Intel's D-Series GPU Enters Customer Sampling With Dell, HPE, ASUS and Supermicro in the Mix as the Agentic AI Push Widens

Intel (INTC) has begun customer sampling of its new D-series GPUs, formerly code-named Crescent Island, expanding its push into AI inference as demand grows for the infrastructure behind autonomous AI agents. Intel unveiled the chips at the OCP Global Summit on Monday, Oct. 12. Sampling is not commercial deployment: Intel has not given production volumes, pricing or revenue projections.

What Intel is selling

The D-series is built to run trained AI models and services, not to train them. Unlike Nvidia's Blackwell server configurations, Intel's GPUs come as PCIe expansion cards that fit into standard servers. The first card, the D270, is rated at 1.3 PFLOPS of FP8 compute and 350 W, and is air-cooled. It uses LPDDR5X memory, with 160 GB on Intel's own card and up to 480 GB on partner cards. Reports note that LPDDR5X bandwidth sits well below what HBM-based designs deliver, so Intel is pitching capacity per watt and per dollar, not peak speed.

Who is lined up

Intel's OEM partners include ASUS, Dell and HPE, according to coverage of the launch. StorageReview, which saw Intel's partner slide, reported 53 logos covering server makers and ODMs including Dell, HPE, Lenovo, Supermicro, Cisco, ASUS, ASRock Rack, Giga Computing and MiTAC, plus software names such as Red Hat, Canonical, vLLM and Hugging Face. Neoclouds were described as sampling recipients, but none was named.

A partner logo shows who is supporting the platform, not who has committed to buy it. No customer has disclosed an order.

The agentic AI angle for CPUs

Intel is also tying the launch to its Xeon business. Its argument is that AI agents, which search databases, navigate websites and manage applications alongside GPU-powered models, need plenty of general-purpose compute. Intel introduced new Xeon-based rack designs the same day, with six partner systems from ASRock Rack, ASUS, GIGABYTE, MiTAC, MSI and Supermicro listed as available today. We did not find independent data showing how much CPU demand agentic workloads are adding, so treat that as Intel's thesis.

The competition

Intel is entering a crowded field. AMD's MI350P is about 40% higher in FP8 compute at 2.3 PFLOPS and up to 600 W, according to ServeTheHome, which also said the D270 is not meant to match Nvidia's Blackwell Ultra. In CPUs, Intel faces AMD, Nvidia's Arm-based Grace processors and Qualcomm's expanding data center portfolio.

What is still missing

Customer sampling means selected customers are testing the chips. There are no independent benchmarks, no volume or availability date, no pricing and no revenue guidance, and server vendors are not expected to ship D-series systems until 2027, per ServeTheHome.