QCOM (Trans): From chipmaker to platform solutions co
Compiled by Dolphin Research:$Qualcomm(QCOM.US) 2026 Investor Day Trans
I. Key takeaways from the earnings backdrop
1. Shareholder returns: Returned a cumulative 40 bn to shareholders over the past five years, and retired 30% of shares over the past decade. Management plans to keep dividends growing at a low-to-mid single-digit pace and return most FCF to shareholders.
2. Materially higher financial targets: FY29 non-handset revenue target raised to 40 bn from 22 bn set 18 months ago, implying a near-doubling and a 40% FY25–FY29 CAGR. FY29 EPS targeted at >18, with a long-term revenue ambition of 100 bn.
3. Data center outlook: FY27 revenue target of 5 bn (including two hyperscalers each >1 bn for custom silicon), FY29 target of 15 bn. Long term, aim for >5% market share within 5–7 years.Custom silicon GPM will run slightly below corporate Avg., but at the OPM level it is accretive.
4. Margins and opex: Long-term OPM targets of 30% for QCT and 70% for QTL. Opex as % of revenue to decline from 23% currently to 19–20%.
5. Mix shift: Handsets to fall below 50% of revenue by FY27 and to ~one-third by FY29. Android handset revenue expected to grow ~5% annually, with potential upside from memory tailwinds and agentic AI not included.
II. Investor Day details
2.1 Key messages from management
1. Corporate strategy
a. Three pillars for the next five years: enter data centers, become a full-stack physical AI platform at the edge, and evolve from a chip vendor to a platform-solutions company (HW + SW + developer ecosystem).
b. Core advantages: a technology/IP portfolio built with >100 bn in cumulative R&D, cross-industry strategic partnerships, and scaled execution (consuming >1 mn advanced-node wafers annually, 75+ tape-outs per year, and shipping 40 bn components per year).
c. Coverage across the compute continuum from ~2 microwatts to ~200 kilowatts.d. Announced the Modular acquisition, framed by the CEO as a potential 'Android moment' or even 'Linux moment', to build an open AI software platform.e. Strategic tie-up with Hugging Face: model ecosystem for Dragonfly data center silicon, model deployment across Snapdragon/Dragonwing/Dragonfly, and a distributed AI framework.
2. Data center
a. Agentic AI changes compute economics: a single agent query can trigger 50x–100x more inference requests and exceed 1 mn tokens. Legacy compute stacks are insufficient, requiring a paradigm shift.
b. Introduced the 'Dragonfly' data center infra brand with a >$1 tn TAM across four lines:- Connectivity: 1st-gen 800G electrical/optical DSP and Coherent Light in production; 224G 2nd-gen by year-end; 448G 3rd-gen in 2028; first hyperscaler anchor secured.- Custom silicon: Won two major hyperscaler programs within six months of ramping the team, with meaningful revenue starting at year-end (FQ1 2027).- AI accelerators: Mid-2027 launch of AI250, the first HBC near-memory compute accelerator; AI300 in 2028 with integrated silicon photonics and next-gen scale-up networking.- CPU (C1000): Mid-2028 launch, >5 GHz (30%+ faster vs. peers), >250 cores, >2 TB I/O bandwidth, AI-native with optional HBC; three SKUs for agentic CPU, general-purpose CPU, and AI head-node CPU, targeting a 200 bn market.
c. HBC breakthrough: re-architects XPU by disaggregating the AI accelerator and placing the XPU directly under DRAM stacks. Delivers SRAM-like performance with HBM-like density/capacity — for ultra-low-latency workloads, 200x capacity/W vs. SRAM-based designs; for high-throughput, 6x bandwidth/W vs. HBM approaches.No expensive silicon interposer required, enabling multiple HPC stacks in standard packaging.
d. Microsoft CEO Satya Nadella confirmed via video that Azure will deploy HBC, citing 'significant gains in cost and performance'.e. Meta CEO Mark Zuckerberg announced a multi-gen partnership, with Qualcomm supplying CPUs for Meta's next-gen server fleet.f. Humain CEO Tareq Amin appeared via video, confirming Humain as Qualcomm's first data center customer.
3. Modular acquisition and software platform
a. Modular is positioned as a portable alternative to NVIDIA's software stack, designed from day one for every AI accelerator.b. The stack comprises the Mojo language (a high-performance low-level programming model), the MAX inference layer, and Modular Cloud for distributed inference.
c. Delivers up to 50% faster AI inference on third-party hardware. d. Turns heterogeneous data centers into 'multi-chip AI token factories', enabling hardware agnosticism.e. CEO Chris Lattner said the platform will evolve into a full OS that is natively distributed, natively accelerated, and natively agentic, with an open-platform commitment.
