what comes next that is truly exciting.

In areas such as quantum computing and robotics – and their application across industries including healthcare and climate science – there is evidence that Europe’s combination of engineering talent and deep customer/market knowledge is laying the foundation for the next wave of society-shaping innovation.

This goes to the heart of what AI sovereignty really means: developing a technology that is both a great product, with no compromise on efficiency and scalability, while also solving the biggest possible problems of today or tomorrow.

Now is the time to act. Sovereignty, particularly AI sovereignty, is an absolute priority for European governments and, increasingly, for customers and the public as geopolitical tensions escalate. We are already seeing many examples of companies in Europe building parallel IT infrastructure where they would usually be dependent on hyperscalers. There is also a growing trend towards on-premise cloud and data solutions, in a bid to create sovereign solutions.

So the question is, can we, those of us outside the US and China, build the best products that solve the most pressing problems? The answer is a resounding yes, but we need to ensure that our approach isn’t purely defensive. True sovereignty is about more than having the right level of regulation to protect ourselves against external systems we can’t control. If we want to be sustainable, we need to build alternatives that are at least as good; if we compromise on that, it will fail.

If we take healthcare as an example of sovereignty. It is one of the most sensitive industries when it comes to data, given how personal and confidential medical data is. If your life were dependent on an AI solution that could help define the best personalized or predictive care, and that solution was not sovereign, would you hesitate? Of course not. It is a big problem, but there are companies out there nearing a solution.

A fantastic example of this is the French scale-up ALAN, which is increasingly disrupting the whole medical insurance market thanks to a truly end-to-end designed business process. We need to play to our strengths in these use-case areas where technology has a life-changing impact. In June this year, it announced that it had raised €480m, valuing the company at €5.5bn.

There are, of course, some factors that might hinder our progress, such as overregulation and the constraints that entrepreneurs face. We cannot ignore the challenge of funding; the US benefits from a $30 trillion pension fund that flows massively into the PE and VC markets. We are far away from that in Europe, and we must find solutions urgently.

There are some positive developments, such as the European Union’s flagship €95.5 billion R&D funding program, Horizon Europe, which supports 200-300 groundbreaking scientific discoveries annually across the EU and the UK. But these initiatives are still relatively small and only represent a step in the right direction.

And Europe’s unique digital market is still not as easy to address as the single US or Chinese market. From my experience on the boards of several European scaleups, many of them find it easier to attack the US market once they’ve gained momentum in their home country than to attack another European country. Ultimately, all of these problems need to be solved.

It is also vital that we choose our battles. We shouldn’t be chasing after American or Chinese LLMs or hyperscalers. We will waste time trying to play catch-up. The reality is that we have a proven track record in use case-driven areas, such as healthcare, and that is where we must focus.

Our approach must be to anticipate where the market is moving. Two areas that are emerging as potentially revolutionary are quantum and robotics. The US is already hedging its bets. President Trump recently signed a bill mandating that the US will have a quantum computer by 2028 and requiring that cyber systems adopt post-quantum crypto-consistent solutions. This is what sovereignty is: it’s not hyperscalers; it’s technology that will significantly impact everyone's lives.

The good news is that in Europe, we have many examples of startups producing world-leading technology in quantum and robotics. Ultimately, the success of sovereignty will depend on our ability to build the best products in Europe to solve our greatest problems. We won’t be sustainable if the ‘sovereign’ alternative isn’t as good as the solutions developed elsewhere. In Europe, we have all the tools to achieve sovereignty. But we must act now.

Manage your data with the best business cloud storage.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

New reports on real-world AI deployments seem to be being published almost daily, but there's one clear message which seems to span them all – shadow AI is a major problem.

The use of unapproved or unauthorized tools by workers is a common theme regardless of business size, sector or geography, and it often stems back to one or two reasons – employers are either being too prescriptive about permitted AI tools and are giving workers a narrow window of unsuitable tools to experiment with, or they lack any clear strategy altogether.

