SOX Index Subsectors: Breaking Down the Semiconductor Stack

What are the SOX index subsectors?

The 30 companies in the PHLX Semiconductor Sector Index (SOX) span the full semiconductor value chain, from chip design and IP licensing through foundry manufacturing, wafer-fab equipment, process control, memory, analog, connectivity and test. Swoopr groups these companies into 11 analytically distinct subsectors based on primary business model and revenue driver: Compute and AI Accelerators, Custom Silicon and Networking, Memory, Foundries, Manufacturing Equipment, Process Control and Metrology, Analog/Power/Embedded, Connectivity/RF/Optical, Semiconductor IP, and Test. These groupings reflect economic function, not ICB classification. Understanding which subsector a company occupies is the first step in separating the index's overlapping demand cycles.

Why subsectors matter

SOX is not a single-category index. Its 30 members span the full semiconductor value chain, and the index simultaneously tracks multiple independent demand cycles. In any given quarter, AI accelerator demand can be accelerating while automotive semiconductor demand is correcting, equipment order books can be expanding while memory pricing is declining, and connectivity chip revenue can be outperforming while mobile RF volumes are flat. Understanding which subsector a company belongs to is the first step in separating these independent signals from the index's aggregate price movement. An investor who reads SOX as a single number loses information that the subsector breakdown preserves.

The subsector lens also clarifies timing. Equipment orders lead actual chip production capacity by 12 to 18 months. Memory pricing moves on supply-demand balance, which is driven by industry-wide capacity decisions made one to two years earlier. Analog revenue tracks automotive production and industrial capital spending, which follow their own multi-year cycles. These independent cadences mean that two SOX components can be moving in opposite directions for entirely rational reasons, and only the subsector context explains why.

ICB classification foundation

Nasdaq's SOX eligibility covers two ICB subsectors: Semiconductors and Production Technology Equipment. All 30 companies qualify under one of these two classifications. Swoopr's subsector breakdown below further subdivides these two ICB categories into business-model-distinct groups to show the meaningful economic differences within the index. The Swoopr subsectors are an analytical convenience, not an official Nasdaq classification. They reflect primary business models and revenue drivers, not ICB codes.

The ICB Semiconductors subsector encompasses chip designers, foundries, memory manufacturers, analog suppliers and IP licensors. The ICB Production Technology Equipment subsector covers wafer-fab equipment makers, process control and metrology tool suppliers, and test equipment companies. SOX's dual-ICB scope is what makes it a full-supply-chain index rather than a pure chip-design index. A Swoopr subsector grouping may draw from both ICB categories (for example, process control tools sit under Production Technology Equipment even though their revenue is tightly correlated with foundry activity classified under Semiconductors).

Compute and AI accelerators (NVDA, AMD, INTC)

NVIDIA, AMD and Intel are the SOX companies whose primary revenue comes from designing high-performance processors for data-center and AI compute workloads. NVIDIA designs the GPUs and AI accelerators that dominate training workloads for large language models and other AI systems. AMD designs competing GPU and CPU architectures for data-center and gaming markets. Intel is an integrated device manufacturer that designs CPUs and accelerators while also operating its own fabrication facilities.

AI changed this subsector fundamentally. GPU architecture emerged as the dominant hardware for training large language models and running inference because parallel matrix operations required by neural network computation map naturally onto GPU architecture. NVIDIA built the CUDA software ecosystem that made its GPUs the default platform for AI researchers years before the current AI boom, creating a software moat that compounds the hardware advantage.

AMD and NVIDIA are fabless: they design chips but outsource manufacturing to TSMC. Intel is an IDM that both designs and manufactures chips, though its manufacturing position has weakened relative to TSMC at leading nodes. This difference in business model means Intel carries the capital intensity of a foundry in addition to the R&D intensity of a chip designer, which affects its margin structure relative to AMD and NVIDIA.

Revenue in this subsector is directly tied to hyperscaler AI capital spending and enterprise GPU cluster build-outs. Intel's traditional PC and server CPU markets provide an additional but slower-moving revenue stream. The subsector can experience sharp revenue swings when hyperscaler build-out pace changes, because customers can accelerate or defer data-center equipment purchases on shorter notice than most manufacturing capital expenditures.

