Nvidia vs. AMD vs. Cerebras: Which Is the Best AI Inference Stock to Buy Today?

While the first phase of the artificial intelligence (AI) trend was all about large language model (LLM) training, the next phase is increasingly becoming about inference -- putting those trained models to work on real-world tasks. This is expected to become the larger of the two markets eventually.

Chipmakers Nvidia (NASDAQ: NVDA), Advanced Micro Devices (NASDAQ: AMD), and Cerebras Systems (NASDAQ: CBRS) are all looking to become the leader in powering inference workloads. Excelling in those computing tasks tends to be more about quick access to memory than raw computing power, and all three are taking different approaches to tackle this issue. As such, let's dig into which looks like the best inference semiconductor stock to buy right now.

Nvidia

Nvidia has long been the AI infrastructure leader: Its graphics processing units (GPUs) have been the main chips used to train AI models. The company has developed a huge moat in this arena by popularizing its CUDA software platform with developers. Most foundational AI code was written in CUDA and optimized specifically for Nvidia's chips.

Inference is a whole other ballgame, though, and the company "acquired" Groq and its language processing units (LPUs) and incorporated them into its CUDA ecosystem to beef up its inference offering. LPUs use a small amount of on-chip SRAM (static random-access memory) to increase inference speeds. Nvidia has essentially created full server racks designed specifically for inference that use a combination of LPUs and GPUs. GPUs packaged with high-bandwidth memory (HBM) take care of the prefill phase of understanding a user's prompt, while LPUs handle the decode phase of giving a response. Since LPUs use SRAM, they can reply with almost zero lag.

Cerebras

Like Nvidia, Cerebras is also using on-chip SRAM to help tackle inference workloads and improve speeds. However, it is doing it in a very different way. Because SRAM is physically bulky, each Nvidia LPU uses a small amount; to handle these workloads, many LPUs need to be interconnected in a huge cluster. Cerebras, on the other hand, has created giant wafer-sized chips about the size of dinner plates that put many standard chips' worth of hardware onto a single slab of silicon. The result is a chip that is 6 times faster than Nvidia's LPUs and 15 times faster than its GPUs.

However, wafer-scale chips come with their own issues. To avoid the impact of costly defects, Cerebras adds extra cores its chips so they can work around any defective areas. Its chips also need special cooling and power management, and as such, Cerebras only sells or rents them out as part of its complete end-to-end server rack CS-3 system. This ultimately makes its offering an expensive premium solution.

That said, the company won a $20 billion deal with OpenAI and also has inked a deal with Amazon Web Services. This could help bring the company's systems into the mainstream.

AMD

Another company looking to gain ground in the inference market is AMD. While it doesn't use on-chip SRAM like Nvidia and Cerebras, its chiplet design allows more memory to be packaged with its GPUs. Meanwhile, its recent acquisition of memory optimization software company MEXT gives AMD a cost-effective way to increase effective memory capacity and reduce costs.

Due to the ongoing AI infrastructure build-out, high bandwidth memory (HBM) is in short supply, and DRAM prices have been surging, increasing the costs of building data centers. MEXT's systems reduce the need for HBM by automatically offloading seldom-accessed data to much cheaper flash storage. MEXT's edge is its predictive AI engine, which analyzes memory access patterns in real time. This allows it to anticipate what data an application will need next and pull it back into HBM right before it is requested. With the integration of MEXT's offerings into its portfolio, AMD can now offer customers full servers dedicated to inference that can save them money.

In addition to its opportunity in the inference space, AMD is also set to benefit from the expected rapid growth of agentic AI, as it is a leading maker of central processing units (CPUs) for AI data centers. Because agentic AI workloads require far more CPUs than other AI processing tasks, the ratio of GPUs to CPUs in data centers -- which was recently about 8 to 1 -- is expected to shrink to around 1 to 1 due to the rise of agentic AI.

Image source: Getty Images.

The verdict

Nvidia is a great stock trading at an attractive valuation, and Cerebras has an opportunity to turn the inference market on its head with its unique wafer-scale engines. However, AMD is the stock I like the most here, as the company is riding two big waves with inference and agentic AI. Its solution is also simpler, and it looks poised to make some nice inroads in this market. It already has some big inference deals in place with OpenAI and Meta Platforms, and there are rumors that Anthropic is about to become a big customer.

With explosive revenue growth potential ahead, AMD could be the biggest winner of the AI inference era.

Geoffrey Seiler has positions in Advanced Micro Devices, Amazon, and Meta Platforms. The Motley Fool has positions in and recommends Advanced Micro Devices, Amazon, Meta Platforms, and Nvidia. The Motley Fool has a disclosure policy.