These reports have already detailed the risks in great depth, but to summarize, using consumer-grade versions of AI apps puts sensitive and confidential workplace data at risk, be it leaks or secondary exfiltration via model training. Hence why companies invest in enterprise-grade versions with additional safeguards.

Too commonly, employers consider shadow AI a disease that plagues their workers. Something that should be stamped out with more effective training or harsher consequences to breaking the rules.

But the reality is that shadow AI is more often a symptom of the boarder workplace culture, and it's the cause of this that I set out to explore when speaking with industry experts and policymakers.

Canva preaches the importance of freedom of choice – the Australian software giant gives its workers full autonomy over the models they want to use, affording them the time to identify the right tools rather than being prescribed unsuitable alternatives.

Beginning with insufficient and unsuitable policies, Zendesk Chief Legal Officer Shana Simmons explained to me in an exclusive interview that many of today's agreements and policies are far too formal and field-specific.

"AI policies often fail because they’re written for lawyers, not for the people expected to follow them," she outlined, "if people can't understand the guidance, their behavior won't change."

Simmons also explained the policies are being stored behind closed doors in hard-to-reach places, like HR folders that workers never, ever check.

"If a policy is buried in a handbook or on a website, it’s not going to reach employees when they need its," she said, noting that policies should actually form part of the UI – or in other words, where the workers already are.

Canva warns us that, "the most common mistake is treating AI training as a curriculum," whereas it should really be seen as an ongoing back-burner activity that's always developed. The company's spokesperson insisted that workers learn through fixing their own problems, not by "sitting through a course on prompting."

In a bid to work out whether it's employees or employers who are at fault (or whether it's shared), I asked whether shadow AI is a reflection of worker misconduct or insufficient tooling.

In response, Simmons stressed that "most people want to do the right thing," agreeing that the most common cause of shadow AI is indeed poor tooling.

"If employees are given the tools they need and are informed of the rules and requirements in a way that’s understandable to them, and technical controls are in place to restrict the riskiest behavior, I’d expect shadow AI to be no greater a problem than any other form of employee misconduct."

The answer then isn't necessarily to approve every new AI application, but understanding why users prefer certain tools over others is key to building suitable policies and safeguards around those.

In certain, low-risk conditions, shadow AI could actually be an important and useful part of feedback, showing organizations where they're falling short and exactly where to invest, but a clear oversight over this is just as important to ensure that no leaks or other threats occur.

Clearly, then, workers need more guidance and support. But does that come in the form of training, policies, access to tools, or something else?

Simmons explained that "training alone is not very effective for developing AI fluency," though giving workers a clear direction and some initial pointers certainly serves as a helpful baseline. Zendesk, for example, has found the greatest success in giving workers time and space to experiment and become accustomed with AI on their own terms.

This particular company's stance was to pause non-urgent work and organize a dedicated internal hackathon to encourage proactive exploration. The result was a marked increase in employees' practical AI skills and better cross-team collaboration, but halting non-urgent operations altogether isn't a necessity and just reflects one initiative.

It's a similar initiative that's being piloted by Canva, which tells its 5,300+ workers to drop tools for a full week and experiment with AI.

"We give our team the room to step back, get out of business as usual, and try something genuinely new," a spokesperson said.

"Practical AI training should go beyond introductory courses and prompting techniques and create space for employees to actually use the tools in a safe, secure environment," Simmons concluded. Piloting AI tools with synthetic data (and therefore, no harmful consequences) ultimately leads to the highest levels of confidence.

Another key area where studies and reports have been split is in whose responsibility it is to upskill and re-skill, whether that's through updated policies, passive training or active experimentation.

As a C-suite exec, Simmons believes the organization should bear the brunt of the responsibility by giving workers access to tools and learning opportunities. Clearly, they must think outside the box and offer a much broader array of support: "not just training, but also ideation and experimentation through initiatives like hackathons, sandboxes, and collaboration opportunities."

