Mara ReyesQualcomm is targeting $15 billion in data-center chip sales by 2029. In 2024, it gave developers models optimized for 45-TOPS PC NPUs. Both moves are technically coherent, which makes the commercial contradiction harder to dismiss.
Key takeaways
- Edge AI has lost its role as the AI buildout’s main commercial narrative because agentic workloads favor pooled utilization, shared orchestration and centralized capacity—even though local compute remains technically useful.
- Running a model on an endpoint is not sufficient to create a market: buyers need frequent, valuable or sensitive decisions that justify redesigning workflows around local execution.
- Robotics offers the clearest durable case for edge compute because machines may need to perceive and act despite network latency or connectivity failures.
- Qualcomm is repositioning from an endpoint-chip vendor into a hybrid compute supplier spanning devices, robots, software, custom silicon and data-center CPUs.
- Qualcomm’s data-center ambitions remain targets rather than demonstrated traction: Meta’s expected Dragonfly C1000 use establishes a prospective deployment, but not shipment volume, contracted revenue or customer breadth.
Buyers need more than a working endpoint model
Phone, PC, vehicle, and machine designers cannot buy their way around power, memory, cooling, and connectivity constraints. A data-center operator can. Qualcomm paired efficient processing with radios so devices could keep more work near the user.
By making the 45-TOPS models available, Qualcomm removed one technical bottleneck for Snapdragon X Elite developers. Developers still had to decide which answers belonged on the device. A local model proves that an endpoint can produce an answer; it does not show that the answer is valuable, frequent, or sensitive enough to organize an application around local execution. Buyers must redesign data collection, judgment, workflow, and action around that answer.
In June 2026, Qualcomm paired the $15 billion goal with a higher 2029 non-handset chip-revenue target, raising it from $22 billion to $40 billion. Qualcomm presented both figures as targets rather than booked sales. The company must now compete for centralized workloads as well as endpoint sockets.
Agents reward busy racks
Agents change the unit of inference economics. A local assistant may answer an occasional prompt with a bounded model. An agent can invoke models repeatedly, use tools, inspect results, revise its path, and coordinate work across services. For an agent, the useful unit is the cost of completing a task across a chain of inferences.
Cloud operators can schedule shared capacity across workloads, pair it with high-performance networking, and support it with a common orchestration layer. The infrastructure has addresses: racks, cooling systems, fiber routes, substations, and contracts for powered capacity. The cloud is a building, and the expensive equipment inside has to stay busy.
On July 1, 2026, The Information reported that OpenAI engineers had told colleagues they had found a way to more than halve inference costs. The report described an internal result, not a production rollout or realized demand. If OpenAI deploys that result, each dollar could buy more agent steps on pooled systems without making every step local.
Barclays’ forecast does not settle where inference will run. It shows the scale of capital chasing pooled utilization: when agents multiply model calls, the cost of completing a task can matter more than the cost of one local answer.
The Dragonfly C1000 gives Qualcomm a concrete test. Qualcomm describes it as a power-efficient data-center CPU built for agentic AI, and Meta is expected to deploy it when production begins in 2028. Qualcomm has not established shipment volume or revenue with that announcement. It has shown a company built around efficient endpoints applying the same logic to Meta’s racks.
Arm already runs on both sides of the network
In a June 25, 2026 interview with Nikkei Asia, Arm executive vice president Mohamed Awad said the company’s architecture accounted for more than half of the hyperscale cloud-computing market. Arm’s 2023 IPO prospectus had estimated a 10% share of the $18 billion cloud-processor market.
Those figures define different markets and cannot form a clean share series. They still establish a substantial data-center presence for Arm-based designs; neither figure says Qualcomm will win there.
Engineers place each task according to its latency, reliability, privacy, connectivity, power, and utilization requirements. They build heterogeneous AI systems from processors in different locations, coordinated by software that decides where work belongs.
When several processor classes can run a workload, schedulers, programming layers, and networks help determine which chip receives it. Qualcomm still has to earn that software control point, especially when buyers can substitute among processors.
Robots keep the strongest case for local compute
A robot acting in the physical world gives local compute its clearest case. Sensors generate local conditions, motors require action, and connectivity can fail. The machine still has to act.
On January 5, 2026, CNET reported that Qualcomm had unveiled Dragonwing IQ10 as a hardware, software, and AI platform for industrial and consumer humanoids. The announcement identifies the intended market; it does not provide deployment counts, safety results, or operating costs.
A platform and processor can make local action possible. A factory or warehouse still has to integrate sensing, judgment, human supervision, maintenance, safety, and machine control into one operating process.
A factory has a clear reason to pay for local perception when a moving machine would otherwise wait on a network round trip. If the robot sends most deliberation to a centralized model and reserves its processor for control, it remains a satellite of the cloud, however sophisticated its board.
Qualcomm wants software to govern the stack
Qualcomm’s planned acquisition of Modular for nearly $4 billion would add a chip-software platform and a proprietary programming language. If the deal closes and the software works across unlike processors, Qualcomm could connect its endpoint compute, robotics platforms, networking, custom silicon, and data-center CPUs.
Qualcomm says Meta will use the C1000 when production begins in 2028, one year before the $15 billion target date. The company still has not established volume, contracted revenue, or customer breadth.
If developers adopt Modular broadly, Qualcomm could influence which chip receives a task and where the economics accumulate. Qualcomm has not yet shown that adoption, but the deal identifies the control point it wants: software that can route workloads across more processors than any single socket win.
Qualcomm gave developers a 45-TOPS endpoint option in 2024. It plans to begin Dragonfly production in 2028, one year before its $15 billion target date. Qualcomm’s edge playbook has brought it back to the building. There, buyers measure efficiency in deployed racks and booked revenue.
Qualcomm’s compressed data-center timetable
- June 25, 2026 — Qualcomm unveiled the Dragonfly C1000, projected $15 billion in data-center chip sales by 2029 and raised its fiscal 2029 non-handset revenue forecast from $22 billion to $40 billion.
- H2 2026 — Qualcomm’s nearly $4 billion acquisition of chip-software company Modular is expected to close.
- 2028 — Dragonfly C1000 production is scheduled to begin, with Qualcomm saying Meta will use the processor.
- 2029 — Qualcomm’s $15 billion data-center chip-sales target comes due, leaving one year after the expected Meta production start to convert deployment into large-scale revenue.
Frequently asked questions
Why are AI agents shifting value toward data centers?
Agents can make repeated model calls, use tools and revise their work, making total task cost more important than the cost of one inference. Cloud operators can spread those calls across shared, highly utilized infrastructure with common networking and orchestration.
Does Qualcomm’s strategy mean edge AI is dead?
No. Local compute remains important when latency, privacy, reliability or connectivity require decisions on the device, with moving robots providing the strongest example.
What does Meta’s Dragonfly C1000 commitment prove?
It shows that Meta is expected to use Qualcomm’s agentic-AI data-center CPU when production starts in 2028. It does not establish deployment scale, revenue or broader customer demand.
Why does Qualcomm want to acquire Modular?
Modular could give Qualcomm a software layer for directing workloads across different processors and locations. That would provide a broader control point than an individual chip sale, although widespread developer adoption has not been demonstrated.
What would justify a distinct edge hardware and software stack?
A deployment must generate decisions that are valuable and time-sensitive enough to require local execution—not merely prove that a local model can run. Buyers must also integrate sensing, judgment, action, supervision and maintenance into an operating workflow.