Why AI exposure classification matters

The AI buildout creates semiconductor demand at multiple layers of the supply chain simultaneously, not just at the GPU layer. A company can be a major AI beneficiary without selling a single accelerator. Classifying SOX companies by their type of AI exposure helps investors understand where AI spending is actually flowing, which companies have the most direct versus derivative exposure, and how different exposure types respond to changes in AI investment pace.

This page uses Swoopr's AI exposure taxonomy, which is an analytical framework, not an official Nasdaq classification. The taxonomy classifies companies by the primary channel through which AI spending reaches their revenue. Many companies appear in more than one category because the same company can have both direct AI compute exposure and AI manufacturing infrastructure exposure.

Direct AI compute

Definition: Companies that sell accelerators or custom silicon whose revenue directly tracks AI training and inference workloads at hyperscale cloud providers.

  • NVIDIA (NVDA): Designs the H100, H200 and Blackwell GPU families that dominate AI training and inference. CUDA software ecosystem locks in developer workflows. Revenue tied directly to hyperscaler AI capital spending. Also serves gaming and professional visualization markets.
  • AMD (AMD): Designs MI300X and successor GPUs that compete with NVIDIA for AI training and inference. Growing presence in inference workloads. Also serves gaming, professional visualization, and data-center CPU markets.
  • Broadcom (AVGO): Designs custom AI ASICs for specific hyperscalers (Google TPUs, Meta's MTIA) alongside merchant networking silicon. ASIC revenue is tied to hyperscaler AI infrastructure investment. Broadcom appears in both this category and the networking category.
  • Marvell Technology (MRVL): Designs custom AI ASICs for specific hyperscaler customers alongside networking silicon. Similar to Broadcom in positioning. Marvell also appears in networking.

Characteristics of direct AI compute exposure: highest correlation with hyperscaler AI capex announcements; revenue can be lumpy and sensitive to capex cycle timing; customer concentration is high (a few hyperscalers represent a large fraction of revenue); new model architectures that change compute efficiency per training run can create demand uncertainty.

AI memory

Definition: Companies benefiting from demand for High Bandwidth Memory (HBM) and high-density DRAM used in AI accelerator systems.

  • Micron Technology (MU): Manufactures DRAM, NAND flash, and HBM. HBM is stacked DRAM placed physically close to the GPU die in a chip package to provide bandwidth that conventional DRAM cannot. AI accelerators require HBM to avoid memory bandwidth becoming a bottleneck for training throughput. HBM commands significantly better margins than commodity DRAM and is supply-constrained during periods of high AI demand. Micron's commodity DRAM and NAND revenue has a weaker AI correlation; the HBM share of revenue is the key AI exposure metric.

Characteristics of AI memory exposure: HBM is supply-constrained and margin-enhancing; commodity DRAM and NAND follow traditional memory cycles independent of AI spending; distinguishing HBM revenue from total memory revenue is essential for accurate AI exposure analysis.

AI networking and interconnect

Definition: Companies supplying the switches, routers, SerDes, PCIe/CXL components, optical transceivers, and related connectivity for AI data centers and AI server clusters.

  • Credo Technology (CRDO): Supplies high-speed active electrical cables (AECs) and SerDes semiconductor IP for connecting components within AI server racks at speeds of 100Gbps to 800Gbps. Revenue is directly tied to AI cluster build-out.
  • Astera Labs (ALAB): Makes PCIe/CXL retimers and smart cable modules that connect accelerators, CPUs and memory within AI server racks and between racks. Revenue directly tied to AI data-center architecture.
  • Broadcom (AVGO): Switch silicon (Tomahawk, Trident families) routes traffic between AI server racks and to storage. Broadcom appears in both compute and networking.
  • Marvell Technology (MRVL): Also supplies switch silicon and optical DSPs for data center networking.
  • MACOM Technology Solutions (MTSI): Supplies RF and optical components for data-center interconnects and telecom; growing optical content in AI cluster inter-rack connections.
  • Coherent Corp. (COHR): Supplies optical transceivers, modules and subsystems for inter-rack and inter-facility connections in AI data centers; also serves telecom networks.

Why networking is critical for AI: AI training requires all-to-all communication between every accelerator in a cluster; as cluster sizes grow from thousands to tens of thousands of GPUs, the interconnect bandwidth requirement grows faster than the compute itself; this makes networking silicon increasingly valuable per dollar of AI infrastructure spending.

AI manufacturing infrastructure

Definition: Foundries, lithography, and wafer-fab equipment companies that produce or enable production of the leading-edge chips used in AI systems.

