Qualcomm’s $15B Bet Moves the Edge Inward

Qualcomm expects $15B in fiscal-2029 data-center chip sales, while Dragonfly C1000 is scheduled to enter production only in 2028. The company built its AI position by moving computation outward, toward phones, PCs, and the network edge. Its central growth target now sits inside data centers bought by a handful of hyperscalers.

The route out of handset dependence runs inward

In 2019, Qualcomm positioned Cloud AI 100 as an edge-inference chip. By 2026, it had announced Dragonfly C1000, a data-center CPU scheduled for production in 2028. Over seven years, Qualcomm shifted the address where it expects its design advantages to earn their return.

Qualcomm first sought to place AI near devices without sending every task back to a centralized system. It paired customer relationships, power discipline, and developer ambitions with edge inference. The company must now make expanding workloads economical inside AI infrastructure, where processors operate within racks, networks, memory hierarchies, and power envelopes purchased by a few very large customers.

Qualcomm raised its fiscal-2029 non-handset revenue forecast from $22B to $40B and assigned $15B of expected data-center chip sales to that horizon. At that scale, data centers become one of the company’s main routes out of handset dependence.

Qualcomm says Meta will use the power-efficient Dragonfly C1000 when production begins in 2028. Even without disclosed volume or economics, a named hyperscale customer gives Qualcomm an announced route into a real deployment rather than a processor waiting for a market.

Meta’s deployment will test whether Qualcomm can carry its endpoint power discipline into racks that must also accommodate memory, networking, and accelerator budgets.

A data-center chip has lost the right to compete alone

A CPU can schedule work, an accelerator can execute it, an interconnect can move it, and memory can feed it. A deployable system needs all four. Compilers must expose the hardware, orchestration software must place workloads, and customers must be able to migrate code without rebuilding their operations around an unfamiliar platform.

Qualcomm is buying those capabilities. Its agreement to acquire Modular for nearly $4B, with closing expected in the second half of 2026, brings in a chip-software platform and proprietary coding language. Its completed acquisition of Alphawave Semi supplied connectivity technology that reportedly figures in talks to design custom chips for ByteDance. Together, the deals aim to make Dragonfly programmable, connected, and adaptable to each customer’s workload.

Qualcomm must now connect processor architecture, memory movement, networking, compilers, and customer support before deployment. Owning more layers raises its entry cost, but it can spare buyers from stitching together unfamiliar tools and increase switching costs after installation.

Customers increasingly define the available choices themselves. OpenAI supplied workload knowledge to Broadcom and took an LLM-optimized inference chip from design to manufacturing tape-out in nine months.

To displace an incumbent, Qualcomm must let Meta migrate code, connect Dragonfly to orchestration tools, and integrate it with existing networking and memory plans without multiplying operational exceptions. Customers will judge Dragonfly as a system before they judge it as a processor.

Inference opens the market by strengthening the buyer

Training created the defining AI clusters, but inference broadens the spending pool because deployed models must keep running after training ends. Barclays projected that inference capital spending would reach $208.2B in 2026 and surpass training spending within two years. Inference workloads vary in throughput, latency, memory use, power, and cost, so different deployments reward different architectures.

Nvidia limits that opening. It held more than 90% of the data-center GPU market when its quarterly data-center revenue reached $22.6B in 2024. Customers had already organized software, networking, procurement, and operations around its platform. Qualcomm, AMD, Intel, and other challengers therefore compete against accumulated operating decisions.

AI workloads have also reopened the CPU layer. By July 2026, Nvidia, Qualcomm, and MediaTek were among the companies challenging the Intel-AMD structure in data centers. Accelerators increased the importance of their surrounding components: hosts must coordinate larger systems, feed more specialized processors, and stay within power budgets that constrain deployment.

Hyperscalers can use inference’s workload diversity to finance alternatives tailored to their own needs. Qualcomm gains more routes into the market, but the buyers funding those routes can demand custom designs, integration support, and better terms.

