Artificial intelligence has turned a handful of semiconductor companies into some of the most closely watched stocks in the world. To follow the story as an investor, you need to understand who buys AI chips, who makes them, and where the bottlenecks are.
Who buys AI chips
Training and running large AI models requires enormous computing power, most of it housed in data centers. The biggest buyers are the hyperscalers, the large cloud providers such as Amazon Web Services, Microsoft Azure, and Google Cloud, along with large internet companies such as Meta that run AI workloads for their own products. Specialized AI cloud providers, AI labs, governments, and large enterprises make up much of the rest.
This concentration matters. When a few customers account for a large share of demand, their spending plans become the single most important input for chip-company revenue. That is why chip stocks often move sharply on a cloud company’s earnings call, even though the chip company itself did not report anything.
The main types of AI chips
- GPUs (graphics processing units): originally designed for graphics, they excel at the parallel math behind AI and have become the workhorse for training models. Nvidia is the dominant supplier, with AMD as the main merchant competitor.
- Custom accelerators (ASICs): chips designed for a specific customer’s workloads, such as Google’s TPUs. Companies like Broadcom and Marvell help hyperscalers design these chips.
- CPUs and networking: AI clusters also need general-purpose processors and very fast networking to connect thousands of accelerators, which creates demand for networking chips, switches, and optical components.
Where the bottlenecks are
Leading-edge manufacturing
Most advanced AI chips are designed by one company and manufactured by a foundry. Taiwan Semiconductor Manufacturing Company (TSMC) produces the majority of leading-edge chips for companies including Nvidia, AMD, and Apple. Capacity at the most advanced process nodes is limited and takes years to expand.
High-bandwidth memory (HBM)
AI accelerators need extremely fast memory placed right next to the processor. HBM is made by a small number of memory producers, principally SK Hynix, Samsung, and Micron. HBM supply has been a key constraint and a major profit driver for those companies.
Advanced packaging
Combining a processor and stacks of HBM into one package requires specialized techniques. Packaging capacity has at times limited how many AI accelerators could be shipped, regardless of how many chips were manufactured.
Numbers that move AI chip stocks
- Hyperscaler capital expenditure guidance: the clearest signal of future demand. See our guide to hyperscaler capex.
- Data center segment revenue for chipmakers, and its growth rate.
- Gross margin: signals pricing power and product mix.
- Forward guidance: AI chip stocks often trade on next quarter’s outlook more than on the quarter just reported.
- Export controls: US restrictions on selling advanced chips to certain countries can remove a portion of addressable demand.
Risks to understand
AI chip demand has grown extremely fast, and the semiconductor industry has a long history of boom-and-bust cycles. Key risks include customers pausing to digest capacity they already bought, a shift from merchant GPUs toward custom chips, pricing pressure as competition grows, and questions about whether AI applications will generate enough revenue to justify the spending. Our explainer on the semiconductor cycle covers how past booms have unwound.
