The Rise of AI PCs: Challenges and Opportunities in 2025

Opinion: Why 2025 Will Be An Important Year For AI PCs

The continuing development of third-party applications and features, combined with the slow but ongoing rollouts of features from Microsoft and Apple, point to the possibility that the nascent but growing AI PC category will face an important test in 2025. It's almost the end of the year, and vendors have yet to make a definitive case for the existence of the AI PC. That shouldn't be seen as a controversial statement because we're much more likely to see vendors pull this off in 2025. After all, the two flagship feature sets for AI PCs -- Copilot+ from Microsoft and Apple Intelligence from, well, Apple -- are not yet widely available to all users. And we have yet to see a groundswell of applications from ISVs that have found widespread traction. [Related: A Larger Wave Of AI PCs Is Coming. Are Businesses Ready?] While Apple Intelligence is set to become available in all Macs powered by Apple's M-series processors, Copilot+ features are only expected to be available in a growing subset of PCs with a neural processing unit (NPU) in the system-on-chip. (Copilot+ should not be confused with Microsoft's Copilot offerings, which are largely cloud-based right now.) Announced back in June, Apple Intelligence became available in beta form in late October for Mac computers with M-series processors, which date back to 2020. This means that users need to have a recent Mac and must sign up for a wait list to gain access to the features, which include things like Writing Tools and an enhanced Siri voice assistant, but they won't have access to other features until a future update. It's not yet known when Apple Intelligence will become generally available, and it's not clear if Apple plans to turn the features on by default at that point. The Copilot+ features from Microsoft, on the other hand, have had a somewhat slow and bumpy rollout, which hasn't been completely intentional. When Microsoft announced the first batch of Copilot+ PCs in May, it did so with Qualcomm as the exclusive chip launch partner. As a result, the initial set of devices from Microsoft and OEMs only used Qualcomm's Snapdragon X processors. While Qualcomm has touted performance and efficiency advantages for the Snapdragon X processors over competing chips from Intel and AMD, the company's silicon is based on the Arm instruction set architecture, which, unlike the x86 architecture underpinning Intel and AMD processors, doesn't have full compatibility with Windows applications yet. It's important to note that Qualcomm has made significant progress in enabling native versions of Windows software, ranging from all the Microsoft Office applications and those from other big names like Adobe, Zoom and Cisco, to apps from smaller players. But there are still gaps in software support, with an emulation layer supporting some but not all x86-based applications -- and the emulated performance isn't perfect. This has, for now, limited the appeal of Arm-based Copilot+ PCs. Then there's the issue of Recall, which was supposed to be the flagship feature of Copilot+ PCs at launch in June until Microsoft delayed it to beef up the feature's security. With Recall, Microsoft promises to provide users with a better way of finding things they accessed or viewed on their computer. The feature accomplishes this by letting users perform a natural language search against screenshots that are taken every few seconds. Shortly after Microsoft revealed Recall in May, privacy and security experts immediately raised concerns about how creating a central store of screenshots could give a bad actor easy access to all sorts of sensitive information. The concerns prompted Microsoft to delay the release of Recall so that it could develop security safeguards and allow time for user testing. As announced by Microsoft, these security enhancements include requirements for Windows features such as virtualization-based security and Windows Hello for biometric authentication. With this extra focus on security, Microsoft only made Recall available to members of the Windows Insider Community for testing on Arm-based Copilot+ PCs in late October. This brings us to another reason the appeal of Copilot+ PCs has remained limited: the availability of Microsoft's advanced AI features on x86-based systems. When Microsoft announced the Arm-based Copilot+ PCs in May, it said any computer carrying its AI PC brand would need to have a minimum NPU performance of 40 trillion operations per second (TOPS) to power Copilot+ features. At the time, there were no x86-based PCs available that could reach that level of NPU performance, but that started to change in the past few months with AMD releasing its Ryzen AI 300 series and Intel releasing its Core Ultra 200V series. However, a problem remained: Laptops powered by these AMD or Intel processors still didn't have the ability to run Copilot+ features because Microsoft had not yet made them available for x86-based PCs, even if they met the NPU performance requirement. That