  • TSMC (TSM): The only high-volume manufacturer of the advanced process nodes (3nm, 2nm) used for AI accelerators. TSMC makes NVIDIA's GPUs, AMD's MI-series, Apple's chips, and most other leading-edge designs. Without TSMC's foundry capacity, AI accelerator production would stop. This is the most critical indirect AI beneficiary.
  • ASML (ASML): Sole supplier of EUV lithography systems required to pattern chips at advanced nodes. Each EUV system costs roughly $380 million. Every leading-edge fab requires multiple EUV systems. ASML's AI exposure operates through foundry investment in advanced capacity.
  • Applied Materials (AMAT): Supplies deposition, etch and CMP equipment for all process steps in advanced chip manufacturing. AI chip complexity requires more deposition and etch steps per wafer.
  • Lam Research (LRCX): Specializes in etch and deposition; dielectric etch for logic and memory is central to advanced AI chip production.
  • GlobalFoundries (GFS): Less direct AI exposure than TSMC; GFS focuses on specialty and mature nodes rather than the leading-edge nodes used for AI accelerators. Some indirect exposure through overall fab spending.

Time lag: equipment orders today translate into production capacity 12-24 months later; this means AI chip demand today drives equipment company revenue in the future, creating a leading indicator relationship.

AI process complexity

Definition: Metrology, inspection, advanced materials, and test companies that benefit because more complex AI chips require more process control steps, tighter material purity, and more thorough testing.

  • KLA Corporation (KLAC): Process control and inspection equipment; as AI chip complexity increases, the number of inspection steps per wafer increases; KLA revenue per wafer manufactured rises with node complexity.
  • Nova Ltd. (NVMI): Optical metrology for film thickness and critical dimension measurement; more process steps per wafer drive more measurement requirements.
  • Entegris (ENTG): Supplies specialty chemicals, filtration systems and handling equipment for semiconductor fabs; advanced AI chip manufacturing requires higher purity materials and more contamination control steps; wafer start volumes at leading-edge fabs drive materials consumption.
  • Teradyne (TER): Automated test equipment; AI chips have extremely high transistor counts and complex functionality that require more test time per chip; growing AI chip volumes drive test equipment demand.

AI power

Definition: Analog and power semiconductor companies that benefit from the extraordinary power consumption of AI data centers and from the need for efficient power delivery to accelerator clusters.

  • Monolithic Power Systems (MPWR): Designs power management ICs for AI server power delivery; voltage regulators for GPU clusters are a high-growth product line; strong positioning in the AI data-center power delivery market.
  • Texas Instruments (TXN): Supplies analog and power chips for data-center power supplies and UPS systems; more diversified than MPWR with industrial and automotive exposure; AI power delivery is one of many end markets.
  • Analog Devices (ADI): High-performance data converters and power management for industrial and data-center applications; growing AI infrastructure content.
  • onsemi (ON): Power semiconductors including MOSFETs and silicon carbide devices for power conversion; growing AI data-center power supply content.

Why AI power matters: a large-scale GPU cluster can consume tens of megawatts; efficient power delivery from the facility level down to the chip level requires sophisticated power management semiconductors at every stage; this drives demand for power ICs, power modules, and power conversion components.

Limited direct AI exposure

Companies with more diversified or less direct AI data-center revenue exposure as of September 2026:

  • Qualcomm (QCOM): Strong in on-device AI for smartphones and PC platforms; limited hyperscale data-center exposure; revenue primarily tied to smartphone and automotive markets.
  • Skyworks Solutions (SWKS): RF front-end modules for smartphones; no significant AI data-center exposure; revenue tied to smartphone production volumes.
  • Qorvo (QRVO): RF components for mobile and defense; no significant AI data-center exposure.
  • NXP Semiconductors (NXPI): Automotive processors, industrial MCUs, RF identification; automotive and industrial cycles are primary drivers; very limited direct AI data-center exposure.
  • Microchip Technology (MCHP): Microcontrollers and analog for embedded applications in industrial, automotive and IoT; very limited AI data-center exposure.
  • Arm Holdings (ARM): Architecture IP that underlies virtually all mobile processors and a growing share of data-center CPUs used for AI inference; Arm benefits from AI through royalties on chips that use its architecture, but indirectly; the revenue model is royalty-based, not tied to AI capex directly.
  • Rambus (RMBS): Memory interface IP and security chips; some AI benefit through HBM interface IP; primarily tied to DRAM volume and interface adoption cycles.

These companies are not "AI losers"; their revenue is driven by cycles independent of AI data-center spending. During periods when AI capex is strong but automotive or mobile demand is weak, these companies can underperform the rest of the SOX index even as the headline index rises.