But beyond that, it's totally on the workers' shoulders to "take those opportunities and run with them." After all, it's not just for the benefit of their organization, but it's also to ensure they stay relevant as work evolves in an AI-first era.

A secondary opinion by Canva also backs this up: "Employers own the conditions: the time, budget, permission to experiment... Employees own the curiosity."

The AI race started off with a pretty clear direction – bigger and better. The first waves were characterized by building bigger models, but it’s all change in the world of artificial intelligence and with enterprises, SMBs and consumers all finding use cases for the technology, the focus has shifted.

Now, AI firms and model developers are looking to realize a much tougher goal. Efficiency. Cost per token, performance per watt, output per input, it’s all about driving maximum efficiency.

One clear divide is between training and inference. While training models still requires huge amounts of resources, inference efficiency is starting to improve, and one company (Rebellions) now believes an opening for inference-first hardware could create a new market.

The company’s racks are said to consume around 16-20kW, compared with around 120kW for leading GPU-based inference systems that, for many use cases, are sheer overkill.

Rebellions’ rack costs are also said to be around one-third of the price, making AI inference more accessible and helping enterprises to deploy AI more widely.

Memory is also another battleground, whereby huge trillion-parameter models are testing the limits of today’s hardware and the intertwined reliance on memory and compute. Something Rebellions says it’s looking to fix by working with the likes of SK Hynix and Samsung to align multiple roadmaps, instead of having to respond to shifts in architecture.

Ultimately, today’s black-and-white chip manufacturing landscape is now evolving, and Rebellions sees two key changes happening simultaneously. Firstly, training and inference hardware is starting to differ more drastically. Secondly, aligning multiple hardware roadmaps across memory, compute and more will drive more efficiency not just across deployments, but in terms of bringing new products to market.

I spoke to Rebellions CEO Sunghyun Park allows me to understand how and why inference and training hardware are starting to separate, as well as the importance of open standards and collaboration in the drive for all-round efficiency.

• The AI chip market seems to be splitting between training-first and inference-first architectures. Why is that happening, and why now?

This split exists because training and inference are fundamentally different problems.

Training is how you build a model. It happens once, involves a small number of organizations, and rewards raw computational flexibility because the workload keeps shifting as research moves forward.

Inference is how you actually use a model: every query answered, every transaction processed, every decision an AI system makes in production. That happens billions of times a day across nearly every industry, and it’s where AI moves from R&D into revenue.

Those two jobs need different physics. Training requires maximum FLOPS. Inference requires efficiency, reliability, and economics that hold up when you’re serving users at scale.

The industry forced a training chip into that second job because that’s what existed. Now that inference has become the larger, more urgent market, that compromise no longer holds. Enterprises and governments are asking how fast they can deploy. That’s why the conversation is splitting now.

• Where does Rebellions fit in that split, and what makes your approach to AI inference fundamentally different from your competitors?

We built for inference, from day one. Most first-generation AI chip companies emerged from the 2016-2017 training boom and adapted their architectures for inference afterward.

We started in 2020 – after that wave – with inference as the only target, which meant designing around what production AI actually needs instead of retrofitting a training chip.

The numbers reflect that choice. Our racks draw 16-20kW versus roughly 120kW for leading GPU-based inference systems, about a sixth of the power, in a market where power is the binding constraint for most operators.

Acquisition cost runs around $10 million per rack versus roughly $30 million, about a third of the cost. Our chiplet-based architecture also scales out rather than betting that a single device can handle a model’s full size, which matters now that production workloads are trillion-parameter mixture-of-experts models instead of the few-hundred-million-parameter models the first generation was built around.

It’s also why our architecture is memory-centric rather than compute-centric. The chiplet approach exists to keep memory close to logic as models scale, not just to add cores.

And we have three years of production deployments behind that architecture, not pilots. That’s the hardest part to replicate: real workloads, running at scale, today.

• A year ago, everyone in AI was talking about chiplets. Now the conversation has shifted to memory. What changed?

