Foxconn and Intel join SambaNova to build rackscale AI infrastructure

The most consequential line in Intel’s Computex announcement was not about a chip. It was about a ratio. As AI workloads move from training to inference, the company argued, the long-standing arrangement of four GPUs to every CPU collapses towards something closer to one to one, and the processor Intel actually sells well moves back towards the centre of the data centre.

That is the bet behind the partnership unveiled in Taipei on 2 June. Intel, SambaNova, and Foxconn said they intend to build rackscale AI infrastructure for data centre, hyperscale, and what Intel calls intelligence centre deployments, all built on Intel Xeon processors.

The companies showed production-ready racks pairing Xeon chips with SambaNova’s SN-50 Reconfigurable Dataflow Units, a combination pitched on inference performance per watt and per dollar rather than raw training horsepower.

Foxconn’s role is the integration layer. The world’s largest electronics manufacturer will provide system integration for the rackscale platform and plans to build a CPU-dense variant for workloads that do not need additional acceleration, including cost-optimised inference, data processing, and hybrid AI.

The two companies also said they would explore collaboration in design services and custom silicon development, the more open-ended part of the announcement and the one Intel will most want to convert into something concrete.

Intel chief executive Lip-Bu Tan framed the moment in generational terms, citing “the rise of inference, agentic, and physical AI” and Intel’s five decades of building foundational technology alongside partners in Taiwan.

The analyst case sat underneath the rhetoric. Creative Strategies principal Ben Bajarin, quoted by Intel, put the shift plainly: where the training era ran roughly one CPU per four GPUs, agentic inference moves that to one CPU to one GPU or fewer.

Foxconn was one name on a longer list. Intel also detailed expanded or new collaborations with Siemens, Hitachi, Echo Neurotechnologies, and Greenstone Biosciences, each aimed at industry-specific silicon.

Separately, a new enterprise inference cloud called Vector Core Compute, formed by Vista Equity Partners and Cambium Capital, demonstrated a fully disaggregated inference system running Xeon for orchestration, SambaNova RDUs for decode, and Nvidia Blackwell GPUs for prefill, with Together.ai as its first commercial customer.

Underpinning the rack story is Intel’s new Xeon 6+ processor, its first data centre CPU built on the 18A process. Intel said a single liquid-cooled rack can deliver 36,864 cores in 32U of space at roughly 100 kilowatts, a density figure aimed squarely at operators trying to host agents without redesigning their facilities.

What the announcement did not include was a dollar figure, an equity stake, or a volume commitment from Foxconn. It is a statement of intent between a chipmaker trying to reclaim relevance in AI, a contract manufacturer with the scale to build whatever the market orders, and a dataflow-chip start-up betting that inference economics will reward something other than the incumbent GPU. Whether the one-to-one ratio holds is the question the whole arrangement rests on. The racks are real. The thesis is still being tested.