Who makes DPUs for AI data centers
Data Processing Units that offload networking, storage, and security functions from the host CPU onto a dedicated, programmable chip.
A DPU takes infrastructure work — packet processing, storage virtualization, security functions — off the server CPU (Server CPUs) so those CPU cycles can go to orchestrating GPUs and other workload instead. It typically shares the board with the NIC or SuperNIC (NICs / SuperNICs) and, in some designs, is combined with it on a single card. Adoption is not universal: it's a design choice tied to a given server architecture and hyperscaler infrastructure stack rather than a component every AI server needs. For investors, DPU exposure is concentrated in a small number of vendors, including at least one hyperscaler that has built its own in-house rather than buying merchant silicon, which caps the addressable market for the merchant suppliers at whichever fleets choose to buy rather than build. The category is worth tracking less for standalone unit growth and more as a signal of which server architectures a given hyperscaler or OEM has standardized on, since DPU attach is usually decided at the platform level rather than per order.
AI delta: Offloading networking, storage, and security work from the host CPU matters more as GPU-dense AI servers push CPU cycles toward scarcity, since every cycle not spent on infrastructure work is a cycle available for orchestrating accelerators.
Companies on this part (5)
| Company | Role | Ownership | Ticker | Price | 1d | Confidence | Note |
|---|---|---|---|---|---|---|---|
| AMD | manufacturer | public | AMD | 517 USD | +2.2% | inferred | Pensando |
| Intel | manufacturer | public | INTC | 86.96 USD | +2.6% | inferred | IPU |
| Marvell | manufacturer | public | MRVL | 229 USD | +3.7% | inferred | OCTEON |
| Microsoft | in house | public | MSFT | 525 USD | +0.5% | inferred | in-house |
| Nvidia | manufacturer | public | NVDA | 211 USD | -0.7% | inferred | BlueField |
Who competes here
The same table read the other way: every company above competes for DPUs sockets. Confidence tags matter — an inferred placement is a lead, not a confirmed rivalry.
Seen in real machines
DPUs is named in 2 of the rack teardowns and supercomputer breakdowns on this site. Each line below is a piece of that machine. The companies shown are the ones named on that line, not necessarily the maker of this part.
- Nvidia GB300 NVL72 · rack teardown
- ConnectX-8 SuperNICs, BlueField-3 DPUs — Nvidia
- AMD Helios (MI455X) · rack teardown
- Pensando Salina DPU (P4-programmable, 16 Arm N1 cores; front-end networking, storage and security offload) — AMD
Nearby parts
- Previous in layer: NICs / SuperNICs
- Next in layer: FPGAs
- Layer: Layer 1 — Compute silicon & packaging
Watch
Independent videos — no affiliation; nothing loads from YouTube until you click.
Latest feed
- [AI Infrastructure Insight] How AI infrastructure will be designed for the future — newsroom:sk-hynix, 2026-10-22Effective real-world AI application requires improvements across faster compute, wider memory bandwidth, higher-performance networking, better storage, and more stable power infrastructure.
- Intel says AI 'gold rush' is over, pushes prompt routing across open hardware — DigiTimes, 2026-10-15Intel announced at the OCP Global Summit that the era of massive, centralized AI models is shifting toward distributed prompt routing across open hardware.
- Xsight Labs Adds E1L to DPU Portfolio with Support for up to 200G — HPCwire, 2026-10-09Xsight Labs has announced the addition of the E1L DPU to its product portfolio, supporting speeds up to 200G.
- Xsight Labs E1L DPU for Lower-Power 200Gbps DPUs — ServeTheHome, 2026-10-08Xsight Labs has announced the E1L, a low-power 200Gbps Data Processing Unit (DPU) designed for high-efficiency networking.
- Full-Stack Semiconductor Solutions for Smart, Secure Industry and Digital Energy — EE Times, 2026-09-27NSING Technologies offers complete semiconductor solutions, including MCUs and hardware root-of-trust chips, for both industrial automation and AI data centers.
- Delos Data Targets Heterogeneous AI with Data Interface — EE Times, 2026-09-24Delos introduced the Apollo chiplet, designed to create a low-latency interface spanning GPUs, accelerators, CPUs, and memory.
- Episil says AI demand marks structural shift in semiconductor epitaxy — DigiTimes, 2026-09-23Episil Precision noted that AI demand is structurally changing semiconductor epitaxy, requiring expansion into materials like silicon, GaN, SiC, and GeSi for various components.
- Taiwan advances major science park expansion to support AI, semiconductor growth — DigiTimes, 2026-09-22Taiwan plans to expand its Longtan Science Park to accommodate the continued growth of AI and semiconductor industries.
- SignatureIP Unveils MesoLink, a Mesochronous Bridge for PCIe/CXL Validation — HPCwire, 2026-09-16SignatureIP launched MesoLink, a new IP core that acts as a mesochronous connectivity bridge, allowing PCIe/CXL validation with full-rate SerDes on the FPGA side.
- Nvidia, MediaTek deepen 20-year chip ties — DigiTimes, 2026-09-14Nvidia invested US$3.5 billion in MediaTek's bonds, expanding their partnership to cover custom AI chips, automotive, and data center needs.
Questions about DPUs
- What are DPUs?
- Data Processing Units that offload networking, storage, and security functions from the host CPU onto a dedicated, programmable chip.
- Which companies make DPUs for AI data centers?
- This catalog lists 5 companies on DPUs, 5 of them with listed equity: AMD, Intel, Marvell, Microsoft, Nvidia.
- How does AI change demand for DPUs?
- Offloading networking, storage, and security work from the host CPU matters more as GPU-dense AI servers push CPU cycles toward scarcity, since every cycle not spent on infrastructure work is a cycle available for orchestrating accelerators.
- How reliable are these DPUs supplier placements?
- The 5 placements on this part break down as 5 inferred. "disclosed" means the company or a filing says so; "reported" comes from trade press or a reference-design list; "inferred" is a taxonomy lead with no citation yet.
Sources: placements without a source URL come from the taxonomy seed and are tagged accordingly. How this data is built.· Markdown · JSON