changed to a small degree on Friday when Microsoft announced that Copilot+ features like Recall, Click to Do and Paint Cocreator would become available to Windows Insider Community members for testing on these recently released x86-based PCs. However, Microsoft has yet to say when Copilot+ features will become generally available on compatible systems. And we don't know when that will happen with Recall on Arm-based PCs. While tech vendors are working with a wide range of ISVs to bring to life new applications and features that will take advantage of AI PCs, it can be hard to find more than a handful of software offerings that are currently available. I learned this firsthand by searching online for software that has been designed to take advantage of the NPU, which has become a requirement for AI PCs in the eyes of major vendors such as Intel, AMD, HP, Dell Technologies and Lenovo. Apple has its own version called the Neural Engine, which has been in its M-series chips since they debuted in 2020. What I found is that even among the third-party applications and features that vendors like Microsoft have highlighted, not all have become available yet. By checking lists of ISVs working with Microsoft, Qualcomm and Intel and then cross-checking the websites of those ISVs, I identified at least eight AI PC applications that are available on some or all AI PCs for a separate CRN article. There are likely more but finding those software offerings is not yet straightforward. The continuing development of third-party applications and features, combined with the slow but ongoing rollouts of features from Microsoft and Apple, point to the possibility that we will see some or many of these offerings become widely available in 2025. If this happens, it will become a major test for AI PCs. A recent survey commissioned by Intel of 6,000 adults in the United Kingdom, France and Germany found that 86 percent "have never heard of or used an AI PC." Assuming there is a similar level of unfamiliarity with the category in other regions, this means vendors have a lot of work to do in highlighting what AI PCs are and what software consumers and businesses can use to take advantage of such devices. After all, consumers and businesses will have fewer options for buying PCs that aren't marketed as AI PCs next year since such devices are expected to represent 43 percent of global PC shipments by then, according to research firm Gartner. And they will only become more pervasive over time, with research firm IDC forecasting that AI PCs will constitute 60 percent of global PC shipments by 2027. As more and more AI PCs land in the laps and desks of users, vendors will get to watch what happens. Will people find useful applications and features and drive adoption to the high levels of popular cloud-based AI applications like ChatGPT? Vendors certainly hope so, which would, in turn, be a boon for their channel partners. If not, they will need to come up with new tricks to revitalize the PC category and keep the opportunities flowing for solution providers.

A Larger Wave Of AI PCs Is Coming. Are Businesses Ready?

As a growing number of AI PC options for businesses hit the market, industry analysts and solution provider executives believe there are multiple impediments to businesses adopting such devices or taking advantage of their AI capabilities, including security concerns. When Microsoft revealed the first batch of Copilot+ PCs with advanced AI capabilities in May, the appeal to businesses was limited: the laptops didn't use tried and trusted x86-based processors, and they mainly consisted of premium designs for consumers. But this will change over the next year as a larger wave of AI PCs become available for business users from a variety of OEMs ranging from Lenovo and HP Inc. to Dell Technologies and others building Windows-based computers. [Related: Microsoft's Windows 10 End Might Spark Start Of New AI PC Spending] "Our general strategy is we innovate in the premium segment, and we make it mainstream. And you'll see a lot of that mainstreaming happening [next] year," said Alex Cho, president of HP's Personal Systems division, in a September interview with CRN. Since AI PCs started to appear in 2023, one of their defining attributes has been the inclusion of a neural processing unit (NPU) to handle sustained AI workloads alongside the CPU and GPU for other kinds of local processing in the computer's system-on-chip. But now the Windows PC ecosystem is coalescing around Microsoft's Copilot+ PCs, which require a minimum NPU performance of 40 trillion operations per second (TOPS) to deliver advanced AI features such as live captions, image generation or the controversial Recall search feature that was made available for testing in mid-November. While the first Copilot+ PCs released earlier this year use Qualcomm's Arm-based Snapdragon X system-on-chips, OEMs are now starting to release PCs with recently released x86-based AMD Ryzen and Intel Core processors that were expected to support Copilot+ features with a free Windows update in November. At the