A typical AI request travels from a device to a data center, where it gets processed on specialized hardware before making the long roundtrip home. This AI in-the-cloud model is the foundation of the trillion-dollar data center boom.
Now comes AI “on-device.” In this setup, all computation happens on the chip inside a PC, laptop or phone—no long journey to a data center required. While the cloud is built to train large language models (LLMs) at scale, devices are typically meant to deliver constant access, privacy and speed.
The implication: As the AI rollout shifts from training LLMs in the cloud to running on local devices, the underlying hardware must be rebuilt to meet new, greater computational demands.
This shift is accelerating with the rise of agentic AI. At the Computex exhibition in June, Nvidia unveiled the RTX Spark, a chip that affords PCs enough horsepower to run AI agents. Days later, Apple announced a fundamental overhaul of Siri, from an answer-fetching voice assistant to one capable of orchestrating tasks across apps without a connection to the cloud.
These were not incremental product updates, in our view. Rather, we believe they marked the start of a deeper architectural rebuild addressing three structural limits to the broader AI rollout:
- Latency. As orders from bots overtake human-generated commands, servers in distant data centers may not be able to keep up, slowing response times. Cloudflare, an internet infrastructure company that helps websites run faster and more secure, estimates that 57% of online search requests are now initiated by bots, compared with 43% by humans. (For context, a human might browse five sites before making a purchase, while an AI service might browse 5,000.)1
- Privacy. Agents need continuous access to sensitive data that organizations may not wish to stream to a third-party cloud provider.
- Cost. Agents consume tokens at a geometric pace. Running relatively simple agentic tasks on specific devices, rather than in the cloud, could reduce AI costs for customers. In the cloud, users pay for every task an agent performs; on a device, they pay for processing power up front, without running up a massive AI tab.
We believe easing these crucial structural constraints will require improved hardware within the two computational layers that sit atop the cloud: the “edge” layer and the device layer.
The Edge Layer
Edge infrastructure is housed in regional facilities that sit between data centers and device users, shortening the distance that AI requests must travel and helping coordinate commands faster than any single device can manage. There are two broad types of players within the edge layer:
- Content delivery networks. These companies, such as Akamai and Cloudflare, speed up website loading times by running thousands of local servers worldwide. Now they aim to repurpose those same footprints to accelerate AI response times.
- GaaS providers. These companies, such as DigitalOcean, lease AI-computing power by the hour. This GPU-as-a-service (GaaS) model aims to accommodate AI usage, which tends to be bursty and machine-driven. As companies deploy more AI agents, we believe they will seek the fastest and cheapest solutions—whether that means renting a nearby GPU or adding computing power on their local devices.
The Device Layer
Increasingly, AI computation is being done on specific devices rather than in the cloud. At the high-performance end, agentic workstations—PCs with enough processing power and memory to run AI models—are emerging as a new product category. Managing immense workloads on-device requires more processing power, memory and storage capacity, and precision “packaging” to bind it all together.
Processors
Traditional designs were built to execute general-purpose tasks one at a time, not to handle armies of AI agents churning at a constant, feverish pace. The fix is to add dedicated AI processing units: circuits built only for AI calculations, distinct from the general-purpose cores that handle other tasks.
We believe rising demand for more on-device processing power is driving a new, multi-year hardware cycle, with market leaders pursuing it from different angles:
- Nvidia and MediaTek provide the RTX Spark platform: the combined chip-and-software foundation that other products are built on.
- Arm Holdings supplies the architectural backbone: the underlying blueprint that super-efficient chips are built from.
- Qualcomm specializes in ultra-low-power design: chips that deliver AI performance while sipping battery life, essential for devices that are always on.
Memory and Storage
Memory is the fast, short-term workspace a chip uses while actively thinking; storage is the slower, long-term space that holds files when a device is off. Agentic AI demands far more of both, hence the stunning rally in the PHLX Semiconductor Index, which shot up nearly 300% between April 2025 and late June 2026, driven by memory leaders including SK Hynix, Samsung Electronics and Micron Technology, and storage leaders such as SanDisk and Kioxia. (Several of these companies make both memory and storage chips.)
Worried that some chipmakers may be adding too much capacity, investors pumped the breaks in July, causing the PHLX to tumble 15% from its peak. Yet we believe we are still in the early innings of this upgrade cycle and that demand for memory and storage will continue to outpace growth in supply, which will take at least a few years to come online.
Packaging
“Packaging” connects chips within a device. Its core element is the IC substrate: a base layer that links the silicon chip to the circuit board via thousands of tiny electrical connections. The global IC substrate market—led by key suppliers including Ibiden Unimicron, SEMCO and AT&S—is projected to grow 6.8% a year and hit $20.6 billion by 2033.2
Increasing Demand, Constrained Supply
The same component makers feeding the data center buildout are the ones now faced with equipping the on-device AI expansion as well. Those two waves of demand crashing on one constrained supply base are driving an historic surge in hardware pricing.
As AI computation moves from the cloud toward the edge and onto local devices, we are keeping a discerning eye on the sector enabling that shift.