5 AI Stocks to Own for the Inference Age

While the first phase of the AI megatrend was dominated by large language model (LLM) training, the second phase is all about inference and agentic AI. This is a major shift in the AI landscape, and with it will come new opportunities.

Let's look at five AI stocks that will benefit from the rising use of AI inference.

Nvidia: Combining GPUs and LPUs

The leader of AI infrastructure, Nvidia (NASDAQ: NVDA), has dominated the market for LLM training. However, it is also well-positioned for the shift to inference. Its "acquisition" of Groq brought with it language processing units (LPUs) developed for inference, which it has incorporated into its CUDA software platform.

This allows it to offer full end-to-end server racks designed for inference that combine its graphics processing units (GPUs) with LPUs. Its GPUs, packaged with high bandwidth memory (HBM), handle the prefill phase of inference by understanding a user's prompt, while its LPUs, with on-chip SRAM (static random-access memory), then take over the decode phase and instantly provide the response. It's a unique approach that will help the company hold onto its status as an AI infrastructure leader.

AMD: Riding two powerful trends

While Advanced Micro Devices (NASDAQ: AMD) played second fiddle to Nvidia as a supplier of GPUs for LLM training, the company is much better positioned for inference. Inference workloads are less technologically complex and tend to be more memory-bound. This suits AMD's chiplet design, which allows its GPUs to be packed with more memory. Meanwhile, its ROCm software platform has improved greatly over the past few years.

The company has two huge GPU purchase commitments worth over $100 billion each, and it's rumored that Anthropic is looking to use its next-generation chip. AMD also has a huge opportunity to take advantage of rising demand for central processing units (CPUs). While the ratio of CPUs to GPUs in servers handling AI training was relatively small, CPUs are needed in much higher ratios in servers that handle inference and agentic AI workloads. As such, the company is riding two powerful trends.

Cerebras: Looking to upend the inference market

Cerebras Systems (NASDAQ: CBRS) has developed chips that can handle inference workloads 15 times faster than the average GPU. The catch is that these chips use a lot of SRAM, which is bulky. Moreover, its massive, wafer-sized chips require special cooling and power management; the company sells or rents the necessary equipment as part of its complete end-to-end server rack CS-3 system. However, this all makes the CS-3 system an expensive, premium niche solution.

That said, the company reportedly has more than $20 billion worth of purchase commitments from OpenAI, which could help its inference solution to go mainstream. If it can become more than a niche player, Cerebras has the chance to upend the inference market.

Image source: Getty Images.

Broadcom: A huge custom chip opportunity

As hyperscalers aim to reduce their inference costs, they are increasingly looking to develop their own custom AI chips, and one of the companies they turn to for design help is Broadcom (NASDAQ: AVGO). Broadcom is a leader in ASICs (application-specific integrated circuits), which are designed for highly specific tasks. While such chips lack the flexibility of general-purpose GPUs, they are generally more energy-efficient and cost-efficient for the narrow range of tasks they are made for.

Broadcom helped Alphabet develop its highly regarded Tensor Processing Units (TPUs), and it continues to benefit from the search leader's massive AI infrastructure build-out. Other hyperscalers have also turned to Broadcom for help developing their own chips, while Anthropic has started using Alphabet's TPUs. Broadcom management has forecast that its custom chip sales will soar to more than $100 billion in its fiscal 2027, while Citigroup analysts are projecting its AI revenue will surge to $180 billion in fiscal 2028. That's just massive growth.

Micron: A memory winner

As noted above, inference workload performance tends to be more memory-bound than compute-bound, which is good news for high-bandwidth memory (HBM)makers like Micron (NASDAQ: MU). The company is one of the three big DRAM makers, along with Korean companies SK Hynix and Samsung. GPUs and custom AI ASICs need to be packaged with large amounts of HBM for optimal performance, and the shift toward inference is only driving demand higher.

With demand for HBM soaring well past what the manufacturers are able to supply, DRAM prices have skyrocketed. This has provided Micron with surging revenues and huge margin expansions. The market for digital memory has historically gone through large boom-and-bust cycles, but right now, it is in one of its biggest up-cycles ever. The big DRAM makers have been able to sign longer-term three- to five-year deals for the first time, helping create better visibility for this notoriously cyclical business.

Citigroup is an advertising partner of Motley Fool Money. Geoffrey Seiler has positions in Advanced Micro Devices, Alphabet, and Broadcom. The Motley Fool has positions in and recommends Advanced Micro Devices, Alphabet, Broadcom, Micron Technology, and Nvidia. The Motley Fool has a disclosure policy.