# Pensando Pensando is a programmable packet-processing engine, acquired in 2022 and now the networking layer of its owner's rack-scale platform. Its interest here is as the clearest single example of a movement running through the whole industry: the infrastructure functions of a cloud, networking, storage access, encryption, and tenant isolation, have been migrating out of host software and into dedicated silicon. > [!note] The one-sentence version > A network engine you program rather than configure, so the plumbing of a cloud runs on its own silicon instead of on the customer's processors. ## What programmable means here Conventional network silicon implements a fixed set of protocols decided at design time. A programmable data plane is specified in a dedicated language that describes how packets are parsed, matched, and acted upon, and that description is compiled onto match-action hardware. The operator can therefore implement a new encapsulation, a new telemetry scheme, or a new policy without waiting for a silicon generation, while retaining hardware throughput and latency; see [[wiki/DPU and SmartNIC|DPU and SmartNIC]]. The functions that move onto such an engine are substantial: virtual switching and overlay encapsulation, stateful firewalling and policy, load balancing, line-rate encryption, telemetry, storage protocol handling, and the remote memory operations that cluster fabrics depend on; see [[wiki/RDMA|RDMA]] and [[wiki/Congestion Control|Congestion Control]]. ## Why the category exists at all Two forces. **Host economics:** at high line rates, infrastructure processing consumes a large fraction of a server's processor capacity, and recovering it is pure margin in a business that sells those processors to tenants; see [[wiki/Host CPU|Host CPU]]. **Isolation:** running the provider's functions on silicon the tenant does not control produces a genuine hardware boundary rather than a software convention, which is what makes dedicated physical machines safely rentable; see [[wiki/Bare Metal Compute|Bare Metal Compute]] and [[wiki/Confidential Computing|Confidential Computing]]. This movement is industry-wide and contested by several suppliers, including the major accelerator vendors, the established network silicon firms, the specialist switch-ASIC designers, and the hyperscale operators building their own. The common conclusion across all of them is that the control points of a cloud belong in hardware, and whoever supplies that hardware holds a position in how facilities are operated. ## Relationships - **class of:** [[wiki/DPU and SmartNIC|DPU and SmartNIC]] - **performs:** [[wiki/RDMA|RDMA]] · [[wiki/Congestion Control|Congestion Control]] - **deployed in:** [[wiki/AMD Helios|AMD Helios]] · [[wiki/Rack and Cluster Integration|Rack and Cluster Integration]] - **enables:** [[wiki/Bare Metal Compute|Bare Metal Compute]] · [[wiki/Data Center Security|Data Center Security]] - **part of:** [[wiki/AMD|AMD]] · [[wiki/Software and Orchestration|Software and Orchestration]] - **source articles:** [[articles/Vertically Integrating an AI Superpower from AI Factory to Citizen|Vertically Integrating an AI Superpower from AI Factory to Citizen]] · [[articles/The Gigawatt Is the New Measure of War|The Gigawatt Is the New Measure of War]] - **collection route:** [[collections/American Infrastructure|American Infrastructure]] · [[collections/American|American Collection]] ## Evidence posture - **Established:** programmable match-action data planes specified in a dedicated language and compiled to hardware; the function set migrating onto such engines including virtual switching, overlay encapsulation, stateful policy, load balancing, line-rate encryption, telemetry, storage protocol handling, and remote memory operations; host processor cost of infrastructure processing at high line rates; hardware-enforced tenant isolation enabling dedicated physical machine rental. - **Established:** acquisition of this engine in 2022 and its role as the networking layer of the owner's rack-scale platform, presented by the vendor as a differentiating low-latency capability. - **Established:** the category contested by multiple accelerator vendors, network silicon firms, switch-ASIC designers, and hyperscale operators building in-house. ## Sources - [Helios Is AMD's First AI System To Rival Nvidia Vera Rubin](https://www.cnbc.com/) — CNBC, October 2026; the 2022 acquisition, the programmable low-latency engine, and its role in the rack-scale platform.