Qualcomm’s $15B Route From AI PCs to Data Centers

Qualcomm’s 45-TOPS NPU cleared the 40-TOPS threshold for a next-generation AI PC, yet it did not create a separate “edge AI” purchase order. The compute arrived inside a laptop refresh. Cloud compute arrives with a contract, a power requirement, and several billion dollars of punctuation.

NPUs disappear into the PC refresh cycle

In 2024, Intel executives said next-generation AI PCs would require 40 TOPS of NPU performance. Qualcomm then distributed models optimized for the 45-TOPS Hexagon NPU in Snapdragon X Elite. The hardware cleared the threshold, developers gained models prepared for it, and PC makers gained another reason to replace existing machines.

For a corporate IT buyer, the budget records a PC refresh while developers inherit a local accelerator. Each order adds inference capacity under the accounting label for a computer. Phone, vehicle, camera, and industrial-system makers use the same route through existing procurement cycles.

Apple and Perplexity route work task by task

Apple made the structure explicit when it described Apple Intelligence as a roughly 3-billion-parameter on-device model paired with a larger model on Apple silicon servers through Private Cloud Compute. The system places work according to privacy, memory, latency, capability, and elasticity.

Two years later, Perplexity split Computer tasks between local and cloud models, emphasizing private data and token efficiency. Application teams lose margin on unnecessary remote calls and user value when underpowered local models fail.

Teams should optimize cost per useful task while meeting requirements for privacy, latency, reliability, memory, power, and model quality. A small local model saves little when poor results erase user value.

Chip benchmarks measure available capacity, not whether an application uses it well. Developers capture the value of local prediction only after redesigning the workflow so routine operations stay near the user and demanding workloads escalate automatically.

Modular turns TOPS into usable capacity

A 40- or 45-TOPS NPU sits idle until developers can build useful applications for it. Qualcomm’s optimized model release gave them a path from silicon capability to product capacity.

Qualcomm agreed to acquire Modular for nearly $4 billion. Modular builds a chip software platform and a proprietary programming language. Its platform translates models across chips, optimizes execution, and helps developers deploy software across Qualcomm’s endpoint business and planned data-center products.

Developers then face portability problems across generations of endpoint silicon, cloud hardware, and different memory and power envelopes. Compilers and runtimes can inspect the task, available models, device state, privacy requirements, and escalation cost before allocating the work.

Dragonfly adds a separately priced revenue line

Qualcomm says Meta will use its Dragonfly C1000 data-center CPU when production starts in 2028. The company also projects $15 billion in data-center chip sales by 2029 and raised its non-handset chip-revenue forecast from $22 billion to $40 billion.

If the Modular deal closes, Qualcomm would sell device silicon, data-center processors, and the software layer that maps models across them. Qualcomm could collect silicon or software revenue as application teams change workload placement.

Qualcomm’s device business earns through unit volumes and silicon content, limiting how directly it can monetize each additional inference task. Data-center processors give the company a separately forecast revenue line tied to centralized build-outs.

Customers still decide whether local AI pays

Customers reward local AI when it changes the device they buy or lowers a provider’s serving bill.

Phones impose hard limits on memory, processing power, and energy. Local models must fit, run reliably, and improve the experience enough to influence a replacement decision.

Cloud providers can point to visible demand. Google said it needed to double AI compute capacity every six months to keep up. Centralized infrastructure translates that demand into power, accelerators, reserved capacity, and revenue forecasts.

An enterprise buyer can put privacy-sensitive or latency-critical work on NPU-equipped PCs and reserve cloud capacity for models that exceed the device. A common toolchain spares application teams from hand-building placement logic across chips, models, and device classes.

The 45-TOPS NPU is easy to miss because it is purchased as part of a PC; the data center is easy to see because it is sold as capacity. They are two destinations within the same workflow. The strategic unit is the stack that reads memory, privacy, latency, power, model quality, and price, then places each task where it belongs.