Hardware Watch: The AI Hardware Shift & The Next AI Device Category
The AI Hardware Shift: From AI PCs to Machines That Can Actually Run Agents
The AI PC Has to Become More Than an NPU
The first wave of AI hardware was defined by branding: AI PCs, AI smartphones, AI everything. The next wave will be defined by capability. The question is no longer whether a device has an NPU or an AI label. It is where AI inference should actually happen—on the device, at the edge, or in the cloud—and what hardware characteristics are required to run autonomous agents rather than simple assistants. The answer will reshape the PC, smartphone, and device markets over the next three years.
Goldman Sachs projects 150 million AI PC shipments in 2026, up 39% year-on-year, representing 59% of all PCs. Omdia estimates that roughly 50% of PCs shipped this year will include NPUs. But the presence of an NPU does not guarantee meaningful AI capability. Memory bandwidth, software ecosystem maturity, and real-world workload support are the actual determinants of whether an AI PC delivers value.
Local AI vs. Cloud AI
The local-versus-cloud debate is often framed as a binary choice. It is not. Local AI offers lower latency, privacy, offline operation, and reduced network dependence. Cloud AI offers greater compute, larger models, continuous updates, and more powerful inference. The emerging hybrid model—exemplified by Lenovo's Think-family devices—spans on-device, edge, on-premises, and cloud computing.
| On-Device | Cloud |
|---|---|
| Lower latency | Greater compute |
| Privacy | Larger models |
| Offline operation | Continuous updates |
| Lower network dependence | More powerful inference |
The hardware that wins will be the hardware that enables seamless task routing between local and cloud resources. A voice assistant should run locally for latency and privacy. A complex document analysis may require cloud compute. An agent orchestrating a multi-step workflow may need both.
The Agent Changes the Hardware Requirement
Agents are not chatbots. They are autonomous systems that perceive objectives, plan multi-step actions, and execute across applications. This changes hardware requirements in fundamental ways. Agents require sustained inference rather than bursty inference—they run continuously, monitoring context and making decisions. They require memory to hold state across sessions. They require battery efficiency because they run in the background.
They require local context—access to files, calendars, communications—to be effective. They require connectivity to cloud services for tasks that exceed local capability. They require sensors to perceive their environment. And they require security to protect the data they access.
Why this matters: The NPU arms race—from 40 TOPS to 180 TOPS—addresses only one of these requirements. The hardware that runs agents effectively will need to be designed around agentic workloads, not just AI inference benchmarks.
GPUs Remain Central
Despite the proliferation of NPUs and specialized AI silicon, GPUs remain the workhorse of AI compute. Q2 2026 desktop discrete GPU shipments reached 12.5 million units—a four-year high—with Nvidia capturing approximately 90% share. AMD holds around 8%, and Intel has climbed to 2%.
Jon Peddie Research notes that buyers are rushing to purchase GPUs before expected price increases, creating a paradoxical market where high prices drive higher sales. The GPU remains central because it is the most versatile AI accelerator: capable of training and inference, adaptable to new model architectures, and supported by the most mature software ecosystem. NPUs are efficient for specific tasks. GPUs are capable of everything.
Beyond PCs: The Next AI Device Category
The AI hardware market is expanding beyond PCs and smartphones. AI wearables—particularly smart glasses—are experiencing explosive growth. Global smart headset shipments grew 83% year-on-year in Q1 2026, with AR glasses and display-free smart glasses growing 136% and 210% respectively. Full-year 2026 smart glasses shipments are projected at 16.1 million units.
AI smartphones are also scaling: IDC forecasts 147 million AI smartphone shipments in China alone in 2026, up 31.6%, with global GenAI smartphone shipments reaching 432 million. Qualcomm has introduced Dragonwing Q-2390 and IQ-2390 processors for consumer IoT, industrial edge, and enterprise terminals. Ambarella and ZEDEDA are partnering to bring cloud-orchestrated AI to billions of edge devices.
