Cristiano Amon, chief executive officer of Qualcomm Inc.
© 2026 Bloomberg Finance LP
The economics of artificial intelligence are shifting, placing the data center under strain. Agentic AI, in which software agents chain together dozens of model calls to complete a single task, has turned inference from a modest tax into the dominant workload.
A single agentic query now generates 50 to 100 times as many inference requests as a conventional prompt, and the infrastructure built for the training era struggles to absorb that volume. The bottleneck has moved beyond raw compute to memory bandwidth, power per token, and the cost of moving data across a rack.
That gap is the opening Qualcomm walked through at its June 2026 Investor Day in New York. The company best known for putting silicon into roughly six billion smartphones declared its intent to become a data center vendor, anchored by a new infrastructure platform called Dragonfly, the strategic acquisition of Modular, and a financial target of fifteen billion dollars in data center revenue by fiscal 2029.
The obvious question, which Qualcomm CEO Cristiano Amon raised before anyone in the audience could, is whether the company is too late to enter a market already dominated by NVIDIA and crowded with in-house designs from hyperscalers. The more interesting question is whether Qualcomm brings anything the incumbents lack.
It does. The case rests on three pillars competitors will struggle to assemble: a memory architecture designed to address the inference bottleneck, a credible software story, and three decades of connectivity engineering that maps directly to the problem data centers face today.
Qualcomm structured its data center strategy around four product lines that layer on top of one another, mirroring the approach it used in automotive, where Qualcomm entered through connectivity and the digital cockpit before expanding into advanced driver assistance.
In the data center, the sequence runs from connectivity silicon, available now from its Alphawave acquisition, through custom silicon ramping in early fiscal 2027 (Qualcomm’s fiscal year begins after the last Sunday in September), AI accelerators in the second half of that year, and Oryon server-class CPUs arriving in mid-2028.
Qualcomm Dragonfly Portfolio
Qualcomm Inc.
The centerpiece of Qualcomm’s accelerated compute strategy is what it calls high-bandwidth compute, or HBC. Conventional AI accelerators pair a processor with stacks of high-bandwidth memory connected by tens of thousands of wires across an expensive silicon interposer, an arrangement that burns power and generates heat as data shuttles back and forth. HBM is also in scarce supply and sits at the center of the current memory supply chain shortage.
Qualcomm is taking a non-traditional approach, placing its processing cores directly beneath a DRAM stack, collapsing the distance data must travel. The company claims this delivers up to 8 times more tokens per watt than traditional GPU configurations and 6 times the memory bandwidth per watt of HBM-based competitors, while eliminating the need for a silicon interposer entirely.
Accelerators based on Qualcomm’s HBC approach are sampling now, with subsequent generations planned on an annual cadence.
Qualcomm Accelerator Roadmap
Qualcomm Inc.
Lending credibility to the architecture is the customer validation Qualcomm presented alongside the announcement. Microsoft CEO Satya Nadella confirmed that Azure will deploy HBC, and Meta CEO Mark Zuckerberg committed to a multigenerational deal for Qualcomm CPUs in Meta data centers. The first HBC accelerator, the AI250, is targeted for mid-2027, with the AI300 following in 2028.
The most challenging part of driving adoption of a new silicon platform is, ironically, the software stack supporting it. NVIDIA has proven this with the dominance of its CUDA stack, the mature software layer that locks developers to its hardware and has given it nearly unshakable dominance in AI model training.
Any company entering this market with strong chips and a weak software stack operates at a deficit. This is the motivation behind Qualcomm’s acquisition of Modular, perhaps the most consequential announcement of its Investor Day. Modular is a software company founded by Chris Lattner, the engineer behind the LLVM compiler infrastructure, Apple’s Swift programming language, and much of Google’s TPU software stack.
Modular has spent more than four years building a portable alternative to CUDA, designed from the start to run AI models on any vendor's silicon. Its stack spans the Mojo programming language, the MAX serving layer, and a distributed cloud serving infrastructure. The company claims inference performance up to fifty percent faster than incumbent options on third-party hardware.
The value of Modular extends beyond enabling Qualcomm’s accelerators to be usable on day one. Modular gives Qualcomm a hardware-agnostic software story by design, which entirely reframes the company’s pitch.
Rather than asking customers to abandon NVIDIA, Qualcomm can offer a layer that runs across NVIDIA, AMD, and its own silicon, positioning itself as the open alternative in a market where hyperscalers are actively seeking to avoid single-vendor lock-in.
The Modular acquisition matters for several reasons:
Qualcomm reinforced the software story with a new partnership with Hugging Face, giving its sixteen million developers a path to deploy open models across Qualcomm platforms.
Qualcomm is entering the data center CPU market amid a structural realignment. Hyperscalers have spent the past several years designing their own processors to cut costs, improve margins, and reduce dependence on Intel and AMD.