The 2028 date is part of the product

Dragonfly arrives inside a construction and supply schedule. Data centers were projected to consume more than 70% of all high-end memory chips produced in 2026, while little new manufacturing capacity was expected before 2027. Memory has become an allocation regime: suppliers decide which systems receive scarce output, customers reserve it early, and a processor can be ready before the memory required to deploy it.

Construction and packaging carry similar lags. JPMorgan said more than 60% of U.S. data-center capacity planned for completion in 2027 was not yet under construction. Nvidia separately committed $1.5B to Amkor through an accord involving a prepayment to bolster Arizona packaging capacity. Buyers and suppliers are advancing money and securing physical capacity before the systems it supports are fully deployed.

Qualcomm gains an opening if Dragonfly enters while memory, packaging, and powered space still constrain deployment. Announced megawatts cannot run chips before buildings and substations are complete. Meta and any follow-on customer will need to secure those inputs before Qualcomm begins production.

A 2028 launch also crosses more of the capital cycle. Memory capacity expected after 2027, packaging financed during the shortage, and data centers moving from plan to construction can begin arriving around the same period as Qualcomm’s platform. Hyperscalers prize efficiency and supply access during shortages, but the factories and facilities financed by those shortages can weaken that advantage.

Projects will finish unevenly across locations and components. A hyperscaler evaluating Dragonfly in 2028 will compare its architecture with the memory, packaging, power, and rack capacity actually available at each site.

Scarcity finances the capacity that can erase it

By 2025, U.S. data-center capacity that was built, underway, planned, or stalled had topped 80 gigawatts. Only built capacity counts as supply; stalled projects remain paper claims, and plans still need substations. The 80-gigawatt pipeline nevertheless shows how strongly scarcity translated into investment commitments before the resulting capacity had a chance to earn a return.

Microsoft reportedly walked away from U.S. and European projects expected to consume 2 gigawatts, which TD Cowen interpreted as evidence of oversupply relative to then-current demand forecasts. After a three-year data-center boom produced surplus computing power in China, the country planned a national network to sell the excess. Together, Microsoft and Chinese operators show how enormous demand can coexist with capacity under the wrong contract or in the wrong location.

Long-duration capacity contracts have therefore become part of the semiconductor story. A reservation can finance construction, a customer commitment can support packaging expansion, and a named deployment can justify a chip roadmap. Contracts transfer risk to whoever must pay when workloads migrate, power is delayed, or hardware earns less than the scarcity-era assumptions embedded in its price.

Operators must match reserved racks to inference products whose revenue covers processors, memory, networking, cooling, power, and depreciation. High utilization alone cannot repair a contract priced for returns the workload never produces.

Google can keep building while Microsoft cancels projects because each operator needs capacity in particular locations, with suitable memory and power, under a viable customer contract. China’s surplus does not release the high-end chips or powered U.S. racks that another operator needs.

Hyperscaler relevance reverses the supplier’s leverage

With endpoint chips, Qualcomm placed its silicon near an enormous number of users. Its data-center ambition brings the company before a much smaller number of buyers with far greater ability to shape what gets built. Meta can anchor a launch. ByteDance can seek a custom design. OpenAI can help define a processor with Broadcom. Microsoft can leave projects that no longer match its forecast.

Before Meta can deploy Dragonfly, it must validate Qualcomm’s compilers against existing code, connect the CPU to its orchestration and networking, and assign powered racks. Qualcomm, in turn, must line up packaging and compatible memory for the production window.

The $15B target requires Meta’s announced use to become volume orders and custom-design talks such as ByteDance’s to become deployments. Those orders must land while memory, packaging, and powered capacity are available.

The company that spent years carrying AI outward in milliwatts now seeks $15B at the center, where the sale is measured in megawatts and the chip is one line on a blueprint drawn by the buyer.