AI Server Stocks

Designs TPU accelerators for Google services and Google Cloud. Alphabet captures the value through cheaper internal compute and paid cloud capacity, rather than selling the chip as a standalone product.

Designs Trainium accelerators for AWS. More Trainium adoption lowers Amazon's dependence on merchant GPUs and turns custom silicon into cloud-compute revenue.

Sells MI accelerators for AI compute and EPYC CPUs, including Turin and Venice, that coordinate the rack. It is one of the few vendors with direct exposure to both the accelerator and CPU budgets.

Sells the high-speed Ethernet switches that turn individual AI servers into a cluster. Larger clusters require more switch ports and more bandwidth between racks.

Licenses the architecture and CPU designs used in merchant and custom server processors. Arm earns license fees and royalties as Arm-based CPUs take a larger share of the AI rack.

Sells PCIe and CXL retimers that preserve signal quality between CPUs, accelerators, memory, and network cards. Faster links and denser racks make this connectivity layer more important.

Designs custom accelerators for hyperscalers and supplies Ethernet switch silicon for the cluster fabric. It benefits when spending shifts toward either custom compute or high-speed networking.

Supplies lasers and optical components used in transceivers between racks. As clusters grow, more traffic must move over optical links rather than short copper connections.

Supplies active electrical cables for short, high-speed copper links inside the rack. Its signal-processing chips extend copper reach as data rates rise.

Sells storage systems that hold training data and model checkpoints beside the AI cluster. This is external storage infrastructure, separate from the SSDs installed inside each server.

Sells all-flash arrays built to feed data to compute-intensive workloads. The company was known as Pure Storage before its 2026 name and ticker change.

Sells storage systems used alongside AI servers, especially in enterprise and private-cloud deployments. More AI workloads increase the need to store and move large datasets.

Sells Xeon CPUs that run the host, control, and general-purpose work around the accelerators. Even GPU-heavy servers still need CPUs to coordinate data and workloads.

Makes lasers and optical components for high-speed transceivers. Demand rises as larger clusters require more optical links between servers and racks.

Designs custom accelerator silicon, SmartNIC processors, and optical DSPs. That gives Marvell exposure to both the compute chip and the network path around it.

Designs MTIA accelerators for recommendation and inference workloads inside Meta's data centers. The benefit appears through lower infrastructure costs, not external chip sales.

Sells HBM placed next to the accelerator and DDR used by the rest of the server. AI systems require much more memory content, with HBM carrying the highest value per bit.

Designs Maia accelerators for Azure AI workloads. Like other hyperscalers, Microsoft captures the value through cloud economics and reduced dependence on external chips.

Supplies the power-management chips that convert and regulate voltage on the server board. Higher chip power and rack density increase the complexity and value of this layer.

Sells shared storage systems and data-management software used to feed AI clusters. Its role is keeping training data available across on-premise and cloud environments.

Sells the merchant GPUs that dominate AI compute, plus Grace and Vera CPUs, Spectrum switches, and ConnectX adapters. NVIDIA captures several layers of the rack as one integrated platform.

Targets the server CPU market with an Arm-based design developed from Nuvia technology. Its pitch is high performance with better power efficiency in CPU-heavy AI workloads.

Supplies HBM for accelerators, DDR for server boards, and NAND for SSDs. This gives Samsung exposure across the memory content of an AI server.

Supplies the NAND flash used in SSDs inside the server. Larger models and datasets increase the amount of fast local storage required per system.

Leads the HBM market and also supplies DDR for the server board. Its HBM position gives it one of the most direct memory exposures to accelerator growth.