The first phase of the current artificial intelligence (AI) boom was all about training the foundational large language models (LLMs) needed to help answer queries and perform tasks. Training these AI models required a ton of processing power, which is where Nvidia (NVDA +1.22%) and its graphics processing units (GPUs) shone.
Not only are Nvidia's GPUs powerful, but the company also smartly seeded its CUDA software platform in places where early AI research was being done. This led to most early foundational code being written on its software and optimized for its chips, giving it a huge moat and making it the biggest AI winner.
While AI model training remains important, this AI-focused market is starting to shift toward inference and agentic AI. New winners will emerge as factors such as memory, central processing units (CPUs), and cost per inference become more important. Let's look at three AI stocks that could become the next big AI winners.
Image source: Getty Images
AMD: An agentic AI and inference winner
The shift toward agentic AI and inference plays to Advanced Micro Devices' (AMD +8.04%) strengths. First, both require much higher CPU-to-GPU ratios: training at about 1:8, inference at around 1:4, and agentic AI at 1:1. AMD is the leader in the high-performance data center CPU market, so this is a huge tailwind for the company. On top of that, data center CPU prices should also be on the rise, as agentic AI needs more cores (sort of like individual workplaces with the chips), which should raise prices.
NASDAQ: AMD
Key Data Points
AMD's chiplet GPUs, meanwhile, pack in more memory, which makes them ideal for inference, which tends to be more memory-bound than processing-bound. With large GPU inference deals in place, the company is set to be a winner in inference and agentic AI.
Micron: A HBM leader
As noted above, inference tends to be more memory-bound than processing-bound, which is great news for dynamic random access memory (DRAM) maker Micron Technology (MU +4.60%). The DRAM market is already undersupplied, which is driving up prices, and the market shift to inference should only keep demand booming.
NASDAQ: MU
Key Data Points
In general, GPUs need to be packaged with a special form of DRAM called high-bandwidth memory (HBM) to perform optimally, and inference uses even more of this memory. While static random-access memory is a competing technology, it is physically large and expensive. With AI chip demand continuing to surge, the outlook for HBM and DRAM remains strong. Meanwhile, the big three DRAM makers -- Micron and Korean companies Samsung and SK Hynix -- have started to lock in longer-term (three-to-five-year) deals, marking a major shift for the memory industry.
Against this backdrop, Micron looks like a big inference market winner.
Broadcom: A custom chip giant
Another winner as the market shifts toward inference is Broadcom (AVGO +1.57%). As a leading provider of application-specific integrated circuit technology, the company helps customers develop custom hardwired chips designed for specific tasks. Increasingly, one of these tasks is inference.
NASDAQ: AVGO
Key Data Points
Broadcom helped Alphabet design its highly successful Tensor Processing Unit (TPU), a huge driver in and of itself. Alphabet is spending a boatload on AI infrastructure, including its own chips. It's also allowing some select customers to buy directly from Broadcom. One of these customers is Anthropic, which has placed a $21 billion TPU order for this year, and the three companies agreed to a partnership for more orders in the future.
Meanwhile, given the success of TPUs, other hyperscalers (owners of large data centers) have also been turning to Broadcom to help them develop their own chips, often with the intent to lower inference costs. Broadcom says it has a clear line of sight to its custom chip business alone, hitting over $100 billion in revenue in fiscal 2027. Throw in its fast-growing networking business, which is benefiting from larger AI clusters, and the company is set to be a big winner in the inference economy.