Custom silicon and networking (AVGO, MRVL)

Broadcom and Marvell design custom application-specific integrated circuits (ASICs) for specific hyperscaler workloads, alongside merchant silicon for network switches and routers. This subsector occupies a distinct position in the AI supply chain: it sits between the general-purpose accelerator layer (NVIDIA, AMD) and the connectivity layer (Credo, Astera Labs), supplying the custom chips that large cloud providers use for AI inference, storage networking and interconnect at scale.

Custom ASIC economics differ substantially from general-purpose chip economics. Broadcom and Marvell work with specific hyperscalers to design chips optimized for their AI inference, storage and networking workloads. This creates long-term partnerships with more predictable revenue than merchant chips sold on the open market. Once a hyperscaler has co-designed and deployed a custom chip, switching to a different supplier requires a new multi-year design cycle, creating meaningful revenue visibility for Broadcom and Marvell.

Networking importance grows as AI cluster sizes scale from hundreds to tens of thousands of accelerators. The bandwidth required to move activations and gradients between accelerators grows faster than the compute itself, making switching and routing silicon increasingly valuable in AI infrastructure. Broadcom's merchant switching silicon (used in open-networking deployments) and its custom networking ASICs both benefit from this dynamic.

Custom ASIC revenue requires deep technical relationships with large customers and cannot be sold on the open market the way a merchant chip can. This limits the addressable customer base but creates stickier, higher-margin relationships with the hyperscalers that are currently driving the largest semiconductor design programs in the industry.

Memory (MU)

Micron Technology manufactures DRAM (dynamic random-access memory), NAND flash storage and High Bandwidth Memory (HBM) stacked close to AI accelerators. Memory is the only SOX subsector where a single company (Micron) represents the entire category by virtue of being the only large-cap U.S.-listed memory manufacturer.

DRAM and NAND pricing is highly cyclical, driven by the balance between manufacturing capacity additions and demand. Pricing can move 50% or more across a cycle. Micron's profitability correlates with industry supply-demand balance, not just with its own production volume. When the industry has excess capacity (as it did in 2022-2023), average selling prices fall regardless of how efficiently Micron runs its fabs. When demand outstrips supply (as HBM demand has recently done), prices and margins recover sharply.

HBM is an AI-specific product that stacks multiple DRAM dies vertically to achieve the extreme memory bandwidth that AI accelerators require. A single NVIDIA H100 GPU uses HBM to deliver roughly 3.35 TB/s of memory bandwidth, far beyond what standard DRAM could provide. HBM commands better margins than commodity DRAM and represents an important shift in Micron's product mix toward differentiated, AI-driven demand.

The reason only one U.S.-listed memory company appears in SOX is structural: Samsung Electronics and SK Hynix, the other two large DRAM manufacturers, are both listed on the Korea Stock Exchange. Neither has a qualifying U.S. listing that meets SOX's Nasdaq Eligible Exchange requirement. Micron is therefore the sole proxy for the global DRAM industry available to SOX.

Foundries (TSM, GFS)

TSMC and GlobalFoundries manufacture chips designed by fabless customers. They own and operate the fabs but do not design chips themselves. This pure manufacturing model separates them from IDMs like Intel that do both. The foundry business is the most capital-intensive segment of the semiconductor industry, with a leading-edge fab costing $20 billion or more to build and equip.

TSMC is the world's most advanced semiconductor manufacturer, producing chips for NVIDIA, AMD, Apple, Qualcomm, Broadcom, Marvell and most major fabless companies at 3nm and 2nm nodes. TSMC's customer list functions as a proxy map of who is winning in chip design: only customers with the most advanced designs need TSMC's leading-edge nodes, and those designs command the highest volumes and prices. TSMC's manufacturing advantage at advanced nodes is so significant that the competitive landscape for leading-edge foundry service has no credible peer.

GlobalFoundries focuses on specialty and mature process nodes for automotive, defense, RF and industrial applications. It does not compete at leading-edge nodes and has no ambition to do so. Its revenue is therefore less tied to AI data-center spending and more tied to automotive production volumes, defense electronics procurement and IoT device production. This gives GlobalFoundries a different cyclical character from TSMC within the same foundry subsector.

