Rising cloud AI costs drive demand for local AI PCs

Full text

Now that AI vendors have us hooked and using generative and agentic tools, pricing models are shifting. Subscriptions are no longer loss-making models designed to attract experimental customers, because they now need to serve a much more important purpose – make back money.

How this might look is still largely being figured out, but already we're seeing major software vendors ditch per-seat pricing in favor of token consumption or outcomes.

What this is likely to mean for corporate and enterprise customers is that AI is about to become a whole lot more expensive – but it's not all bad news because these companies also have the opportunity to gain more control over certain spends.

Step forward the AI PC, which is increasingly capable of running small models locally and handling basic and even some mid-level generative tasks without ever needing to spend a cent in the cloud. A few months ago, my social media was even filled with consumers and knowledge workers grabbing Mac minis to run Openclaw's AI agent locally.

A brief overview of how the consumption-based model could look

Shashi Upadhyay, Zendesk's President for Products, Engineering and AI, explained the reason behind the shift we're seeing in pricing models – traditional metrics are becoming less relevant and new metrics are emerging. "We believe software value should align directly with customer success, not headcount," he told me in an interview.

Upadhyay also criticized per-seat models for charging customers for raw AI, whether their problem gets solved or not. But a shift to outcome-based pricing relies on the vendor and customer agreeing "on the exact result that triggers payment."

On the surface, it looks like this could be a variable that might differ on a vendor-by-vendor basis, but thinking about it more deeply, customers would ultimately be able to determine their own meaningful outcome, meaning that they only ever pay for success and never pay for failures.

Distinguishing the needs of local processing vs. cloud compute

With this emerging hybrid split of use cases in mind, it would mean enterprises can pay a one-time set price for an AI PC and workers will have unlimited access to on-device processing for tasks like summarization, transcription, image background removal and more.

It leaves cloud – the more unpredictable expense – only for the tasks that require extra compute, such as huge enterprise applications that handle centralized data, or running the latest and most powerful models.

I set out to determine what exactly makes an AI PC and how analysts expect this shift to impact the relevant markets, but Omdia’s research director for PC and tablet research Ishan Dutt warned me that we're still living through this transition, so quite how the end looks is yet to be determined.

For example, Dutt explained that so-called 'AI-ready' PCs just 18 months ago would've had sub-10 TOPS NPUs. But then came along Microsoft's own classification of Copilot+ PCs with around 40 TOPS. Just recently at CES 2025, Intel, AMD and Qualcomm all showed 50+ TOPS NPUs, and we're already seeing hints of 75+ TOPS.

Market intelligence firms like Omdia would typically classify an AI PC as one possessing a Neural Processing Unit (NPU), "designed for running AI workloads locally alongside the CPU/GPU," Dutt told me, but this is clearly a new category whose ceiling is still being pushed by chipmakers, and whose baseline is still being written.

Other, more traditional metrics are also relevant in the world of local processing, with Omdia implying that 16GB or memory is barely sufficient these days. Anything more than lightweight tasks is more likely to benefit from 32GB+.

"In practice, 'AI-capability' is really a function of three hardware considerations (NPU TOPS, RAM, and increasingly GPU for generative/creative workloads), not a single number," Dutt told me in an exclusive interview. "I’d expect the bar to keep moving as agentic, always-on background AI workloads become the reference use case rather than chat-style assistants."

AI PC demand vs. the effects of the device refresh cycle

Having determined the parameters that broadly define an AI PC, I wanted to understand whether enterprises are actually kitting staff out with them today. I've read many a study about the increasing unpredictability of cloud-based AI pricing, and data even shows that AI PC shipments are rising. But is this indicative of a shift away from cloud-exclusive processing, or is it just that more PCs now classify as AI PCs anyway, and generic refresh cycles are impacting these numbers with false positives?

"NPUs are now the baseline across new Intel, AMD and Qualcomm platforms," Dutt said, before reminding that me that Macs have had NPUs since the 2020 move to Apple silicon, which came around two years before ChatGPT went public.

Omdia's market analysis projects that the AI PC (equipped with an NPU) market share among all PCs will reach 77.5% by 2030, compared with 17.3% in 2024. Analysts also note that the Windows 10 end of life concentrated a large number of upgrades over the past year or two, bringing AI PC market share up considerably.

Dutt explained that "clearer outlining of how hardware upgrades unlock future on-device AI functionality" could further impact the cadence of refresh cycles.

AI PCs have a pricing problem of their own

"Consumption-based cloud AI pricing does create a genuine total cost of ownership argument for shifting inference workloads on-device," Dutt agreed, however the PC market has its own challenges that are also leasing to pricing instability.

While it may seem like we've been living through a chip shortage for half a decade, we've actually been in two – and this second one is probably much worse.

The earlier 2020-2022 shortage was a "pandemic-driven demand/logistics mismatch," which normalized pretty quickly. This second shortage is a "deliberate capacity reallocation as memory makers are shifting DRAM/NAND wafer capacity toward HBM and high-capacity DDR5," Dutt told me. In other words, the problem we're looking to solve (AI-driving pricing instability) is being driven by itself (enterprise AI consumption). A Catch-22, if you like.

With Apple recently doing the unthinkable and raising prices across most Mac models, or cutting entry-level models altogether, Omdia sets out the challenges that lay ahead: 60% of channel partners expect to experience either delays (53%) or outright cancellations (7%) of their customers’ hardware refresh plans.

Already, 70% of channel partners are expecting major shortages across some (11%) or most (59%) of their PC hardware products.

The future is hybrid

Returning to my original theory that many organizational AI workloads will shift to local processing amid an evolving cloud subscription landscape, it's increasingly clear that a rise in AI PC shipments doesn't necessarily correlate with genuine customer demand.

While shifting some processing to run locally has huge merits when looking at the cloud compute market in isolation, the PC market's own challenges complicate strategies further. The answer will likely involve a hybrid approach to both sides, but how that hybrid and split will look will vary case-by-case.

But more crucially, cloud processing isn't set to go anywhere, and it will continue to complement local workloads with the type of power and performance that few organizations could ever afford to acquire. Upadhyay reminded me that cloud compute's benefits span "centralizing multi-channel customer data, real-time ticket routing, unified reporting, and seamless API integrations across hundreds of tools."