same time, Apple in October started to roll out its first advanced AI features under the umbrella name of Apple Intelligence, which is available not only to Mac computers with its M-series system-on-chips but also iPads and iPhones released from the past few years. This means businesses will have a growing number of options for AI PCs, including those with lower price points for mainstream users. A February forecast from research firm Gartner said AI PC shipments were expected to more than double in 2025 to 116 million units, which would account for 43 percent of estimated total PC shipments that year. But while the share of AI PCs available in the market is expected to grow significantly next year, industry analysts and solution provider executives believe there are multiple impediments to businesses adopting such devices or taking advantage of their AI capabilities. Bob Venero, CEO of Fort Lauderdale, Fla.-based solution provider Future Tech Enterprises, told CRN that the topic of AI PCs has yet to come up in conversations with customers, which include Fortune 500-type companies, about refreshing their fleets of PCs. "It's not like, 'Hey, we want to start rolling out AI PCs in the environment,'" said Venero, whose company is No. 76 in CRN's 2024 Solution Provider 500 list. But Venero (pictured above) also sees his company as sort of an anomaly because of its focus on large enterprises. He suspects AI PCs will get picked up quicker by smaller businesses, particularly startups who are nimbler and want to be on the cutting edge. "Because they can build around it, No. 1. No. 2, [it's] less risk for them. And No. 3, they need they can leverage the AI capabilities to do more with less, where they're not as mature in their company's journey," he said. For wider adoption of AI PCs to occur, Venero thinks solution providers need to work closely with customers on how they can use advanced AI applications without creating any security or privacy risks, which he believes are the top concerns among businesses. "It all comes around to the security, the guardrails and the risk vector associated with can they be deployed within an organization with the right controls to avoid having any" intellectual property, healthcare information or otherwise sensitive information leaking, according to the solution provider executive. Jack Gold, principal analyst at Northborough, Mass.-based J. Gold Associates, told CRN that while he expects 85-90 percent of new PC purchases by enterprises to be AI PCs within three years, he also sees security concerns as the main reason for businesses not wanting to buy such devices or use their AI capabilities for now. "A lot of companies don't know if they want to give AI to their end users," he said. While PC and chip vendors have said that AI PCs can provide a more secure AI experience than cloud-based services by processing data locally, Gold said the advent of AI-powered desktop applications could still open up considerable security concerns. The main issue, according to Gold, is the same as one of the AI PC's central appeals: the promise that AI applications could help users more easily access, summarize or analyze proprietary company data. While that may improve productivity and convenience, it could also make such data more easily available to hackers who gain access to a user's PC through social engineering tricks, including phishing. "If you're able to generate more data locally, then you're also increasing your exposure," he said. This is a problem Microsoft has had to address with its Recall feature, which is meant to help users more easily find things they have done on their Copilot+ PCs by letting them perform a natural language search against screenshots that are taken every few seconds. When Microsoft revealed Recall as the flagship feature of Copilot+ PCs, privacy and security experts immediately raised concerns about how creating a central store of screenshots could give a bad actor easy access to all sorts of sensitive information. The concerns prompted Microsoft to delay the release of Recall so that it could develop security safeguards and allow time for user testing. As announced by Microsoft, these security enhancements include requirements for Windows features such as virtualization-based security and Windows Hello for biometric authentication. While Gold thinks security concerns may hold up businesses from fully embracing AI PCs for now, he believes most of them will eventually come around as vendors like Microsoft iron out issues with applications or features like Recall. "Companies forever have had employees use tools that the employees think make their life easier, but the end result could lead to security issues," he said. Outside of security issues, another pressing topic for AI PCs is software availability. While Microsoft and some ISVs have started to push out applications that can take advantage of the AI PC's local processing capabilities, there aren't too many of them right now. "Right now, a lot of IT managers, a lot of the tech vendors themselves are a bit hard-pressed to elucidate