4. Auto, Industrial & Robotics
a. Auto:
- FY26 ARR of 6 bn, 23 consecutive quarters of double-digit YoY growth, and a FY29 target of 10 bn.- Design-win pipeline at 65 bn (up from 45 bn two years ago), with content per car up 8x from Gen 3 to Gen 5.- 500 mn+ Snapdragon-powered vehicles on the road, 90 mn cockpits, working with 70+ OEMs and 100+ Tier-1/Tier-2s.
- Gen 5 uses a mixed-criticality architecture to run cockpit and ADAS on a single chip. Commercially runs a 30 bn parameter model in the cockpit while simultaneously running L2–L4 stacks.- ADAS footprint expanded to 25 OEMs; Ride Pilot AD validated in 60 countries; Stellantis chose the full Digital Chassis (SOP '28).- New growth vectors: Robotaxi (HBC Gen 2 accelerator in 2028), in-cabin token acceleration with offline federated inference, vehicle AI/ML with on-device ML ops via EdgeImpulse, and satellite connectivity.
b. Industrial & embedded IoT:
- Indirect revenue to grow 77% from FY24 to FY26, serving 38,000+ customers with 200+ HW solutions, 35+ distributors, and 45 global GSIs.- Dragonwing spans industrial, commercial, and mobile across 12 verticals.- Three acquisitions to strengthen the developer ecosystem: Arduino (33 mn developers), EdgeImpulse (edge ML ops), and Foundries (industrial-grade Linux management).
- Launched Arduino VENTUNO Q in Aug.: 40 TOPS AI, octa-core, 12 cameras, safety island, full upstream Linux support, and agentic development with Claude/Codex/Cursor.- Full-stack video AI from camera silicon to edge AI boxes to video AI services, covering retail (shrink prevention, shelf intelligence), energy & utilities (autonomous drone inspection, auto shut-off), and oil & gas across upstream/midstream/downstream.
c. Robotics:
- In under nine months, secured 100+ partnerships and has shipped across multiple robot form factors.- Introduced a 'three-computer' architecture: System 2 (inference brain/mixed-critical AI workloads), System 1 (actuation/motion planning), and System 0 (reflex/millisecond control/decentralized nervous system), with Qualcomm uniquely architecting across all three domains.
- Dragonwing IQ10/IQ9/IQ8 are in production, with a six-layer full stack spanning compute, next-gen OS, simulation, data pyramid & flywheel, foundation models, and HW reference designs.- Built in-house simulation and robotics FMs supporting multimodal inputs (vision, depth, tactile, natural language) and training via simulation, behavior cloning, teleoperation, and RL.
- First key partner NEURA Robotics provides full reference designs for cognitive arms and humanoids.
- Edge devices & 6G
- a. Paradigm shift: agents become the center instead of phones, and devices become agent endpoints. Each device will serve two 'users' — humans and agents, with China likely at the forefront of agentic use cases.
b. Smart glasses as a conviction category: ~600 mn eyeglasses ship annually, with <1% smart-glasses penetration. Qualcomm has 40+ designs in development; the core use case is 'see what I see, hear what I hear'.
c. Partner ecosystem: Microsoft Project Solara (agent-first devices), Amazon Alexa (home + mobile AI experiences), Google Books Chromebook (Snapdragon + Gemini), and Hark (new personal intelligent devices).
d. Hybrid AI demo: local + cloud routing matches pure-cloud performance with better economics. Annual token demand is projected to rise 40x from 2026 to 2030.
e. Three 6G themes: uplink video at scale turning everyone into a 'walking camera', networks as token-generation and sovereign-AI nodes with carriers selling AI token capacity, and RF sensing that treats every RF as a radar for uses like drone detection.
f. New edge-device architecture needs: an orchestrator-class CPU, new inference architectures, a mobile HBC co-processor, enhanced sensing, and next-gen modems.
6. Financial framework
a. Past five years: revenue doubled to 44 bn and EPS tripled. QCT Auto CAGR was 44%, Android handsets grew 12% CAGR, delivering growth in a 'mature' market.QCT earnings grew 4x, or 2x the pace of revenue.
b. FY29 revenue bridge: 15 bn data center, 10 bn Auto, and 14 bn+ across IoT (personal AI & compute + industrial networks & robotics). Android handsets grow ~5% annually, and licensing remains stable with 4G/5G unit growth.
c. Handsets <50% of mix by FY27 and ~one-third by FY29.d. Completed 35 acquisitions in five years as opex rose just 6% (opex/revenue fell from 31% to 23%).e. TAM expanded to 1.7 tn.f. Post-FY29 growth engines: data center, robotics, industrial upgrade cycles, ADAS & L4 autonomy, new personal-AI categories, and 6G.
