Here's How Intel Can Still Be an AI Winner

Chip giant Intel (INTC -5.27%) has tried and failed to break into the lucrative AI accelerator market that's dominated by Nvidia. The company's Gaudi family of AI accelerators had potential, and Intel priced them aggressively, but sales were snarled by an immature software ecosystem. Intel missed its own AI chip sales estimates for 2024, and it recently shifted gears. Falcon Shores, originally a follow-up to Gaudi 3, is cancelled as a commercial product, and Intel is now focusing on rack-scale AI solutions that likely won't be ready until 2026. While Intel is effectively a non-factor in the AI accelerator market, the company's CPU business could pick up some of the slack. As the AI industry matures and workloads shift from training AI models to running those models, Intel's Xeon server CPUs are a potent weapon in the AI wars. Cheap AI could work in Intel's favor The recent revelation that Chinese start-up DeepSeek managed to train an AI model that compares favorably with the best models from U.S. companies while costing far less to train and run could ultimately work in Intel's favor. While training advanced AI models requires powerful accelerators like those from Nvidia, that's not necessarily the case for inference, or actually running those models. Smaller, less capable AI models can already be run effectively on CPUs, particularly CPUs that have some AI acceleration capabilities built in. Intel announced new members of its Xeon 6 family of server CPUs on Monday, extending the lineup to cover lower price points and specialized use cases. The Xeon 6500 and 6700 series chips are aimed at data centers. These server CPUs are far more efficient and denser than Intel's previous-generation chips, and the company claims that customers can achieve up to 68% lower cost of ownership compared to 5-year-old systems. While these chips can be paired with AI accelerators as part of an AI training cluster, Intel also noted that they deliver up to 50% greater AI inference performance compared to the most recent server CPUs from rival AMD. Intel also launched its Xeon 6 for network and edge chips, which target radio access network and other edge computing applications. These chips are up to 70% more power efficient than their predecessors, and they pack an AI punch as well. Intel noted that a 38-core video edge server system is capable of running AI inference on 38 simultaneous camera streams. Not every AI use case requires the most advanced AI model trained on all the world's data. As more capable AI is crammed into smaller, cheaper models, CPUs with built-in AI acceleration can be an important piece of any company's AI strategy. A big market up for grabs As the AI industry matures, the initial mad dash to deploy AI solutions will give way to serious discussions about returns on investment. Intel's attempt to focus on cost efficiency with its AI accelerators fell flat due to software issues, but a similar strategy could work with its CPUs. For applications involving small, fine-tuned AI models capable of being run without pricey AI accelerators, servers full of Intel's latest Xeon 6 CPUs could be the most cost-effective solution. In other cases where AI accelerators are necessary, Intel's CPUs can still be part of the equation. IDC estimates that total annual spending on machine learning and analytics will reach $361 billion by 2027, with $153 billion of that total coming from generative AI spending. As capable AI models become cheaper to run and more efficient, a growing share of this spending could go to infrastructure without high-end AI accelerators. Being shut out of the AI accelerator market, Intel's overall AI opportunity is smaller than it would have otherwise been if the company had managed to turn Gaudi into a success. But even with that setback, Intel still has a horse in the AI race.