Intel pitches right-sized inference systems with D-series GPU and Xeon racks at OCP Summit

Intel’s strategic argument is that inference, not training, will define enterprise AI economics, and that agentic AI will bring the CPU back to the center of the data center. For three years, AI infrastructure has been dominated by bigger models, bigger clusters and larger checks written to a small group of suppliers. Intel said most organizations currently send every request to a frontier model, even when the task does not require one. Its alternative is an orchestration layer that evaluates each request by required depth of intelligence, context length, latency, data sensitivity and sovereignty, cost and energy. Deep research may justify a frontier model, proprietary documents may need to stay inside the enterprise, and routine coding, summarization or operational tasks can run on a smaller open model at a fraction of the cost.

Dubé said open models are expanding deployment options. “Open models are expanding the options enterprises have for deploying AI,” he said. “To realize their full potential, organizations need infrastructure that can match different workloads with the right intelligence, compute resources and economics.” Intel cited Vercel AI Gateway data showing that open-weight models’ share of observed token traffic rose from roughly 31% in early July to about 78% by late September. SiliconANGLE reported that one gateway is not the entire market, but the data aligns with what IT leaders say about rising token bills and concerns over sending sensitive data to external models.

The Intel GPU D-series, formerly code-named Crescent Island, is the hardware embodiment of that thesis. First disclosed at last year’s OCP Summit, it is built on the Xe3P architecture and delivers 1.3 petaflops of FP8 compute and 1.5 terabytes per second of memory bandwidth in a 350-watt, air-cooled PCIe form factor. The design supports up to 480 gigabytes of LPDDR5X memory. Intel’s own branded card carries 160 gigabytes, and partners can build higher-capacity cards.

Intel chose LPDDR memory instead of expensive, supply-constrained high-bandwidth memory, trading peak bandwidth for capacity and cost. The company says the D-series supports models up to 300 billion parameters on a single card and more than 1 trillion parameters in an eight-GPU system. Most enterprise data centers were not built for liquid-cooled racks drawing more than 100 kilowatts, so a 350-watt card that fits a standard air-cooled server lets a company start with one card, move to an eight-way server, and scale to rack-level systems when the workload demands it.

Intel also used the summit to highlight Xeon rack work for agentic AI. Agentic AI turns inference into a multi-step workflow spanning retrieval, reasoning, tool calls, database queries and application execution. Every agent generates tokens and a significant amount of general-purpose work that runs on CPUs. Intel showcased partner-built Xeon rack designs, many based on Open Rack v3, with select Xeon 6+ configurations scaling beyond 46,000 cores per rack at 133.45 kilowatts.

Xeon 6+ with dense, efficient cores targets high-concurrency agent workloads, while Xeon 6 with performance cores handles tasks that depend on single-thread speed. Intel argues that the right metric is no longer core count but how many concurrent agents a rack can sustain at target latency. The result is a heterogeneous system: Xeon for orchestration, data processing and tool execution; the D-series for open-model inference; and SambaNova as an optional fast-token engine where speed is the priority.