On-device agents: architectural truth or Qualcomm's growth story?

On the first day of Computex 2026, Qualcomm CEO Cristiano Amon declared that the shift to agentic AI makes local edge computing "unavoidable." NVIDIA launched RTX Spark the same morning, naming OpenAI, Anthropic, and SpaceX as early adopters. Microsoft's Copilot+ PC platform had already encoded a minimum on-device AI processing requirement. The coordination is notable. The question is whether it reflects genuine architectural convergence or a jointly constructed narrative that happens to serve all three companies' interests — and two of those three companies' interests considerably more than the other.

What was actually said. Amon's framing at Computex was direct: "Current device architectures were built for user-initiated actions, not always-on autonomous agents." The implication is structural. A user-initiated action generates a discrete API call — you ask, the cloud responds, the session closes. An always-on autonomous agent, by contrast, requires continuous perception, continuous inference, and continuous action. If that inference runs in the cloud, the per-call economics (the model by which AWS Bedrock, Azure AI, and Google Vertex generate AI revenue) become a liability rather than a feature. The cloud-inference model, Amon is suggesting, was built for the wrong paradigm.

This is a substantive claim. It may also be correct. But it is worth noting who is making it.

Qualcomm needs this story to be true. Qualcomm's QCT (Qualcomm CDMA Technologies) handset revenue for fiscal Q2 2026 was $6.0 billion. QCT IoT (which includes PC chipsets) was $1.4 billion. The company declined, on its earnings call, to break out AI-specific PC chipset revenue from that IoT line — which is itself a signal that the number is not yet large enough to be promotional.¹ The modem business faces long-run structural pressure: Apple is building its own modem, MediaTek is competitive, and the smartphone upgrade cycle is not what it was. On-device AI processing — specifically the NPU (neural processing unit, a chip designed to accelerate AI workloads locally) throughput story built around Snapdragon X Elite — is Qualcomm's primary growth narrative for the PC market.

The "agentic AI requires local compute" thesis, if it becomes the industry consensus, expands Qualcomm's addressable market materially. The company's incentive to believe this thesis is, to use a precise term, approximately 100%.

That is not evidence the thesis is wrong. Incentive-motivated arguments can still be correct arguments. But it is worth holding Amon's framing against the actual fiscal disclosures, which do not yet support the claim that on-device AI is a material revenue driver.

NVIDIA's position is structurally different. NVIDIA launched RTX Spark on the same morning as Amon's remarks — a local AI inference (running AI models on hardware you own, rather than calling a remote server) platform for personal computers. The simultaneity looks like coordination. But NVIDIA's incentive structure is not Qualcomm's. NVIDIA holds a dominant position in data-centre GPU sales regardless of where inference ultimately runs. If the market moves to on-device, NVIDIA has RTX. If it stays in the cloud, NVIDIA has H100s and B200s. NVIDIA is hedging; Qualcomm is betting.

The RTX Spark early-adopter list is its own puzzle. OpenAI and Anthropic are named. Both companies' primary revenue model is API-based cloud inference — the model that the on-device thesis structurally threatens. Their appearance as local-inference advocates is either a hedge (supporting on-device to remain relevant wherever inference migrates) or a signal that the local models they're shipping are sufficiently parameter-constrained (small enough in scale) that they don't cannibalise their own cloud revenue. A 7B parameter (a rough measure of model complexity and capability) on-device model does not replace a frontier cloud model for complex agentic reasoning. It complements it, handles simpler tasks locally, and routes harder calls back to the API. That is not the death of the per-call model; it is an extension of it.

Microsoft's role is structural rather than rhetorical. Microsoft has not made a formal Computex statement endorsing on-device agentic AI. It doesn't need to. The Copilot+ PC platform specification, published in May 2024, mandated 40+ TOPS (tera operations per second — a throughput measure for AI chip performance) of NPU capability as a baseline Windows PC requirement.² That architectural requirement is already baked into the PC supply chain. Microsoft doesn't need to say "on-device AI is the future" at a trade show; it already wrote that into the platform standard. This structurally validates the Qualcomm and NVIDIA framing without Microsoft having to take a public position.

What Microsoft has not disclosed: whether Copilot+ NPU utilisation is material. A required capability is not a used capability. The adoption rate and actual workload profile of Copilot+ NPUs in the wild is unknown.

The inference-economics question, mapped. Inference economics — the cost structure of running AI models at scale — currently favours the cloud on a per-call basis for most enterprise use cases. Cost-per-token for frontier cloud models has fallen sharply through 2025 and into 2026. On-device inference is faster improving but starts from a different cost base and a different capability ceiling. For genuinely complex agentic reasoning (multi-step planning, real-time web access, tool use), cloud models remain both cheaper and more capable than on-device alternatives.

The on-device economic case sharpens under two conditions: first, if agents run continuously (generating very large call volumes that accumulate into significant cloud spend); second, if on-device model quality closes the capability gap with frontier models. Neither condition is established yet. The first is a plausible future for some consumer use cases. The second is a harder technical problem.

The hyperscalers are not obviously threatened today. The question is whether the trajectory of on-device model capability changes that arithmetic within two or three hardware generations.

What to watch. Qualcomm's fiscal Q3 2026 earnings (due late July) should be the first data point where management either breaks out AI-specific PC chipset revenue or continues to absorb it into the IoT line. Continued non-disclosure is informative. On the NVIDIA side, RTX Spark adoption rates and whether frontier-lab local model offerings gain traction with enterprise buyers will test whether the weak or strong form of the on-device thesis is operative. And watch for the first hyperscaler response — AWS, Azure, or Google publicly quantifying what share of their agent workloads run locally versus cloud would reset the analytical baseline entirely. None of them have disclosed this, which is itself a gap worth noting.

The coordinated Computex messaging is a market-structure move. Whether it maps to architectural reality depends on data that does not yet exist. Amon may be right. He is also the person who most needs to be right, which is worth keeping in the margin.

Footnotes

¶