The next AI device category is not a single product. It is a proliferation of form factors—glasses, phones, wearables, industrial terminals, robots—each optimized for different agentic workloads.
Hardware Economics
The economics of AI hardware are becoming more challenging. Memory has become the single largest cost driver: Microsoft's AI PC definition requires 16GB of memory and 40 TOPS of compute, which has pushed memory's share of bill-of-materials costs from 15% to 35%. DRAM and NAND prices are projected to rise more than 130% by the end of 2026 as memory manufacturers prioritize high-margin HBM and enterprise SSDs over consumer DRAM.
| Cost Factor | Current Trend | Impact |
|---|---|---|
| Memory | DRAM and NAND prices projected to rise 130%+ by the end of 2026 | Memory's share of BOM costs has risen from 15% to 35% |
| AI silicon | NPU TOPS increasing; cost per TOPS declining | Silicon is not the primary cost constraint |
| Battery & thermals | Higher power requirements for sustained inference | Form factor constraints limit agentic workloads on mobile devices |
IDC estimates that high memory costs could cause global PC shipments to decline approximately 9% in 2026. The AI hardware boom is colliding with a memory supply crunch. Component costs—memory, AI silicon, battery, thermals—are rising faster than consumer willingness to pay. The companies that manage this cost pressure while delivering genuine AI capability will be the ones that win the next upgrade cycle.
What to Watch Next
- AI PC adoption: Whether 150 million shipments in 2026 translate into actual AI usage, not just hardware sales, will determine the category's credibility.
- NPU utilization: The gap between NPU TOPS and actual workloads running on NPUs will indicate whether the hardware is being used effectively.
- Local inference: The share of AI tasks processed on-device versus in the cloud will shape the next generation of hardware requirements.
- AI wearables: Smart glasses growth rates will show whether wearables become a mainstream AI category or remain a niche.
- GPU shipments: The leading vendor's share and the impact of memory prices on GPU pricing will determine the economics of AI compute.
- Memory costs: DRAM and NAND pricing trends will dictate whether AI hardware becomes more affordable or remains premium.
- Agentic-device launches: The first devices designed specifically for autonomous agents—rather than retrofitted with AI features—will define the next hardware category.
The AI hardware race is not about adding NPUs to existing devices. It is about designing machines that can run autonomous agents.
The first wave of AI hardware was defined by branding. The next wave will be defined by capability. Agents require sustained inference, local context, battery efficiency, and security—characteristics that NPU marketing does not address. The hardware that wins will be the hardware that enables seamless task routing between local and cloud resources, optimized for agentic workloads rather than AI inference benchmarks.
Memory costs are the wild card. With DRAM and NAND prices projected to rise more than 130% by the end of 2026, the AI hardware boom is colliding with a supply crunch. The companies that manage this cost pressure while delivering genuine AI capability will win the next upgrade cycle.
Sources
- Lenovo StoryHub — Hybrid AI for Business: New Think-Family Devices, Displays, and Security Solutions
- Tom's Hardware — Desktop graphics card shipments hit four-year high of 12.5 million
- Goldman Sachs — AI PC shipment projections 2026
- Omdia — PC shipment NPU penetration estimates
- Jon Peddie Research — GPU shipment data
- IDC — AI smartphone and PC shipment forecasts
- Qualcomm Dragonwing Q-2390 and IQ-2390 processor announcements
- Global smart headset shipment data, Q1 2026
The CODEW Stat
150 million AI PC shipments projected in 2026 — 59% of all PCs. Desktop discrete GPU shipments hit 12.5 million in Q2 2026, with the leading vendor holding a 90% share. Smart glasses shipments grew 83% year-on-year. And DRAM/NAND prices are projected to rise more than 130% by the end of 2026.
Reviewed by Erwin Castro
on
Wednesday, September 16, 2026
Rating:
