Qualcomm sold roughly $0.3 billion of data center silicon in fiscal 2026. By 2029, it plans to sell $15 billion. That gap, fifty times the current base, is the whole story of its June Investor Day and the reason the market moved.
It bought Modular, the AI-native software platform, for $3.9 billion in stock, and now claims a silicon-agnostic software layer that attacks NVIDIA's CUDA lock-in
Two hyperscaler deals are set to generate at least $1 billion in revenue within a year, with shipments starting by the end of 2026
This is a company that left the server market in 2018 and is now telling hyperscalers it can be a second supply chain for AI compute. The details matter, because the gap between the pitch and the silicon is where the real risk sits.
The fifty-times gap
Qualcomm's data center revenue target of more than $15 billion by fiscal 2029 is not an extrapolation. It is a jump from a near-zero base. CFO Akash Palkhiwala said the two hyperscaler customers already signed are expected to produce at least $1 billion in combined revenue within a year, with initial shipments beginning by the end of 2026.
For context, the rest of the company's diversification plan: $10 billion from automotive, more than $14 billion from IoT, $8 billion from industrial, networking and robotics, and $6 billion from personal AI and compute. Handsets drop to roughly one-third of chip revenue. The non-handset target was doubled to $40 billion.
Data center revenue ambition
From roughly $0.3 billion in fiscal 2026 to a target above $15 billion in fiscal 2029. · Investor Day, 2026
The software purchase
It agreed to buy Modular, the AI infrastructure software firm, for about $3.9 billion in stock, completed in late July. · Company announcement, 2026
Revenue already locked in
Its CFO said the two signed hyperscaler customers should generate at least $1 billion within a year. · Data Center Knowledge, 2026
The numbers are ambitious, but they are not the only read of the event. Its argument is that agentic AI changes what the data center actually needs, and that the biggest silicon supplier is architected for the wrong workload.
The deal that makes it real
Meta is one of the two hyperscaler customers. The second is unnamed. Meta committed to a multi-generation collaboration under which Qualcomm will supply data center CPUs for its next-generation server fleet, with the first-generation Dragonfly C1000 going into production in the second half of 2028.
Mark Zuckerberg appeared at the event to frame it in Meta's terms. "Our goal is to deliver personal superintelligence to everyone in the world," he said. "That's why our work with Qualcomm is so critical."
Our goal is to deliver personal superintelligence to everyone in the world. That's why our work with Qualcomm is so critical.— Mark Zuckerberg, Founder and CEO, Meta
Meta's own statement was more measured. A spokesperson told Data Center Knowledge the company is "embracing a flexible, portfolio-based approach, combining hardware from a range of partners with our own rapidly advancing MTIA silicon program." In other words, it is one component in a portfolio, not the replacement for Meta's in-house effort.
That caveat matters. The deal gives it validation from a hyperscale operator and revenue to fund development. It does not make it a default choice. Moor Insights analyst Matt Kimball put it directly: "One customer win doesn't change the server CPU market overnight. But it definitely expands the conversation."
Three chips, one memory bet
The product line is called Dragonfly. The CPU is the C1000, built on Qualcomm's Oryon architecture with more than 250 cores, sustained frequencies above 5 GHz, PCIe Gen 7 and CXL support. It claims about twice the performance-per-watt of existing server CPUs, aimed at the sequential reasoning and context-switching that agentic workloads push onto the CPU rather than the GPU.
The inference side is the AI300 accelerator, a third-generation rack-level design supporting air and direct liquid cooling. It carries the second generation of High-Bandwidth Compute (HBC), a near-memory architecture using 3D-stacked silicon to attack the data-movement bottleneck.
HBC is where the technical bet lives. The company describes it as combining SRAM-class performance with HBM-class capacity, and its own design targets are a 6x improvement in bandwidth per watt versus HBM at the card level, and a 200x improvement in capacity per watt versus SRAM at the rack level. These are company design targets, not measured third-party benchmarks. An intermediate product, the AI250 with first-generation HBC, reaches 133 Tbps per card and samples commercially in mid-2027. The AI300 and second-generation HBC follow in 2028.
Why memory became the bottleneck
Buying the software layer
The most important move may be the one that was not silicon. Qualcomm agreed to acquire Modular, the AI-native software platform, for about $3.9 billion in stock. Modular's pitch is a silicon-agnostic compute layer: write once, deploy across CPUs, GPUs, and accelerators from different vendors. The acquisition completed on July 29.
It frames this as building bridges instead of moats. It is also, plainly, a response to CUDA. NVIDIA's software stack is the reason developers stay locked to NVIDIA silicon, and any serious challenger needs an answer at that layer. Whether Modular's technology, which runs CUDA-derived code on other hardware, is that answer is unproven at hyperscale.
The portfolio also includes the Alphawave Semi acquisition from 2025, which brought in the interconnect and chiplet IP that ties the platform together, and a custom silicon engagement with an unnamed hyperscaler, with initial shipments scheduled for December 2026. Tony Pialis, who led Alphawave, now runs its data center business.
The second supply chain
Qualcomm is not entering a vacant market. NVIDIA dominates, AMD is shipping its own rack-scale platforms, and every major hyperscaler is designing custom silicon in parallel. Amazon, Google and Meta are all building their own accelerators. The question is whether there is room for a merchant supplier of CPUs and accelerators that is not NVIDIA.
That is exactly the opening it is targeting. Hyperscalers do not want a single point of supply for the most expensive part of the data center. A credible second source, with a software layer that makes switching cheap, changes procurement leverage even if no switch ever happens.
| Parameter | Dragonfly platform | NVIDIA rack-scale |
|---|---|---|
| Primary pitch | ✔ performance-per-watt, inference + CPU | ✗ raw training throughput |
| Software layer | ◐ Modular (silicon-agnostic, unproven at scale) | ✔ CUDA, mature ecosystem |
| First production | ✗ C1000 in H2 2028 | ✔ shipping now |
| Memory bet | ◐ HBC near-memory, design targets only | ✗ HBM supply-constrained |
The scar from this beat is fresh. As we wrote in August, when OpenAI shipped its Jalapeño inference silicon, the ASIC-vs-GPU cost-per-token war is the actual competition in AI infrastructure right now. It is a different kind of challenger: a merchant supplier, not a captive program, offering hyperscalers an off-the-shelf second source.
Can Qualcomm close a fifty-times gap by 2029?
Probability: 55% — the $1 billion first-year base from two customers is already signed, and a credible second silicon supply chain is structurally attractive to hyperscalers
✅ Arguments for
The custom silicon engagement with a second hyperscaler ships as soon as December 2026, an early proof point
Modular gives it an answer to CUDA that AMD and Intel lack
Confirmation criteria: the second hyperscaler is named in 2027, or a third customer is announced
❌ Arguments against
HBC performance figures are company design targets, not measured benchmarks
NVIDIA's ecosystem and roadmap, including its own rack-scale platforms and software, remain the default
Disconfirmation criteria: no named second customer by end of 2027, or a first shipment slip into 2027
Signals to track
Whether the first custom-silicon shipment for the unnamed hyperscaler lands in December 2026
The AI250 commercial sampling milestone in mid-2027 and any third-party HBC benchmark
Whether the second hyperscaler customer is identified and which of the majors it is
How fast Modular's software is adopted on rival silicon, which would validate the silicon-agnostic thesis
Development scenarios
🟢 Optimistic scenario (25%)
Implications: it becomes the credible second source, and data center silicon splits into NVIDIA plus a diversified merchant tier
🟡 Base-case scenario (50%)
Implications: data center becomes a real segment but not a structural threat to NVIDIA; Meta remains the marquee CPU customer
🔴 Pessimistic scenario (25%)
Implications: handset business carries the bet, and the $3.9 billion Modular purchase looks expensive
Qualcomm has done the hard part once already: it walked away from servers in 2018 and built a mobile monopoly instead. The question now is whether it can do the return trip faster than the market expects, and whether the second supply chain it is building turns out to be real. The first data points arrive by the end of this year.