Intel’s GPU D-Series, previously codenamed Crescent Island, officially launched at the 2026 Open Compute Project Global Summit, marking the company’s latest attempt to challenge Nvidia and AMD in the AI inference market. The flagship D270 model, built on Intel’s Xe3P architecture, delivers 1.3 petaflops of FP8 compute performance and 1.5 terabytes per second of memory bandwidth, with a 350W thermal design power (TDP). The card uses LPDDR5X memory, offering up to 480GB of capacity—triple the 160GB on Intel’s reference design—while avoiding the high costs and supply constraints of high-bandwidth memory (HBM).
Technical Specifications and Design Choices
The D270 features 32 Xe3P cores, 256 XMX matrix engines, and 32MB of unified L2 cache, optimized for AI inference rather than traditional 3D graphics. Unlike Nvidia’s HBM-based GPUs, which require liquid cooling, the D270 operates in air-cooled server racks, aligning with Intel’s focus on cost-effective, scalable deployment. The card’s LPDDR5X memory configuration—160GB for Intel-branded units and up to 480GB for partner variants—prioritizes capacity over peak bandwidth, a trade-off Intel argues suits memory-bound inference workloads.
Intel emphasized the D-Series’ compatibility with its Xeon processors, aiming to create heterogeneous systems
that combine CPUs and GPUs for agentic AI tasks. The company highlighted partnerships with OEMs like ASUS, HPE, and Dell, as well as software support for frameworks like vLLM and Hugging Face. However, the D270’s 1.5TB/s memory bandwidth lags behind HBM-based competitors like Nvidia’s Vera Rubin, which achieves 3.7TB/s, raising questions about its performance in bandwidth-intensive scenarios.
Intel’s Xe3P architecture includes 512 KB of L1$ per Xe core and 16 MB total, along with 32 MB of unified L2 cache. The design eliminates conventional 3D-graphics and ray-tracing hardware, dedicating silicon area to compute and AI processing. The D270’s 350W TDP and air-cooled PCIe 5.0 x16 form factor allow deployment in standard server infrastructure, contrasting with Nvidia’s liquid-cooled systems. Intel’s reference card ships with 160GB of LPDDR5X, while ODM partners can configure up to 480GB, a capacity exceeding any other PCIe accelerator announced to date.
Market Positioning and Competitive Landscape
Intel positioned the D-Series as a value-oriented alternative to Nvidia’s Blackwell and AMD’s MI450X, targeting enterprises seeking cost-efficient inference solutions. The company noted that AI inference now accounts for two-thirds of AI compute demand, with memory capacity—rather than raw FLOPS—becoming a critical differentiator. However, Intel faces challenges: Nvidia retains over 80% of the data center GPU market, and AMD’s roadmap includes competing offerings.
The D-Series’ launch comes amid Nvidia’s supply constraints for HBM-equipped GPUs, creating a window for alternatives. The success of agentic AI depends not on a single component but on open heterogeneous infrastructure,
Intel's Nicolas Dube stated, emphasizing the need for ecosystem support.
Intel’s GPU D-Series enters a market where it has struggled to gain traction. The Data Center GPU Max family, code-named Ponte Vecchio, failed to establish a meaningful presence, and subsequent efforts were shelved. The D-Series represents an interim product, following the cancellation of Rialto Bridge in 2023 and Falcon Shores in 2025. Intel’s focus on inference aligns with the growing demand for agentic AI, which requires systems to handle multi-step tasks rather than single prompts.
Partnerships and Ecosystem Development
Intel has partnered with a broad ecosystem of OEMs, ODMs, and software vendors to roll out D-Series systems. Key partners include ASUS (2357.TW), HPE (HPE), Dell (DELL), and Supermicro, as noted in source 1. The company also collaborates with independent software vendors and AI-native companies to integrate the D-Series into existing workflows. Intel’s software stack includes support for Meta’s Muse Glimmer, a 30-billion-parameter open-weight model designed for local agentic workloads.
At the 2026 Open Compute Project Global Summit, Intel showcased rack-scale systems built to the OCP Open Rack v3 specification around Xeon 6+ processors. These configurations reach over 46,000 CPU cores per rack at 133.45 kW, as detailed in source 5. The company also highlighted its 116 contributions to OCP since 2011, emphasizing its long-term commitment to open standards.
Financial Performance and Stock Market Reaction
Intel’s stock has surged 183% year to date, driven by demand for its server processors and the resurgence of interest in its GPU initiatives. The company’s stock closed at $101.35 on October 12, 2026, down 3.20% for the day, but remained up 175% over the past 12 months. Analysts noted that Intel’s focus on AI infrastructure aligns with broader market trends, though challenges in GPU adoption persist. The stock’s performance reflects investor optimism about the D-Series’ potential to diversify Intel’s revenue streams.
Intel’s GPU D-Series also faces competition from Arm-based CPUs, which are gaining traction in data centers. Nvidia’s Grace chip has enabled standalone CPU racks, while Qualcomm has entered the fray with Arm-based solutions. Intel’s Xeon 6+ processors, which can scale to 46,000 cores per rack, compete with these offerings, but the company’s GPU strategy remains critical to its long-term goals in AI.
Customer Sampling and Future Availability
Customer sampling for the D270 began in late 2026, with general availability expected in 2027. Intel’s roadmap includes rack-scale systems using Xeon 6+ processors, aiming to integrate GPUs and CPUs for multi-tenant AI services.
However, the company has not yet disclosed pricing or detailed performance benchmarks, leaving questions about its competitiveness. A 2026 source noted that the D270’s software ecosystem and real-world performance
will determine its viability, with third-party benchmarks and customer deployments likely to provide clarity.
Intel’s Chief Technology Officer Sachin Katti, emphasized the shift from static training to real-time inference, stating, Scaling these complex workloads requires heterogeneous systems that match the right silicon to the right task, powered by an open software stack.
The D-Series is designed to support models with up to 300 billion parameters, with eight GPUs capable of handling trillion-parameter models. However, Intel has not provided independent benchmarks or detailed model configurations to validate these claims.
Despite the challenges, Intel remains committed to its GPU strategy. The company has not announced pricing or a general availability date for the D-Series, but its partnerships and technical specifications suggest a focus on enterprise adoption. As agentic AI continues to evolve, the success of the D-Series will depend on its ability to deliver scalable, cost-effective solutions that meet the demands of modern data centers.