A SmartNIC is a network interface card with its own processing — FPGAs, Arm cores, or network ASICs — that offloads networking tasks from the server’s CPUs. A DPU (Data Processing Unit) is the current form of that idea: a purpose-built device, led by NVIDIA’s BlueField family, that offloads networking, storage, and security processing and presents itself to the host as infrastructure rather than as a peripheral. The market exists because AI-era servers spend a growing share of their CPU cycles on networking overhead — encryption, RDMA transport, virtual switching, storage protocols — and hyperscale operators would rather buy dedicated silicon for it than burn accelerators’ neighbors on it. 650 Group tracks this market quarterly; our SmartNICs research report covers share, forecast, and vendor detail. Here is the vendor landscape in brief.
NVIDIA is the reference vendor: its BlueField DPUs (BlueField-2 and BlueField-3 shipping today, BlueField-4 announced) combine its networking stack with Arm cores and are bundled into AI server designs — the pairing of Palo Alto Networks’ NGFW on BlueField-2 was an early security proof point. AMD entered through the $1.9B Pensando acquisition and fields the Pensando DPU line alongside its Xilinx FPGA heritage (Alveo and SDAccel-era parts), giving it both fixed-function and programmable routes. Intel fields the IPU (Infrastructure Processing Unit), built on its Mount Evans ASIC and Agilux FPGA paths, aimed at its hyperscale customers’ infrastructure offload. Marvell fields the OCTEON family — the long-running Cavium heritage — spanning NP-based SmartNICs and DPUs across telco and cloud. Broadcom, the merchant silicon leader, offers the Stingray/NetXtreme line and ties its NICs to its Tomahawk switch franchise; its advantage is the end-to-end Ethernet story.
Beyond the top five, the specialist and OEM field includes Napatech (acceleration appliances for network monitoring and security), Ethernity Networks (FPGA-based NICs for telco and security appliances), Chelsio (iWARP/RDMA offload NICs), Realtek at the value tier, and the heritage names — Myricom, Solarflare (now part of Xilinx/AMD), and StarTech — that seeded the category. The hyperscalers also build their own: Amazon’s Nitro DPU family powers AWS instances, Google has its own infrastructure processing efforts, and Microsoft works with the merchant vendors on its SmartNIC designs. That build-versus-buy tension — hyperscalers designing internal DPUs while NVIDIA sells BlueField into the same accounts — is the defining commercial dynamic of the market.
We would offer buyers three guideposts. First, define the workload before the vendor: RDMA-heavy AI fabrics point one way, security appliances another, telco edge a third. Second, weigh the software stack — the DPU is only as useful as its DOCA, IPU, or vendor SDK ecosystem, and that is where the practical differences sit today. Third, watch the generational cadence: BlueField-4, Intel’s next IPU, and AMD’s post-Pensando roadmaps land through 2026-2027, and share will move with them.
We believe the DPU sits at the most interesting boundary in the industry — the line between computing and networking — and that its market will grow faster than the server market around it for the next several years. Our quarterly share and forecast data for the full vendor set is in the SmartNICs report.
Tags: #SmartNIC #DPU #NVIDIA #BlueField #AMD #Pensando #Intel #IPU #Marvell #OCTEON #Broadcom #650 Group
650 Group researches several of the technologies we covered in this explainer, including Data Center, Co-Location & Cloud and Enterprise, and more.