The chiplet conversation was about architecture: breaking a chip into modular pieces that scale independently, rather than betting everything on a single monolithic die.

That mattered because it let the industry move past an assumption the first generation of accelerators made in 2016 and 2017, that a single device would always be big enough to run any model.

That assumption broke once mixture-of-experts and trillion-parameter models arrived.

The memory conversation is the layer underneath that. Once the architecture problem is solved, the constraint becomes physical: can you actually get enough high-bandwidth memory (HBM) to build what you’ve designed?

HBM is 3D-stacked memory, and how closely you can physically stack it to compute is as much of a bottleneck as raw supply.

Every AI accelerator company is competing for the same limited supply right now, and demand has outpaced what memory makers can produce. That’s the memory-logic co-design problem: architecture and memory supply are no longer separable decisions.

We’re in a different position because our investor relationships were built around supply, not just capital. Our memory partners are also investors, and we co-design our memory architecture directly against their roadmaps rather than simply purchasing off them.

Our chiplet architecture also develops against our foundry partner’s process roadmap. When the rest of the industry was fighting for allocation, we already had a seat at the design table through those relationships.

That’s memory-logic co-design, not just secured supply. The shift from chiplets to memory tracks has moved the real constraint: from architecture to physical supply.

• Your stack runs on open-source frameworks like vLLM, PyTorch, Kubernetes, and OpenShift, tools many enterprises already use. Does that make adoption relatively plug-and-play, or is that an oversimplification?

It’s mostly true, but ‘plug-and-play’ undersells how deliberate that was, and oversimplifies in one specific way.

We built entirely on open standards: vLLM, PyTorch, Kubernetes, and Red Hat OpenShift. We’re one of only two chip companies in the PyTorch Foundation, and the only AI accelerator company fully integrated with OpenShift.

A developer who already knows how to run inference on existing infrastructure already knows how to run it on ours. There’s no proprietary runtime to learn and no migration project. That part really is close to plug-and-play.

The first generation of AI chip companies each built proprietary software stacks, and hundreds of millions of dollars went into software that didn’t survive. We came to market once the open source ecosystem had matured and chose to build on it instead of forking it.

Where it oversimplifies is assuming that means zero integration work. Production deployments still require validating performance at your specific workload and scale, and that takes real engineering time, no matter how compatible the stack is.

What open standards remove is lock-in risk and retraining cost, not the deployment work itself.

• What are the biggest challenges organizations face when trying to run AI inference outside of hyperscaler platforms?

Most organizations aren’t built like hyperscalers, and much of the available inference infrastructure assumes they are.

The first challenge is physical. Most enterprises, telcos, and governments already have data centers. They can’t wait two to three years or spend $600 million-plus on new ones, and they can’t retrofit for liquid cooling without major cost and disruption.

So the practical question is whether inference hardware runs on what they already have: standard racks, air cooling, and existing power budgets.

The second is sovereignty. Organizations increasingly want to bring compute to where their data already lives, rather than move sensitive data to wherever compute is hosted.

That’s partly regulatory, partly operational, but either way, cloud-only inference creates a dependency a lot of operators are no longer comfortable with.

We built specifically for that gap. Our systems run at 4-5kW per server on standard air-cooled infrastructure, no facility redesign required.

SK Telecom has run on our hardware for nearly three years, scaling from a small cluster to close to 100 racks and now processing 50 million API transactions a day, entirely inside their existing network.

KT Cloud runs real-time inference on highway CCTV systems nationally. Both show you don’t need hyperscaler-scale infrastructure to run AI at hyperscaler-relevant volume.

• Cerebras' IPO, and potential listings from others in the space, have put AI infrastructure in the spotlight. What do these IPOs signal about the market, and where might the hype be outpacing deployment reality?

It tells the public markets that AI infrastructure is a durable, investable asset class, not just a venture-backed bet. That validation benefits the whole sector, including us: it says purpose-built AI silicon is real, differentiated from general-purpose GPUs, and worth independent capital.