Foundry economics are driven by utilization rates (what share of fab capacity is running), process-node mix and wafer pricing. Capital expenditure cycles determine long-term capacity. Underutilization creates rapid margin compression because foundry fixed costs are extremely high. Conversely, when utilization is high and leading-edge capacity is constrained, foundries command premium pricing.

Manufacturing equipment (AMAT, LRCX, ASML)

Applied Materials, Lam Research and ASML make the machines that deposit, etch, clean, pattern and inspect materials on silicon wafers during chip manufacturing. These companies sit one step upstream from foundries in the production chain: before TSMC can make a chip, it must buy and install the equipment that processes the wafers.

ASML occupies a uniquely powerful position as the sole global supplier of extreme ultraviolet (EUV) lithography systems. EUV machines are required to print circuit patterns at the most advanced process nodes. No credible alternative EUV supplier exists anywhere in the world. This creates structural pricing power: a foundry that wants to manufacture at 5nm or below must buy from ASML, and ASML sets the price. Each EUV system costs roughly $200 million and takes years to manufacture. ASML's backlog and delivery schedule is one of the most closely watched leading indicators in the semiconductor industry.

Applied Materials supplies deposition systems (CVD, PVD, ALD), chemical mechanical planarization (CMP) equipment and other critical process steps. It is the largest wafer-fab equipment company by revenue and serves nearly every step in the chip manufacturing process. Lam Research specializes in etch and deposition equipment, with particular strength in memory manufacturing and advanced logic nodes.

Equipment order books typically lead actual chip production capacity by 12 to 18 months, because fabs must be ordered, installed and qualified before they can process wafers. Equipment company revenue is therefore a leading indicator of future semiconductor supply capacity. Rising equipment orders in one quarter signal production capacity additions 12 to 18 months later.

Export control exposure is a significant risk for all three companies. U.S. and allied government restrictions on exporting advanced semiconductor equipment to China have directly affected the revenue that ASML, Applied Materials and Lam Research can generate from Chinese customers, which had been a major market for all three.

Process control and metrology (KLAC, NVMI)

KLA Corporation and Nova Ltd. make inspection and measurement tools used during semiconductor manufacturing to detect defects and verify dimensions at each process step. Every advanced chip manufacturing process involves dozens of deposition, etch, implant and anneal steps, and the wafer must be inspected at multiple points along the way to catch defects before they propagate into finished chips that fail final test.

KLA is the leading supplier of process control equipment and holds a dominant position in wafer inspection and patterned wafer inspection. Its tools are deployed at multiple points in a fab's production flow. KLA's revenue growth tends to track the intensity of advanced-node manufacturing: more complex processes require more inspection steps per wafer, which increases KLA's revenue per wafer started even without an increase in total wafer volumes.

Nova Ltd. specializes in optical metrology, measuring film thicknesses, compositions and critical dimensions using light-based techniques. Nova is smaller than KLA but serves similar inspection and measurement markets, with particular strength in thin-film measurement used at deposition and etch steps.

As process nodes shrink, individual defects or small dimensional variations have a larger impact on yield because the tolerances are tighter. More inspection steps per wafer are required at each successive node. This means the process control equipment market grows faster than the total wafer-fab equipment market when leading-edge production intensity increases. AI chip manufacturing drives demand for process control through the same channel: more complex AI chip designs mean more process steps and higher defect detection requirements at each step.

Analog, power and embedded (TXN, ADI, MPWR, ON, NXPI, MCHP)

Texas Instruments, Analog Devices, Monolithic Power Systems, onsemi, NXP Semiconductors and Microchip Technology design and manufacture analog integrated circuits, power management chips and embedded microcontrollers. This is the largest subsector by company count in SOX and serves the broadest range of end markets: automotive, industrial, medical, consumer electronics and factory automation.

Analog chips differ fundamentally from digital chips. They process continuous real-world signals (temperature, pressure, sound, current, voltage) rather than binary digital data. They are designed into specific customer products and are difficult to replace once designed in because changing a power management chip or analog front-end requires re-qualifying the entire product. Switching costs are therefore high. Analog chips have much longer product life cycles than digital chips, sometimes 10 to 20 years, and margins tend to be strong and relatively stable because pricing competition is less aggressive than in commodity digital markets.

