SambaNova Systems — AI supply-chain exposure

SambaNova is a fabless designer of AI chips built around a reconfigurable dataflow architecture — its SN40L — rather than the general-purpose GPU design that dominates training. It sells them not as loose chips but as full-stack systems, and runs an inference cloud that serves large models directly. That places it in the accelerator layer as a challenger, competing on how efficiently it can serve a trained model rather than on raw programmability.

The structural bet is that inference — running models in production — is a different and ultimately larger workload than training, and that a purpose-built dataflow design can serve big models at lower cost and power than repurposed GPUs. Its hook is architectural specialisation paired with a systems-and-cloud go-to-market that sidesteps the incumbent's software lock-in. The exposure is that it draws on the same foundry and memory supply as everyone else, and must win share from an entrenched ecosystem to matter.