Qualcomm forecasts $15 billion in data center revenue by 2029

The forecast came on the Q4 2026 earnings call. The shape of the ramp matters more than the endpoint. Qualcomm expects roughly $650 million in the first quarter of 2027, $5 billion by the end of 2027, and $15 billion by 2029. For scale, The Verge put total revenue for 2026 at $10.4 billion, so the 2029 data center target exceeds its entire 2026 revenue.

The phone business is shrinking. Smartphone sales fell around 20 percent compared with 2025.

Qualcomm's data center push runs through partnerships with Amazon and Meta announced earlier this year. The company says data centers will eventually account for around 30 percent of its business. That makes the forecast a target rather than booked business, and buyers of AI compute should read it that way.

The benchmark AMD has set

AMD has set a far higher bar. It says it expects more than 80 percent of its revenue eventually to come from data centers, against Qualcomm's roughly 30 percent. The gap is partly ambition and partly arithmetic. Qualcomm still has a shrinking handset business to defend, and data centers are only forecast to reach around 30 percent of its business.

Memory stacked on the compute die

The silicon is still in validation. Qualcomm said it had completed initial validation of its High Bandwidth Compute memory architecture, which stacks layers of DRAM on top of a small compute die. Putting memory there cuts the energy spent moving data, according to ServeTheHome.

The Dragonfly AI250 accelerator is scheduled to start sampling in 2027.

Cheaper packaging is part of the pitch. Because less bandwidth is needed between the stacked die and the main processor, Qualcomm can use traditional 2D organic substrates instead of CoWoS or EMIB. The stacking still requires through-silicon vias, so manufacturing does not get simple.

Inference is where the capacity goes

Others are already selling inference capacity on hardware that exists. SambaNova and Supermicro said on 4 November they will build systems pairing Nvidia GPUs for prefill with SN50 RDUs for decode, starting with 10,000 chips. They argue operators can add inference capacity without waiting for buildouts constrained by power and cooling, and extend the life of GPUs they already own.

Argonne bought the same argument. Argonne National Laboratory said the same day that it added the Tara system to its inference service. Tara carries 2,688 Nvidia GH200 superchips, roughly 100 times the GPU memory previously available to the service. The lab also expanded Metis from two to six SambaRacks, for 96 SambaNova SN40L processors. For buyers, the question is less which vendor wins than whether inference capacity arrives before new power-constrained buildouts do.

What happens next

Two dates frame the next year. Dragonfly AI250 sampling is due in 2027. Qualcomm's forecast sets the end of 2027 as the deadline for $5 billion in data center revenue.