2.2 Q&A
Q: How will revenue scale toward the 15 bn data center target, and can the supply chain support it?
A: The ramp from 5 bn in FY27 to 15 bn in FY29 aligns with the product rollout. By FY29, all four lines — CPU, accelerators, custom silicon, and connectivity — contribute.CPU volume ramps from late FY28, so FY28 revenue bridges roughly between 5–10 bn. FY29 is only the starting point of the full portfolio; multi-gen deliveries in FY30–FY31 should accelerate the curve.
On supply, the 5 bn revenue for FY27 already has wafers and memory secured. Qualcomm consumes significant advanced-node capacity, and suppliers are willing to bet on us.The memory commitments needed for HBC are also in place.
Q: What enables the C1000 CPU's >5 GHz performance lead?
A: It is architected from the ground up rather than by simply stuffing in more compute. The Oryon core is transformational, performs exceptionally well at 5 GHz, and is built with automated tools for annual iteration rather than hand-crafted design.While rooted in mobile compute lineage, the server-class design started from scratch for performance leadership. Even with a 2028 launch, it should lead on compute and I/O.
Our CPU journey ran PC vs. Apple's M-series first, then mobile, then safety-grade automotive, with C1000 the next design. We have shared the 2028 CPU spec with hyperscalers, and feedback was 'hard to believe — when can we get silicon?'Combined with native AI inference via HBC, this is a 'game over' combo. Demand for CPUs is massive now, but when head-to-head competition arrives in 2028, we expect to be very competitive.
Q: Where is the software maturity for data center, and what does Modular add?
A: We have long focused on inference, especially disaggregated inference accelerators in data centers, and support standards like Triton. The AI 100 program matured the stack to where new models can run on our accelerators within 24 hours.We continue to embrace open ecosystems.
Modular brings a truly modern platform. CUDA was designed ~20 years ago; Modular was built for heterogeneous, distributed compute and openness from day one.It has shown strong performance on NVIDIA and AMD hardware and on CPUs. Our dual-track strategy is to keep supporting today's standards while building a higher-performance, easier platform with Modular for the industry, spanning data center and edge.
Q: Which customers will deploy Modular, and what are the hurdles?
A: If NVIDIA were the only inference option, only NVIDIA would be shipping — that is not the case. Google has TPU, NVIDIA acquired Groq, and multiple architectures are emerging, with weaker moats in inference vs. training.That creates opportunity: heterogeneous fleets need a developer-abstraction layer.
Some will say 'I'll stick with CUDA on NVIDIA', but many juggle 3–5 different stacks. Leading AI players are coming to Qualcomm saying they need to push workloads to the edge since cloud tokens are better used elsewhere.That inevitably brings diverse hardware. Our vision is an open, scalable, easy platform for customers who manufacture at scale, from edge to cloud, and it will remain open. Open, level systems will win — that is our bet.
Q: What is the FY27 5 bn data center revenue mix and which customers drive it?
A: Custom silicon is the largest piece, with two global hyperscalers each contributing >1 bn. AI accelerators contribute from year-end, and connectivity from the Alphawave acquisition adds some.Humain will also be a key customer. So the three main lines are custom silicon (largest), accelerators (ramping at year-end), and connectivity.
Q: How does customer concentration evolve from FY27 to FY29? Is the 15 bn FY29 target achievable with existing customers?
A: We have high confidence. The FY29 view rests on active, multi-line, multi-gen engagements with a diversified customer base.Discussions have moved from megawatts to gigawatts; with full-fleet deployments, a few GWs can deliver the FY29 target.
Q: With chipset variants potentially doubling to 250–500, how will Qualcomm manage complexity and quality?
A: 'Quality' is in our name. We have the lowest defect rates with mobile customers, and Apple counts Qualcomm among its top-quality suppliers.In auto, we keep setting records from tape-out to SOP.
Snapdragon mobile chips must ramp fast without Intel-style binning — Samsung will not sell 'fast' vs. 'regular' Galaxies. Every die must meet the same spec, reflecting our manufacturing maturity.With exploding DC demand, customers report rework and failure issues elsewhere; we view reliability as a key differentiator. Alphawave accelerated customer traction under Qualcomm due to stronger execution and capacity commitments.Our FY27 5 bn outlook is already backed by wafer and memory reservations.
Q: What is the CPO (co-packaged optics) roadmap in connectivity?
A: Alphawave has invested >5 years in silicon photonics and CPO. We plan to introduce first-gen silicon photonics in the AI300 family, enabling direct optical scale-out — no copper-to-optics conversion, with major power savings.Initial deployment targets 2028 for scale-out, with subsequent expansion to scale-up networking.
<End of text>
Risk disclosure and statement:Dolphin Research Disclaimer and General Disclosure