three killer use cases, especially ones that are sort of enterprise-facing," said Linn Huang, research vice president for displays and devices at research firm IDC. However, Huang said Microsoft and other vendors, like Intel, have been pushing to enable and generate excitement for a wide variety of third-party AI PC uses cases. Intel, for instance, is working with more than 100 ISVs to enable over 300 AI-powered features on AI PCs through its AI PC Acceleration Program. "There's a lot of AI features that are going to be coming out in the pipeline next year," he said. But until that groundswell of killer apps happens, Huang thinks most AI PC purchases by businesses now should be seen as future proofing. "Buying an AI PC, it might not do the things you hoped it would do, but in 12 months' time, it's going to do a lot of things that you otherwise wouldn't have imagined," he said. Another issue hindering interest in AI PCs is general lack of awareness, which Intel recently highlighted in a survey of 6,000 adults in the United Kingdom, France and Germany. In the Intel-commissioned survey released last month, it was found that 86 percent of consumers "have never heard of or used an AI PC -- yet 40 percent would consider one for their next upgrade once they understand the benefits." Harry Zarek, president of Ontario, Canada-based Compugen, No. 60 on CRN's 2024 SP 500 list, said his company is focusing on educating its customers and "understanding how they see AI as a strategy in the business." "We're encouraging [customers] to dabble, to put their toes in the water, so to speak, and begin to understand what it can do for you and the whole bunch of very simple use cases and workloads and activities that can be enhanced by the use of AI, whether it's Copilot or OpenAI or whatever you want," he said.

The Edge AI Renaissance

The promise of edge AI lies in its ability to have an immediate impact on real-life problems for businesses, leading to a wide-open field for innovative solution providers. When Wi-Fi wasn't cutting it, MSP Step CG helped Toyota with a private 5G solution to improve operations on the manufacturer's factory floor and power its edge AI applications. Now there are 20 more industrial and health-care deals with big-name brands in the pipeline in addition to the deal with Toyota for fast-growing Step, according to Ed Walton, CEO of the Covington, Ky.-based company. "It's really picking up momentum. Anywhere we're deploying private cellular, it is enabling edge applications with AI," Walton said. [RELATED: The 2024 Edge Computing 100] Step has built out a robust private networking practice and has been deploying private cellular networks for its enterprise customers with the help of its technology partners Ericsson and Celona. The MSP is integrating that technology with traditional wired and wireless infrastructure from companies including Hewlett Packard Enterprise and Juniper Networks. Step has seen a massive uptick in private cellular use cases, especially for industrial environments. Private cellular and 5G, combined with edge computing, are the power behind AI at the edge where fast speeds and low latency are required, Walton said. Edge AI, or using AI in an edge computing environment to bring the processing of AI functions closer to where the data is being generated, is top of mind for a variety of technology vendors, cloud providers and hardware makers alike. It's also an emerging area of opportunity for solution providers. Worldwide spending on edge computing is expected to reach $228 billion in 2024, an increase of 15 percent over 2023, according to IDC. The research firm has further forecast global spending on edge computing to skyrocket over the next several years to $378 billion in 2028. Where edge AI will make the largest impact isn't in the carpeted enterprise. It's inside the factories, mining operations, airport baggage handling areas and hospital corridors where quick decision-making is required, according to Shahid Ahmed, group executive vice president of new ventures and innovation for solution provider giant NTT Data. "That's the world we are living in with edge AI. It's very different. It has zero time for cloud interaction and has to be done all locally, not only for security reasons, but for latency reasons. That's what our edge AI does, [and] we think it's a phenomenal opportunity for this part of the world where we haven't seen a lot of AI transformation [as opposed to] the carpeted space," Ahmed said. To that end, NTT Data in July unveiled its ultralight Edge AI platform, a fully managed service that lets businesses deploy AI applications at the edge using smaller, more efficient language learning models. With U.S. headquarters in Plano, Texas, NTT Data today is working with a number of businesses to improve their operations by employing AI at the edge. Its pipeline for this business is rapidly growing every week, according to Ahmed. One such manufacturing customer is employing computer vision in the form of a 4K camera that can capture events over the factory's thousands of square meters in real