Capital flowing into the sector is necessary, but it doesn’t by itself determine who wins. The companies that define the next decade of this market won’t necessarily be the most-funded ones.

They’ll be the ones with real fundamentals: production customers, deployment scale, proven economics, durable supply chain relationships. That’s a different filter than fundraising size, and it’s the one that matters once public market scrutiny starts.

We’ve been building toward that filter since 2020, with production deployments and supply relationships with our foundry and memory partners that we secured before the rest of the industry was fighting over the same allocation. The capital is a tailwind for everyone serious about inference.

Whether it gets deployed well is a separate question, and one the market will answer over the next few years.

• Looking ahead, what trends are you watching most closely across AI infrastructure and inference over the next 12–24 months?

The buildout happening inside existing infrastructure keeps accelerating. Most enterprises, telcos, and governments aren’t waiting for new data centers.

They’re deploying inference into facilities they already have, and I expect that to become the bigger story even though it gets less attention than hyperscaler headlines.

Additionally, memory remains the physical constraint. HBM supply hasn’t caught up with demand, and as models keep moving toward trillion-parameter mixture-of-experts architectures, that pressure increases rather than eases.

Companies with secured, strategic supply relationships will have a real advantage over the next two years, not just a cost one.

Chiplets are how we get there. We’ve already mass-produced a highly advanced 4-chiplet package – a level of integration Nvidia has struggled to reach.

Reporting this year indicated Nvidia had built and demonstrated a four-chiplet Rubin Ultra design, then canceled it in favor of a dual-die architecture over manufacturability concerns: a four-die single package pushes roughly 7.5-8x past reticle limits on yield and cost. That’s the foundation.

The next layer we’re building on is performance optimization of HMB3E (3D-stacked memory) in close collaboration with memory and compute co-designed together from the start, not bolted on after the fact.

Also on my radar is the fact that as more companies in this space go public, capital will get valued against production fundamentals rather than funding rounds.

And the efficiency point matters: as inference gets cheaper per token, demand doesn’t shrink, it expands, because new use cases become viable at lower cost. That’s been true of every computing platform in history, and I don’t expect AI inference to be the exception.

For quite a long time now, I've been concerned about how Microsoft is going to monetize Windows going forward. While the world's dominant desktop operating system remains available for a one-off fee (assuming you aren't an upgrader who can get it for free), there's a good chance this could change in the future.

Why do I think that? To me, it feels somewhat inevitable that eventually, Microsoft is going to look for a way to shift Windows from an upfront payment to a subscription model for consumers. That regular monthly income stream piling up in the coffers is the end game for most big tech companies and their products these days for good reasons in terms of the profits to be made.

Of course, rumors about Microsoft looking to charge a subscription fee for using Windows to consumers (as opposed to businesses, where there's already a subscription option) have been floating around for years. True, they've all been dubious and sketchy in nature, or indeed proven outright incorrect, but this idea keeps bubbling up, and I believe we're witnessing a development with AI right now that indicates how Microsoft might have an ideal opportunity to make this pivot at some point down the road.

Last week, Windows Central spotted that the Copilot feature Deep Research is getting the axe, to be replaced with a new feature: Researcher. The idea behind both pieces of functionality is roughly the same (researching and creating detailed reports complete with citations), but the difference is Deep Research was free, whereas Researcher requires a subscription to Microsoft 365 Premium.

This follows Microsoft Whiteboard (where AI assists you in brainstorming ideas) getting changed so personal accounts can no longer use it — the app now requires being signed up for a Microsoft Business account. On top of that, Outlook's Meeting Insight functionality was just shifted behind a paywall, being transformed into an admittedly beefier AI feature, but one that needs a Microsoft 365 Copilot license.

So, there's a trend towards turning previously free AI features into paid ones, presumably as Microsoft rejigs its Copilot offerings and figures out what works well — and what's being used — and whether any of that can be charged for.