AWS Graviton, Google Axion, and Microsoft Cobalt are custom Arm-based server CPUs, and the trend has reached a tipping point. According to Arm's majority shareholder, Softbank, Arm-based processors now account for 50% of new hyperscaler CPU compute. AWS supports this, claiming that its Arm-based Graviton processors represent more than half of all new CPU capacity it has added over the past three years.
Analyst firm Counterpoint projects that Arm-based CPUs could account for roughly 90% of host CPU deployments in custom AI ASIC servers by 2029, up from about 25% in 2025. The driver is performance per watt, the single metric that matters most in power-constrained facilities, where Arm designs deliver 30% to 60% better energy efficiency than comparable x86 parts.
Qualcomm’s C1000 CPU steps directly into this current. The company claims cores running above five gigahertz, more than thirty percent faster than competing designs, with configurations exceeding 250 cores, all built on the Oryon architecture Qualcomm developed in-house through its acquisition of Nuvia.
Qualcomm offers two paths to hyperscalers that competitors rarely combine. It will sell its own CPUs, as the Meta deal demonstrates, and it will design custom silicon for hyperscalers who want bespoke parts. It reported two major hyperscaler wins, each expected to exceed one billion dollars in fiscal 2027.
The custom silicon business is where Qualcomm’s manufacturing scale becomes a weapon. The company consumes over a million leading-node wafers annually and ships forty billion components, a scale that lets it commit capacity and guarantee yield in a supply-constrained market where smaller design houses cannot.
This market is also among the industry’s most competitive. In its most recent earnings report, Broadcom reported $10.8 billion in AI-related revenue, up a staggering 143% year over year. Marvell, a smaller player in this segment, reported 28% growth and is guiding to 40% growth for the current year.
The competitive risk is real and should not be understated. NVIDIA’s Grace and Vera CPUs, AMD’s EPYC line, and the hyperscalers’ internal programs leveraging companies like Broadcom and Marvell, all compete for the same sockets. Arm itself has begun shipping finished silicon that could compete with its licensees.
The element of Qualcomm’s strategy that is hardest for rivals to replicate is the one most deeply rooted in its history. Qualcomm built its business on wireless connectivity, authoring the standards that moved the industry through 3G, 4G, and 5G.
Connectivity is now the third major bottleneck in the data center, after memory and software. As AI compute roughly doubles every two years, so does the connectivity needed to link clusters of accelerators. The industry is in the middle of a transition from copper interconnects to optical solutions, and the company that masters low-latency, high-bandwidth data movement holds a structural advantage. This is behind NVIDIA’s aggressive investment in the space.
Qualcomm’s recent Alphawave acquisition placed it directly in this fight with electrical and optical SerDes technology, die-to-die interconnects, and co-packaged optics that the company has been developing for more than five years.
Qualcomm is already in production with 800-gigabit electrical and optical DSPs and has a roadmap to 448-gigabit connectivity and next-generation co-packaged optics that bring photons directly into the compute package.
Its connectivity portfolio is anchored with a lead hyperscaler win and generates revenue today, making it the one Dragonfly product line already contributing to the top line.
Qualcomm’s data center ambition is no longer aspirational. The company guided to five billion dollars in data center revenue in fiscal 2027, with the bulk from custom silicon, scaling to fifteen billion by fiscal 2029. It stated that the wafers and memory needed to reach the fiscal 2027 figure are already secured. That is a meaningful signal in a supply-constrained market and separates Qualcomm from vendors making claims unbacked by committed capacity.
The execution risks center on timing and proof. AI accelerator and CPU revenue arrive in the back half of the forecast window, so early revenue depends heavily on custom silicon and Alphawave connectivity rather than the HBC architecture that defines the long-term thesis.
The larger story is that the agentic AI era rewards exactly the capabilities Qualcomm spent decades accumulating: efficiency, memory-aware architecture, disaggregated design, and connectivity. The incumbents built their dominance in a training-centric world that is giving way to an inference-centric one, and that transition is where insurgents find room.
Whether Qualcomm ultimately reaches its $15 billion revenue target will hinge on execution over the next three years. It must bring a fundamentally new memory architecture to market, integrate Modular into a developer ecosystem capable of competing with entrenched software stacks, and demonstrate that it can translate early hyperscaler design wins into sustained production deployments. These are meaningful challenges, and investors should expect the road to be measured in milestones rather than quarters.
Even so, Qualcomm enters this market with clear competitive advantages. It has assembled a portfolio that spans custom silicon, CPUs, accelerators, connectivity, and an increasingly credible software platform. The data center market is only one prong of Qualcomm’s growth strategy, with the company equally focused on expanding into automotive, industrial, and robotics.
AI’s next phase is defined by inference economics, making Qualcomm’s entry well-timed to leverage its decades of expertise in efficient computing, system integration, and connectivity. These are significant differentiators for the company, making a credible case that the next generation of AI infrastructure may increasingly resemble the engineering problems Qualcomm has been solving for more than thirty years.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。