time. It's analyzing automated guided vehicles, goods and whether people are wearing safety gear moving through different areas of the floor. "It's basically ingesting terabytes of data per hour, and it's simultaneously taking action and correlating all those actions into a single AI model. It knows if somebody wore a hard hat or didn't wear a hat, if a door was left open, how many boxes were put onto the pallet and ready to be shipped, all in one single capture, and it's able to take action simultaneously," he said. "It's one single camera. That's the power of edge AI." The same process without computer vision would require separate devices capturing the different areas of the facility, including views of the pallets, people and doors, Ahmed said. "Not only do you reduce all those sensors and IoT devices and take them out of the picture because you've got one single machine vision camera, but it can take immediate action to alert the right people," he said. The promise of edge AI lies in its ability to have an immediate impact on real-life problems, Ahmed said. "I think a lot of focus is on the big AI models at the moment, but none of them really has a way of scaling down to where you can get dynamic, time series data that requires very quick-response times to take actionable measures or provide tools that can immediately provide some benefit to that factory worker that makes their life a lot easier," he said. "I call this 'blue-collar AI.' It's gritty, it's raw, and it's real." Edge computing adoption has largely been niche, particularly seen in industries such as retail and manufacturing for real-time applications. But AI at the edge is evolving, and solution providers like World Wide Technology (WWT) are increasingly seeing demand for emerging use cases that can't tolerate even seconds of latency, including computer vision and safety in manufacturing. "If I'm a user executing a chatbot prompt, and my interface to that is text in a browser, I don't really care if that takes one second or 10 seconds. I don't really need edge AI. It can go back to wherever -- the cloud or the data center -- wherever I've got my LLM inferencing," said Neil Anderson, vice president of cloud, infrastructure and AI solutions for St. Louis-based WWT. "But if I'm trying to do something like computer vision where I've got some cameras locally in a store where I'm trying to do facial recognition or customer behavior recognition, that's where you really need something at the edge where you can process that locally. You're not going to want to drag all that video back to a central place to process it. You need a distributed approach." WWT has helped customers implement safety measures in industrial manufacturing settings. "[The customer might say,] 'I want to use computer vision for security and to keep people safe.' If they're sticking an arm somewhere they shouldn't be, we can detect that and stop the machine. That requires ultra-fast response time, so you're going to do that at the edge," Anderson said. It's critical that solution providers understand the specific use case their customers have in mind. From there, they can determine the architecture requirements, Anderson said. WWT, an AI-first company according to its co-founder and CEO Jim Kavanaugh, offers AI workshops for customers. The first thing that's discussed is the use case, Anderson said. "'What are you trying to do?' The second thing is, 'OK, what does the architecture look like to be successful to bring that use case to life?'" he said. Platform manufacturers are especially keen on edge AI to help offload features that don't have to be processed in the cloud or a data center. Cisco Systems' Webex platform for collaboration, as an example, has an AI-powered background noise removal feature that automatically filters unwanted or unexpected noise for users. The San Jose, Calif.-based tech giant also partnered with Nvidia in 2023 for the creation of the Room Kit EQX, a packaged offering for medium and large conference rooms that consolidates audio, video and compute components into a single unit that's powered by Nvidia's Jetson edge AI platform and is bringing advanced AI capabilities to Cisco's portfolio of collaboration devices. Security concerns abound for edge AI, simply because it's yet another attack vector that can be exploited, solution providers said. "To the extent that you're putting data on devices or you're putting data at the edge, you have to be protecting that data wherever it is, otherwise you're at risk," said Steve Wylie, senior vice president and general manager, East, for Irvine, Calif.