Now, we all know AI agents are going to be the 'next big thing' (TM) in Windows 11, certainly if Microsoft has its way, and the idea is for these AI entities to be doing more and more within the OS.

We won't just have an agent for changing Windows settings — and I mean a proper incarnation of this already existing idea, which really can adjust a host of options based on a simple request to "make my laptop battery last longer" or similar — but we will have a small crowd of them. Maybe a troubleshooting agent, for example, which can take a problem that you're battling in Windows and use some genuine AI smarts to help solve it. Or perhaps a creativity agent which can tackle a host of image or video-related tasks, or more broadly help with organizing your projects and photos.

I think the catch will be that eventually, as we've seen with some aspects of AI and Copilot, Microsoft will start shuffling some of these agents behind a paywall — especially considering that more powerful capabilities like these in-depth AI tricks will cost Microsoft a fair bit to drive in terms of cloud resources. It's certainly conceivable that Microsoft could end up charging a small monthly fee to use these premium agents, maybe as separate add-ons in the hope that you'll bolt more of them onto Windows for a cumulatively greater benefit to its coffers.

We could ultimately be looking at a kind of modular, AI-focused OS, and to me, this seems like the easiest way for Microsoft to transform Windows into a subscription model, because it'll happen slowly. You don't need a troubleshooting AI agent to run Windows at all, but it'll be a nice thing to have — certainly for less tech-savvy types — in case problems do arise in the OS.

As more bits and pieces that might start off free drift behind a paywall — as these AI features are built up and become more compelling — people may eventually be tempted to buy a bundle of them. And before you know it, you've signed up for a Windows subscription of sorts (although the free version of the OS will, of course, still be available).

In the end, there may be a couple of bundles — tiered subscriptions, in other words — and let's not forget Microsoft's potential ambitions for a cloud PC model for consumers, either.

This has been a possibility raised in past rumors and leaks (and again, this is something which has already happened in the enterprise world), and if you put all this together, you're looking at a kind of Netflix model, if you will: a streamed OS with several paid subscription tiers. Albeit with a free basic tier for Windows – although who's to say that may not become ad-supported, or indeed more ad-supported, as Windows 11 already does a fair old line in adverts and promos. (Although admittedly Microsoft is cutting back on that front in its crowd-pleasing efforts to fix Windows 11).

This is just my opinion, naturally, and yes, maybe I got rather carried away with the extrapolation at the end there. It's also true that some big question marks remain hanging over the wider notion I've put forward.

As I just mentioned, Microsoft is very much bending over backwards to please Windows 11 users right now, so any possible timeline for this shift may be pushed way back in view of that. Bringing in some form of subscription is hardly going to be a well-received move, even if implemented in small, delicate increments as I'm guessing it would be.

The other obvious sticking point is that Microsoft needs to make its AI agents in Windows worth having. They need to be objects of desire and come packing genuinely useful AI abilities, otherwise clearly, people won't pay for the privilege of having them on their Windows desktop. They also need to be secure and trustworthy so they don't end up throwing spanners in the works of your Windows installation.

That could be the biggest hurdle for Microsoft to overcome, because as it stands, when the idea of AI features being paywalled in Windows 11 has been raised (in rumors and the like), it's been actively welcomed by the more skeptical out there. The cynics are more than happy to have everything AI-related locked away from them, and as non-paying users, this would give them an AI-free Windows 11 desktop.

Again, this plays into any potential move along these lines having to happen further into the future, but I think Microsoft does have this as an eventual goal. If you ask yourself the question: if Microsoft could charge a subscription for Windows 11, would it? The answer is clearly yes, a thousand times over. But the actual, real question here is whether Microsoft thinks it could successfully get away with such a plan without sparking a large-scale defection from its desktop OS.

We can keep our fingers crossed that a monthly charge for Windows isn't coming, but frankly, I think it's likely. Or even just a matter of time — perhaps a lot of time, granted — whether that subscription pertains to AI add-ons, or a cloud PC offering, or Microsoft finds another way to spin this.