Revenue drivers differ substantially from AI-driven subsectors. These companies' revenue tracks automotive production volumes, industrial capital spending, factory automation activity, medical device production and consumer electronics unit volumes. These cycles are partly independent from data-center AI spending and are often out of phase with the AI compute cycle. The analog sector went through a severe inventory correction in 2022 to 2023 as customers who had over-ordered during supply shortages worked down excess inventory. This correction was distinct from and out of phase with the AI compute boom occurring simultaneously.

Within the subsector, company profiles differ. Texas Instruments is the largest analog and embedded company and is deeply focused on industrial and automotive markets with tens of thousands of product lines and a large direct sales force. Analog Devices concentrates on high-performance signal processing for industrial, healthcare and communications infrastructure. Monolithic Power Systems specializes in high-density power management and has meaningful AI data-center exposure through power delivery for GPU clusters. onsemi focuses on power semiconductors for automotive and industrial applications, particularly electric vehicle power conversion. NXP Semiconductors is strongly positioned in automotive and IoT applications. Microchip Technology serves embedded control applications across industrial, automotive and consumer markets.

Connectivity, RF and optical (QCOM, SWKS, QRVO, CRDO, ALAB, MTSI, COHR)

This subsector groups seven companies with distinct connectivity roles: wireless RF for smartphones and base stations, high-speed wired connectivity for AI infrastructure, and optical components for data transmission. The grouping reflects a shared theme of transmitting signals between devices and systems, but the revenue drivers differ substantially across the group.

Qualcomm, Skyworks Solutions and Qorvo supply chips for smartphones, base stations and wireless devices. Qualcomm is the dominant mobile applications processor and modem supplier: its Snapdragon platform is used in most premium Android smartphones and a growing share of Windows laptops. Skyworks and Qorvo supply RF front-end modules that handle signal amplification, filtering and switching for wireless transmission in smartphones. All three companies are primarily tied to the smartphone upgrade cycle and mobile unit volumes. The smartphone RF market has been a mature and competitive market for several years, limiting growth rates relative to the AI-exposed parts of the index.

Credo Technology and Astera Labs are AI connectivity companies that have emerged as direct AI infrastructure beneficiaries below the accelerator layer. Credo makes high-speed active electrical cables and SerDes (serializer/deserializer) chips for connecting components within AI server racks. Astera Labs makes PCIe and CXL retimers and smart cable modules that extend signal integrity in AI data centers. Both companies' revenue is primarily tied to AI data-center capital spending rather than consumer electronics, making them among the most direct AI beneficiaries in SOX outside of NVIDIA and AMD.

MACOM Technology Solutions makes RF and microwave components for defense, telecom and data-center optical applications. Coherent makes optical transceivers and modules for data-center interconnects and telecom networks. As AI clusters grow in size and require longer-distance high-bandwidth connections between racks, buildings and facilities, optical interconnect becomes a critical bottleneck, and both MACOM and Coherent benefit from this demand. The defense segment of MACOM's business adds a revenue stream less correlated with commercial semiconductor cycles.

Semiconductor IP (ARM, RMBS)

Arm Holdings and Rambus license processor architectures and interface specifications rather than manufacturing chips. Customers pay upfront license fees and royalties per chip shipped. This business model generates highly recurring revenue with very high operating leverage: each additional royalty unit requires no incremental manufacturing investment.

Arm Holdings licenses the ARM instruction set architecture (ISA) used in virtually every mobile chip worldwide and increasingly in data-center CPUs, automotive processors and AI inference chips. Arm does not manufacture any chips. Almost every smartphone application processor in the world uses an ARM-based design under a license from Arm Holdings. As the royalty rate per chip rises with the increasing value and capability of chips using ARM architecture, Arm's royalty revenue grows even without an increase in unit volumes. The shift toward AI inference at the edge and in embedded systems is expanding the addressable royalty base for Arm's architecture.