Using AI exposure classification

AI capex spending can accelerate and decelerate independently from the broader semiconductor cycle. In 2023, consumer electronics and automotive semiconductor demand corrected sharply through inventory digestion, while AI data-center spending was accelerating. Analysts tracking SOX as a single number saw conflicting signals; those tracking the AI-exposed and non-AI-exposed subsectors separately could separate the two trends.

The taxonomy also helps with time-horizon analysis: Direct AI Compute and AI Networking companies see revenue move quickly with hyperscaler capex announcements. AI Manufacturing Infrastructure companies see revenue move with a 12-24 month lag as fab investment translates into capacity. AI Process Complexity companies track closely to leading-edge wafer production volumes. AI Power companies see more gradual revenue growth correlated with total AI data center power consumption.

Companies in multiple categories (Broadcom, Marvell, TSMC) have multiple revenue lines with different lag structures and different sensitivities. Analyzing them as single-category companies loses this granularity.

Frequently asked questions

Which SOX companies have the most direct AI exposure?

Using Swoopr's AI exposure taxonomy, the companies with the most direct exposure are in the Direct AI Compute category: NVIDIA (GPU accelerators for training and inference), AMD (GPU challenger for AI workloads), Broadcom (custom AI ASICs for hyperscalers), and Marvell (custom AI ASICs). In the AI Networking category, Credo Technology and Astera Labs have revenue that is almost entirely tied to AI data-center build-out. TSMC has the most critical indirect exposure: it manufactures virtually all leading-edge AI chips, so AI demand flows entirely through TSMC's foundry capacity, but TSMC's revenue also depends on non-AI customers.

What is High Bandwidth Memory (HBM) and why does AI need it?

High Bandwidth Memory is a type of DRAM that stacks multiple memory dies vertically and connects them through silicon vias to create an extremely wide memory interface. Where conventional DRAM uses a 64-bit interface per channel, HBM uses a 1,024-bit interface or wider, delivering far more data per clock cycle to the processor. AI accelerators like NVIDIA's H100 and H200 use HBM because training large neural networks requires moving enormous quantities of data between memory and compute at extremely high bandwidth. Without sufficient memory bandwidth, the GPU's compute units sit idle waiting for data, making training slow regardless of compute throughput. HBM is manufactured by Micron, SK Hynix and Samsung; Micron is the only HBM supplier listed in SOX.

Why is ASML considered an AI beneficiary if it doesn't make AI chips?

ASML makes the EUV (extreme ultraviolet) lithography machines that are required to manufacture chips at advanced process nodes. Every leading-edge AI accelerator, including NVIDIA's Blackwell GPUs, is manufactured at 3nm or 2nm process nodes, which require EUV lithography. ASML is the sole global supplier of EUV systems. When AI demand drives hyperscalers to order more GPUs, TSMC (which manufactures those GPUs) must invest in more EUV capacity, which means ordering more ASML machines. This chain makes ASML an AI infrastructure beneficiary, but with a lag of 12-24 months from when AI chip demand accelerates to when it translates into ASML equipment orders and delivery.

How does Arm Holdings benefit from AI without selling chips?

Arm licenses processor architecture (instruction set architecture, or ISA) and physical IP to chip designers, who incorporate Arm designs into their chips and pay royalties per chip shipped. As AI inference deployments grow, the number of chips using Arm architecture grows: smartphones running on-device AI models use Arm application processors; cloud inference servers increasingly use Arm-based CPUs like AWS Graviton and Ampere Altra; automotive AI systems use Arm cores. Arm's royalty rates per chip tend to rise with chip sophistication, so the value of AI chips incorporating Arm IP is higher than commodity embedded chips. Arm benefits from the AI cycle through rising volumes and rising per-chip royalty rates, not by selling anything directly to AI data centers.

Can semiconductor companies benefit from AI without selling to hyperscalers?

Yes. Several paths exist for AI benefit that do not require direct hyperscaler sales. Equipment companies (ASML, Applied Materials, Lam Research) sell to foundries like TSMC, which then manufacture AI chips for hyperscalers; AI drives fab investment, which drives equipment demand, without any equipment company interacting with a hyperscaler directly. Process control companies (KLA, Nova) sell to fabs whose production complexity increases with AI chip designs. Materials companies (Entegris) sell to fabs whose consumable consumption increases with advanced node production. Test companies (Teradyne) sell to chip manufacturers and OSATs who test finished AI chips. Power management companies (Monolithic Power, onsemi) sell to server OEMs and rack power supply manufacturers who build the infrastructure that AI clusters run in.

References