Qualcomm said on June 24 that Microsoft and Meta will adopt its newest AI chips, a breakthrough for a company trying to move beyond the smartphone market that built it. It also said it is developing custom silicon for two additional, unnamed hyperscale cloud providers.
Microsoft will use a chip architecture Qualcomm calls high-bandwidth computing, or HBC. The design leans on the kind of inexpensive, widely available memory used in phones and laptops—standard LPDDR-class DRAM—rather than the premium high-bandwidth memory that Nvidia’s accelerators require. Meta, meanwhile, will deploy Dragonfly C1000, a Qualcomm-designed processor built specifically for AI data centers.
The announcements are a bet that the economics of AI inference—the stage where trained models answer questions—will reward frugality. Running an AI model in production costs money on every single request, and memory is the largest line item. Qualcomm’s pitch is simple: most inference workloads do not need the fastest memory in the world, only enough bandwidth at a fraction of the price.
That puts Qualcomm in direct competition with a market Nvidia has dominated. Nvidia’s GPUs, paired with high-bandwidth memory made by SK Hynix, Samsung and Micron, remain the default for training and for the heaviest inference loads. But a growing share of AI traffic is lighter: summarizing documents, generating captions, answering support questions. For those workloads, analysts said, a cheaper memory configuration can cut total cost of ownership sharply, and cloud providers are increasingly price-sensitive as their AI bills mount.
The company’s approach echoes the strategy it used to disrupt the mobile industry. In phones, Qualcomm’s processors won by integrating modems, graphics and AI accelerators into a single chip, letting handset makers ship complete devices without buying from multiple suppliers. In data centers, the company is offering a similar package: standard memory, efficient design, and a roadmap that promises frequent improvements. Its engineers argue that matching memory bandwidth to the actual workload, rather than buying the fastest memory available, is the discipline the industry has been missing.
Qualcomm’s own CPU for data centers, the Dragonfly line, is central to the pitch. Naming a product after a predator of smaller insects was deliberate, executives have said, and the C1000 aims at the power-hungry processors that dominate server fleets today. Meta’s decision to use it for AI workloads gives the product a marquee customer and a proof point it lacked.
The custom-chip work for the two unnamed hyperscalers follows a pattern set by Google and Amazon, which have designed their own accelerators to reduce dependence on Nvidia. Broadcom and Marvell have built substantial businesses helping cloud providers design such chips, and Qualcomm is now offering the same service with its own intellectual property. Custom silicon does not threaten Nvidia’s dominance by itself, but it erodes the economics: every chip a cloud provider designs in-house is a GPU sale that never happens.
Analysts were measured in their response. Qualcomm has announced data-center ambitions before, and shipping in volume is harder than signing a customer. “The win is real,” said one analyst who covers the chip industry. “The question is how many chips Meta and Microsoft actually deploy, and how fast Qualcomm can scale supply.”
The timing is favorable. AI workloads are shifting from training to inference, and inference is where cost discipline matters most. Cloud providers are under pressure to show that AI spending produces profit, which pushes them toward cheaper ways to serve the same requests. Qualcomm’s memory strategy is aimed squarely at that pressure, and its history of shipping chips in the hundreds of millions of units gives it manufacturing scale most data-center newcomers lack.
Qualcomm also has something else most challengers do not: a balance sheet built on the phone business. The company generates billions in quarterly profit from handset royalties and chips, which lets it fund data-center development without depending on venture capital or early customer prepayments. Rivals that entered the AI chip race earlier have had to raise money or lean on partnerships; Qualcomm can simply write the checks.
The company has not disclosed shipping volumes or revenue expectations, and executives said only that production would ramp through next year. Supply is also a question: Qualcomm outsources manufacturing, and foundry capacity for AI chips is tight across the industry.
For Microsoft, the decision fits a pattern of hedging. The company is Nvidia’s largest customer, but it has also invested in its own accelerators and in custom designs through partnerships. Adopting Qualcomm’s architecture gives it a second, cheaper option for inference at a time when its AI capital spending is under investor scrutiny.
For Meta, the Dragonfly C1000 becomes part of a strategy of in-house infrastructure built around open hardware designs. The company has long pushed for alternatives to the dominant suppliers, and its AI recommendation systems—the engines behind its ad business—run at enormous scale, where small per-request savings add up.
None of this displaces Nvidia next quarter. But the direction is clear: the biggest buyers of AI compute are assembling a portfolio of options, and Qualcomm has just inserted itself into the list.