Rambus licenses memory interface IP, particularly for high-speed DRAM interfaces including DDR5 and HBM interfaces, and sells security chips and cryptography solutions. Rambus occupies a specific niche in memory interface standards: chip designers who want their memory controllers to work optimally with the latest DRAM standards often license Rambus IP to implement those interfaces. Rambus is considerably smaller than Arm but benefits from the increasing importance of high-speed memory interfaces as memory bandwidth requirements grow with AI workloads.

The main risk for IP licensing business models is whether licensees can negotiate lower royalty rates in future contract cycles or develop competing architectures. Arm's scale and the breadth of the software ecosystem built around ARM instruction sets creates substantial barriers to architectural substitution.

Test (TER)

Teradyne makes automated test equipment (ATE) used to verify semiconductor device functionality after manufacturing. Teradyne also operates a collaborative industrial robotics business through its Universal Robots subsidiary, which partially diversifies its revenue away from the pure semiconductor cycle.

Every chip must be tested before shipping. As chips become more complex, the time required to run a complete test suite increases and the test equipment must become more capable to exercise all the functionality in modern designs. An advanced GPU or AI accelerator has tens of billions of transistors and must be tested for logic correctness, memory functionality, power management, thermal behavior and numerous other parameters before it can be shipped to a customer.

AI chip test demand is a specific growth driver for Teradyne. AI accelerators have extremely high transistor counts and complex functionality. Testing one advanced GPU requires sophisticated ATE systems capable of running millions of test patterns across multiple test domains. Increased AI chip volumes drive demand for test equipment both through higher unit volumes and through the longer test time per unit that more complex chips require.

Test equipment spending is semi-cyclical relative to chip production volumes. It is more deferrable than wafer-fab equipment (a fab can delay ordering an additional tester more easily than it can delay ordering a deposition tool needed to run production) but less deferrable than materials because chips must be tested before shipment regardless of the economic environment. Test equipment revenue therefore correlates with chip production with a short lag and with somewhat higher variability than materials revenue.

Subsector economic comparison

Different market conditions affect SOX subsectors in different ways and at different times. The following describes the primary channels through which four major market conditions propagate through the subsector map.

AI capital spending expansion benefits Compute and AI Accelerators (NVIDIA, AMD) most directly and with the shortest lag, because hyperscalers order GPU clusters before building the infrastructure around them. Custom Silicon and Networking (Broadcom, Marvell), AI Connectivity and Optical (Credo, Astera Labs, MACOM, Coherent) and AI Memory (Micron HBM) benefit shortly after, as the data-center infrastructure around the accelerators is built and deployed. Manufacturing Equipment (ASML, Applied Materials, Lam Research) and Process Control (KLA, Nova) benefit with a 12 to 18 month lag as fabs invest in capacity for advanced AI chip production. Analog and Power companies benefit selectively through AI data-center power management demand (Monolithic Power, onsemi, Texas Instruments). Mobile RF companies (Qualcomm, Skyworks, Qorvo) have limited direct AI data-center exposure.

Semiconductor inventory correction primarily affects Analog, Power and Embedded companies and Mobile RF companies. These subsectors supply products into end markets where distributors and end customers build inventory buffers. When demand growth slows, customers work down excess inventory before ordering new chips, creating a demand gap that can last several quarters. AI-specific companies are somewhat insulated from this dynamic during active AI build cycles because hyperscalers tend to pull forward AI chip purchases rather than accumulate distributor inventory in the traditional sense.

Geopolitical export restrictions on semiconductor technology primarily affect Manufacturing Equipment companies. ASML EUV exports to China have been restricted by Dutch export controls coordinated with the United States. Applied Materials and Lam Research face restrictions on exporting advanced etch and deposition tools for leading-edge chip production to China. These restrictions create geographic revenue risk but also redirect chip production to other geographies, which can benefit foundries and equipment companies serving those regions.

Rising interest rates generally compress semiconductor equity valuations because higher discount rates reduce the present value of future earnings streams. The compression tends to be most severe for high-multiple AI-exposed companies (NVIDIA, AMD, Arm Holdings) and IP companies (Arm, Rambus) because their valuations embed the most earnings growth in future periods. Analog and equipment companies with more stable near-term earnings typically see smaller multiple compression in rising rate environments.

Frequently asked questions

References

• Nasdaq: PHLX Semiconductor Sector Index Overview
• Nasdaq: SOX Index Methodology (PDF)