Nvidia NVDA
Nvidia designs the GPUs that train and serve virtually every frontier AI model. Its H100 and Blackwell platforms set the...
AI chip designers.
The GPUs and custom accelerators that train and serve every frontier AI model. Chip designers sit at the center of the AI hardware stack.
AMD AMD
AMD is the main alternative AI accelerator supplier as labs and hyperscalers look to diversify away from a single chip v...
Broadcom AVGO
Broadcom designs custom AI accelerators for Google (TPUs) and other hyperscalers, alongside high-speed networking silico...
Marvell Technology MRVL
Marvell designs custom AI accelerators and data-center networking chips used across cloud and AI platforms. Like Broadco...
Intel INTC
Intel is pushing its Gaudi AI accelerators and Intel Foundry Services as it works to recapture relevance in the AI chip ...
Qualcomm QCOM
Qualcomm designs the chips inside most Android smartphones and increasingly targets on-device AI inference. Its Snapdrag...
What this part of the stack does.
Every AI model is trained and run on accelerators: chips built to do enormous numbers of calculations in parallel. There are two kinds. General-purpose GPUs, sold by Nvidia and AMD to anyone, and custom chips (ASICs) designed for a single customer's workloads, such as Google's TPUs or OpenAI's in-house accelerators. Broadcom and Marvell earn their AI revenue largely by co-designing those custom chips and supplying the networking that links thousands of them together.
None of these companies makes its own leading-edge chips. They design them, and TSMC manufactures and packages them, which is why the foundry and memory categories move with this one.
Who leads, and how the others compete.
Nvidia dominates. Its advantage is as much software (the CUDA platform developers have built on for almost two decades) and complete systems (chips, networking and racks sold together) as raw chip speed. It books more than half of TSMC's advanced packaging capacity, the step that limits how many AI chips can be made.
AMD is the main GPU alternative and has won the two largest labs as customers: OpenAI agreed in October 2025 to deploy 6 gigawatts of AMD chips, taking a warrant for up to about 10% of AMD, and Anthropic agreed in July 2026 to buy up to 2 gigawatts, with AMD investing up to $5 billion in it. Broadcom's custom-chip business grew with Google's TPUs, which Anthropic also rents at multi-gigawatt scale, and with OpenAI's own accelerator. Qualcomm, best known for phone chips, announced data-center inference chips in October 2025. Intel has struggled to win AI accelerator business and is focused on rebuilding its manufacturing.
What moves these stocks.
Above all, how much the largest buyers spend. Amazon, Microsoft, Alphabet and Meta together planned roughly $725 billion of capital spending for 2026, most of it on AI data centers, and the AI labs have signed multi-year compute contracts worth hundreds of billions more. Earnings calls from those buyers, and new lab deals, often move chip stocks more than the chipmakers' own results.
Supply matters too. Advanced packaging and high-bandwidth memory have both been bottlenecks, and US export rules decide which chips can be sold to China.
Risks to weigh.
A handful of customers account for most AI chip demand, so a pause by one hyperscaler or lab shows up quickly. Custom chips are designed to take share from general-purpose GPUs. Some deals swap equity for orders, which critics call circular. And chip spending has always been cyclical: periods of shortage have been followed by gluts.