Intel's AI Chip Strategy Takes Shape Around GPUs and Inference Workloads
Intel is building its challenge to Nvidia around GPUs designed for inference and agentic AI workloads, combined with CPUs and other processors working in tandem. CEO Lip-Bu Tan has made this strategy a top priority since arriving last year, signaling a shift away from years of failed attempts to compete in the accelerator chip market.
The company spent billions over 15 years developing accelerator chips to compete with Nvidia, only to pivot repeatedly as products failed to gain traction. The Xeon Phi processors, the Gaudi chips for AI training, and the cancelled Falcon Shores GPU are examples of initiatives that either underperformed or were abandoned.
Dominic Daninger, vice president of engineering at Intel systems integration partner Nor-Tech, said this history creates skepticism in the channel. "One of the things with Intel we've seen so many times was that they were out there a year or two with a particular product line, and if it's not sticking, it's gone," he said.
A New Direction: Crescent Island and Annual GPU Releases
Intel revealed its updated strategy in October with a 160-GB data center GPU code-named Crescent Island. The company positioned it as power- and cost-optimized for inference workloads running on air-cooled servers, with sampling expected to begin in the second half of 2026.
The GPU features Intel's Xe3P microarchitecture, optimized for performance-per-watt, and 160 GB of LPDDR5X memory. Intel said it plans an annual release cadence, matching similar commitments from Nvidia and AMD.
Anil Nanduri, Intel's vice president of product management for data center AI accelerators, acknowledged past failures. "We did not meet the needs of the frontier AI training market, and we didn't meet the market needs," he said at CES 2026 in January.
Heterogeneous Computing Returns as Core Strategy
Intel is pursuing a heterogeneous approach-using multiple processor types in a single system to optimize performance and cost. This strategy involves pairing compute-optimized GPUs for the "pre-fill" stage of language models with memory-bandwidth-optimized GPUs for the "decode" stage.
Sachin Katti, who led Intel's AI strategy before departing for OpenAI in November, said this approach could deliver significant efficiency gains. In testing, Intel ran pre-fill on an Nvidia GPU and decode on an Intel accelerator, achieving a 70 percent improvement in performance per dollar compared to single-vendor systems.
The strategy requires open software architecture that doesn't force developers to change their workflows. "If we can build such a heterogeneous infrastructure, then we can optimize that performance-per-dollar by making sure that the right part of that agentic workload runs on the right-priced hardware," Katti said in September.
Rack-Scale Solutions and Data Center Integration
Intel cancelled its Falcon Shores GPU to focus on a successor for rack-scale AI solutions. The next-generation product, initially code-named Jaguar Shores but now called the "Shores product line," aims to compete with Nvidia's Grace Blackwell GB200 NVL72 platform.
Tan identified rack-scale solutions as a top priority when he joined Intel. "There's no question we need to strengthen our position in the cloud-based AI data center market by developing competitive rack-scale system solutions," he wrote in Intel's annual report.
At CES 2026, Nanduri referenced the Shores product line on Intel's road map as the next step after Crescent Island, though the company has provided limited details on timing and specifications.
Key Hires Signal Commitment to GPU Development
In January, Intel hired Eric Demers, Qualcomm's longtime GPU leader, to head GPU engineering and data center GPU solutions. The company also hired Nicolas Dubé, a former HPE and Arm executive, to lead data center systems and solutions.
These appointments followed Intel's decision to fold its AI group back into the data center business unit under former Arm executive Kevork Kechichian. Tan said the merger ensures "tight coordination across CPUs, GPUs and platform strategy."
Kechichian wrote to employees that "AI and the modern data center are fundamentally linked," with customers standardizing on complete AI platforms spanning compute, networking, and software.
SambaNova Partnership and Reference Architecture
Intel announced a multiyear collaboration with AI chip startup SambaNova Systems in late February. Tan is chairman and investor in SambaNova, and Intel Capital participated in the company's $350 million Series E funding round.
In April, Intel and SambaNova detailed a reference architecture combining Intel's Xeon 6 CPUs, SambaNova's RDUs (Reconfigurable Data Units), and GPUs to handle different components of agentic AI models. The design uses GPUs for pre-fill, RDUs for decode, and CPUs for tool calling and system orchestration.
A SambaNova spokesperson said OEMs can use the reference architecture to build their own implementations. The companies positioned the design as requiring less energy and fitting into existing air-cooled data centers compared to Nvidia's Vera Rubin platform.
AI Workstations Emerge as Secondary Market
Outside data centers, Intel is expanding its Arc Pro B-series GPUs for AI workstations. The company expanded the lineup in March, positioning the GPUs as optimized for multi-user and multi-agentic AI workloads with strong price-to-performance for inference.
Nanduri said the Arc Pro B-series fits the emerging market for autonomous AI agents. "We have a very nice product that fits the right price point, fits the right capabilities and can give you a multi-card scaling that takes the graphics workstation to the next level," he said.
Execution Remains the Critical Question
Daninger from Nor-Tech said Intel's strategy could work only if the company delivers competitive performance at reasonable prices. "It would probably have to have a heck of a lot more performance than anything the competitor would have to make people interested," he said.
Tan acknowledged the challenge when he arrived at Intel. "It won't happen overnight, but I know we can get there," he said, referencing the company's need to learn from past mistakes.
Intel's latest earnings call in January showed strong demand for server CPUs due to agentic AI workloads, but Tan reaffirmed the company's commitment to accelerators. "The accelerator remains central to frontier AI, and we will continue to participate, innovate and partner in that category," he said.
For executives evaluating infrastructure investments, Intel's strategy represents a shift toward heterogeneous systems optimized for inference rather than training. Whether the company can execute remains uncertain given its track record, but the focus on performance-per-dollar and open standards suggests a different approach than previous attempts. For deeper understanding of AI strategy at the executive level, see AI for Executives & Strategy.
Your membership also unlocks:
