Foxconn and Intel partner with SambaNova to develop rackscale AI infrastructure.

      The most significant statement in Intel's Computex announcement did not pertain to a chip but rather a ratio. The company suggested that as AI workloads transition from training to inference, the typical setup of four GPUs for every CPU will shift closer to a one-to-one ratio, bringing Intel's widely sold processors back towards the center of data centers.

      This concept underpins the partnership revealed in Taipei on June 2nd, where Intel, SambaNova, and Foxconn announced plans to create rackscale AI infrastructure for data centers, hyperscale endeavors, and what Intel refers to as intelligence center deployments, all utilizing Intel Xeon processors.

      During the presentation, they showcased production-ready racks that combine Xeon chips with SambaNova's SN-50 Reconfigurable Dataflow Units, emphasizing their performance in inference per watt and per dollar rather than focusing solely on raw training capability.

      Foxconn's contribution involves serving as the integration layer. As the largest electronics manufacturer globally, it will provide system integration for the rackscale platform and intends to develop a CPU-rich variant for workloads that don’t require additional acceleration. This includes cost-optimized inference, data processing, and hybrid AI.

      The two companies also expressed intentions to explore collaboration in design services and custom silicon development, a more flexible aspect of the announcement that Intel is likely keen to solidify into tangible results.

      Intel’s CEO Lip-Bu Tan presented the moment as pivotal, referencing "the rise of inference, agentic, and physical AI" alongside Intel's long history of collaborating with partners in Taiwan.

      The underlying analysis supported this rhetoric. Ben Bajarin, a principal at Creative Strategies and quoted by Intel, clearly articulated the transition: while the training phase typically involved one CPU per four GPUs, agentic inference suggests a shift to one CPU for every GPU, or even fewer.

      Foxconn was just one name among several others. Intel also highlighted expanded or new partnerships with Siemens, Hitachi, Echo Neurotechnologies, and Greenstone Biosciences, each focused on industry-specific silicon.

      Moreover, a separate endeavor called Vector Core Compute, created by Vista Equity Partners and Cambium Capital, demonstrated a fully disaggregated inference system utilizing Xeon for orchestration, SambaNova RDUs for decoding, and Nvidia Blackwell GPUs for prefill, with Together.ai as its inaugural commercial client.

      At the core of this rack initiative is Intel's new Xeon 6+ processor, marking its first data center CPU developed on the 18A process. Intel stated that a single liquid-cooled rack could deliver 36,864 cores in 32U of space at around 100 kilowatts, a density that targets operators looking to host agents without needing to redesign their facilities.

      What was absent from the announcement was a monetary figure, any equity stake, or a volume commitment from Foxconn. This represents a statement of intent from a chipmaker seeking to regain significance in AI, a contract manufacturer with the capacity to produce whatever the market demands, and a dataflow-chip startup betting that the economics of inference will favor alternatives to the current GPU leaders. The viability of the one-to-one ratio remains the key question underpinning this arrangement. The racks are tangible, but the hypothesis is still under evaluation.