-based solution provider giant Trace3. As excitement and momentum builds around potential game-changing use cases, there are steps that enterprises need to take now to prepare for when they do have a viable edge AI opportunity, he said. "They need to know, 'Where's your data? Is it in a centralized place, or at least, do you know what's where, and is it clean? Can you access it?' There are things that people can be doing with their environments to prepare themselves so that when they have a viable use case, they can act on that use case," Wylie said. Trace3 has placed a big bet on AI, as evidenced by upward of 80 percent of the team having been trained on it. The businesses that are thinking about AI the right way are doing some level of planning now with the help of Trace3, Wylie said. "Our first step was getting the team to speak the language of AI. The next piece is then helping clients speak the language of AI. The whole reason for doing that is so people can then understand the real capabilities of AI, what they can really do with it, and preparing so they can put their business in a place to be ready to adopt those use cases when their business is ready to consume them," he said. Wylie and his team are doing red-teaming exercises to show clients where their cybersecurity gaps are related to AI technology, especially in a distributed environment or network edge, he said. "We're showing clients [why] you can't just implement [Microsoft] Copilot out of the box and just hope that it's secure. If you're not fine-tuning it to your environment and making sure holes are closed, you're putting yourself at risk," he said. "With any of these capabilities, you're opening up access to a data lake, or access into some element of your environment that has a ton of your data. You can really get the keys to the kingdom." Ericsson Enterprise Wireless Solutions, which now includes Cradlepoint, is focused on building AI into the network. AI requires data, and to get that data to the edge, good connectivity is king, said Donna Johnson, head of marketing for Plano, Texas-based Ericsson Enterprise Wireless Solutions. Similar to solution providers, manufacturing is a popular use case that Cradlepoint is seeing where there is a desire for edge computing using AI. The problem is, connectivity usually isn't sufficient to send the data back to the data center or a cloud for processing at many manufacturing facilities, Johnson said. "The connectivity often isn't high-performance enough, and it's subject to intermittent outages, even micro-outages in the Wi-Fi network. Our work on private 5G is a great example of how we're building a really high-performance, high-capacity network that allows us to collect all these data points from the manufacturing process, at the edge, and do analysis on it that allows [businesses] to actually correct potential errors and find safety issues in real time, and hold on to the data for later analysis for potential issues," she said. It's a huge ask from customers, and a big opportunity for solution providers, Johnson said. "No one company has the full solution here, so I think that integration role is going to be huge for the channel to be able to bring the different pieces together in a unique way for individual customers," she said. As executives read about edge AI and learn how it could work for their businesses, it's been "call after call" for Albertson, N.Y.-based solution provider Vandis, according to CTO Ryan Young. Many customers, especially those in the manufacturing space, have been running AI in centralized data lakes like Microsoft Azure, Amazon Web Services or Google Cloud, but the costs associated with running AI in those environments are becoming too high. These businesses are looking to defray those costs by putting AI at the edge. By moving to the edge, there's also the added benefit of modernizing what were once very "low-tech" networks, Young said. "Historically, manufacturing has had very low-throughput networks that are really just there to support the barcode scanners and forklifts. Now, all of a sudden, it's, 'We're going to put compute out there, and we need at least a 10-Gig backbone and we need 10-Gig ports to connect the compute.' We're seeing those early customers building the foundation," he said. Vandis is evaluating its customers' data to determine what they need and how they can piece together a solution, he said. "[HPE] Aruba [Networking] has a product. Juniper has a product with Mist, Palo Alto [Networks] and Fortinet have products, but these are point products. The overall data is not being brought together and completed," he said. "The journey we're on right now with our customers is completing their data." To do that, Vandis is using ReadyWorks, a low-code, agentless SaaS platform to pull data across disconnected systems into one source. "When you're looking at manufacturing or shipping or supply line constraints, you have all the data there for the AI to be able to make decisions on," he said. "We're starting to work with these [customers] on that completeness of data so that when they're ready to make that next step, they're in a really solid position to actually run AI models against that data set." That's where vendors like HPE, with its edge AI clusters, will come into play, he added. "They're preloaded with trained GenAI models that you just pump your data into and start your query," Young said. "I think that's going to be kind of an easy button for a lot of our customers who don't have the skill set or the wherewithal to go and build their own AI models."