One area where Intel (INTC -3.67%) has seriously dropped the ball has been artificial intelligence. Nvidia essentially owns the market for AI accelerator chips, and AMD is at least competitive with its Instinct line of data center GPUs. Nvidia is generating tens of billions of dollars in revenue each quarter from its data center GPUs, and AMD is producing a few billion per year. Meanwhile, Intel fell short of its goal to sell $500 million worth of AI accelerators in 2024.
Missing the boat
It didn’t have to be this way, but mistakes made long ago have hampered Intel’s AI efforts. The company was working on a discrete GPU project codenamed Larabee way back in 2008, and its architecture would have been well suited for the massively parallel computing tasks required to train AI models. Larabee was canceled, however, setting the stage for Intel to be blindsided by the AI revolution and the surge in demand for GPU accelerators over the past few years.
Intel’s current AI accelerator lineup comes from Habana Labs, an AI chip company Intel acquired in 2019. The Gaudi family of AI chips aren’t traditional GPUs, although they’re similar in nature. The latest Gaudi 3 chip offers solid performance, and Intel has been aggressive on pricing. Gaudi could have been a success story for Intel, but an immature software ecosystem is hurting sales. Developers who have been working with GPUs for years don’t have experience with Gaudi’s distinct architecture, something that’s been impossible for Intel to overcome.
Software immaturity is starting to become a theme for Intel in the graphics and AI space. The company tried again in the discrete GPU market in 2023 with its Arc Alchemist gaming graphics cards, but they failed to sell well due to faulty software drivers and myriad game-breaking bugs. The company has stuck with it, greatly improving its software over time, and its second-generation graphics cards are faring much better.
Don’t expect big things in 2025
If Intel’s experience in the graphics card business is any indication, it’s going to take time for the AI software ecosystem to improve enough for Gaudi to be a major success. Another wrinkle is that Intel’s AI accelerator roadmap is complicated. Intel actually already sells data center GPUs under the Max brand name, which power the Aurora supercomputer at Argonne National Laboratory. Neither the Max family of GPUs nor the Gaudi family of AI accelerators will be extended. Rather, the next-generation Falcon Shores will be a traditional GPU that reportedly integrates some of Gaudi’s unique features.
Falcon Shores is expected to launch by the end of 2025, although given Intel’s current turmoil, there are no guarantees. Gaudi 3 sales should ramp up this year compared to last year, but it’s unlikely sales are going to be anywhere close to what AMD is raking in with its data center GPUs given the software issues.
Intel has had some successes, although they’ve been few and far between. One major win was a deal with IBM to put Gaudi 3 chips in IBM’s cloud datacenters and integrate them into IBM’s watsonx AI platform. That should help boost Gaudi sales this year, although the size of the deal hasn’t been disclosed.
A bigger AI opportunity
Intel may eventually find success selling data center GPUs that double as AI accelerators, but it’s probably not going to happen this year. In the long run, manufacturing AI accelerators for other companies may prove to be a more lucrative endeavor.
Intel is betting big on leveraging its decades of manufacturing expertise and tens of billions of dollars in manufacturing investments to build a viable foundry business. The company is on the cusp of completing its initial process roadmap, with the Intel 18A process scheduled to go into volume production next year. Intel will make some of its own CPUs on Intel 18A, and it’s already won a few high-profile customers.
Those customers include Microsoft, which will use Intel 18A to produce an undisclosed chip, and Amazon, which will tap Intel 18A for an AI fabric chip. If Intel can prove to potential customers that Intel 18A is the real deal by successfully making its own chips and chips as part of these early partnerships, more AI-related chips could be running through Intel’s foundries in 2026 and beyond.
Long story short, Intel’s AI accelerator revenue probably won’t be all that impressive in 2025 as the company works through software issues and solidifies its roadmap. However, looking further ahead, the foundry business may ultimately be a larger opportunity for Intel to scoop up some AI spending.
