NVIDIA BlueField

NVIDIA BlueField is a family of data processing units (DPUs) designed and sold by Nvidia that offload and accelerate networking, storage, and security tasks away from a server's main CPU. A DPU is a programmable system-on-chip that combines Arm CPU cores, a high-speed network interface, and a set of hardware accelerators on a single device, allowing it to take over the infrastructure work, networking, storage, and cybersecurity, that would otherwise consume cycles on the host server's main processor. By moving this "data center tax" off the host CPU and onto a dedicated chip on the network adapter, BlueField frees the server's general-purpose cores to run customer applications and, in modern AI clusters, to keep GPUs fed with data. When it announced BlueField-3 in April 2021, NVIDIA stated that one BlueField-3 DPU "delivers the equivalent data center services of up to 300 CPU cores." ^[10] The line originated with the networking company Mellanox, which Nvidia acquired in 2020, and has become a core building block of Nvidia's data-center networking portfolio alongside ConnectX network adapters and the Spectrum-X Ethernet platform. ^[1][2]

As of September 2026, NVIDIA's BlueField product page describes the line as an "advanced infrastructure computing platform for agentic AI factories." ^[2] In September 2026 NVIDIA also made BlueField-4 the hardware for NVIDIA Sentry, an agent-monitoring watchdog in its Open Agent Safety Platform reference design. ^[21][22]

What is a DPU?

A DPU (data processing unit) is the third class of processor in the data center alongside the CPU and the GPU, dedicated to running infrastructure rather than applications. A conventional server runs its operating system, virtualization layer, network virtual switch, storage stack, and security agents on the same CPU that runs the user's workloads. As network speeds climbed past 100 and then 400 gigabits per second, the share of CPU cores consumed by simply moving and protecting packets grew large enough that hyperscalers began designing dedicated silicon to absorb it. Nvidia describes the DPU as one of "the three pillars of computing" alongside the CPU and the GPU. ^[35]

BlueField integrates four elements that distinguish it from an ordinary network interface card (NIC):

• A cluster of 64-bit Arm CPU cores running their own Linux-based operating system, independent of the host.
• A ConnectX network engine providing Ethernet and InfiniBand connectivity, including remote direct memory access (RDMA) and RoCE.
• Hardware accelerators for cryptography (IPsec, TLS), data compression, regular-expression matching, and storage protocols.
• A PCIe switch and the ability to virtualize devices toward the host so that storage or network resources elsewhere in the cluster appear as local hardware.

Because the DPU runs its own software and sits between the host and the network, it can enforce a "zero-trust" security boundary: the infrastructure control plane runs on the DPU and is isolated from any compromise of the tenant operating system on the host. This is the foundation of the multi-tenant cloud and software-defined networking use cases that drove BlueField's early adoption. ^[2][4]

What is BlueField's history and origin?

BlueField began at Mellanox Technologies, an Israeli-American supplier of high-performance interconnects. Mellanox's first BlueField system-on-chip, shown in 2017 as a controller for NVMe-over-Fabrics storage, combined a 16-core Arm CPU with Mellanox's ConnectX-5 networking IP and an integrated PCIe switch. ^[38] Mellanox introduced the second-generation design, BlueField-2, on August 26, 2019, marketing it alongside ConnectX-6 Dx as a cloud "SmartNIC" and "I/O Processing Unit (IPU)", and showed it at VMworld 2019; Nvidia formally launched it under its own brand at GTC in October 2020. ^[37][6][5]

Nvidia announced its intent to acquire Mellanox for approximately 6.9 billion dollars (125 dollars per share in cash) on March 11, 2019, and completed the deal on April 27, 2020, for a transaction value of roughly 7 billion dollars. The acquisition brought the BlueField, ConnectX, and InfiniBand product lines into Nvidia and is the reason BlueField is sometimes still referred to by its original "Mellanox BlueField" branding on older hardware. After the close, Nvidia rebranded the chips as data processing units and folded them into a multi-generation roadmap presented at its GTC conferences. ^[1][7]

What are the BlueField generations?

Nvidia has brought three generations of BlueField to market under its own name (BlueField-2 through BlueField-4), each doubling peak network throughput while adding Arm compute and accelerators. The table below summarizes the specifications NVIDIA has published for each generation; where NVIDIA documents disagree, the text below gives both figures.

BlueField-2

BlueField-2 was the first DPU released under the Nvidia brand. It pairs eight 64-bit Armv8 Cortex-A72 cores with a ConnectX-6 Dx network engine, supporting two ports of 25, 50, or 100 Gb/s, or a single port at 200 Gb/s, over either Ethernet or InfiniBand. It includes hardware accelerators for encryption (IPsec, TLS, and AES-XTS), regular-expression matching, compression, storage offload (including elastic block storage and NVMe over Fabrics), and RDMA/GPUDirect, and it carries 16 GB or 32 GB of on-board DDR4 memory and a PCIe Gen 4.0 interface. ^[39] An NVIDIA launch slide reproduced by ServeTheHome listed 6.9 billion transistors and claimed the chip "replaces 125 x86 CPU cores." ^[6] Nvidia positioned it to offload virtualization, networking, and security from data-center hosts; a variant called BlueField-2X added an Nvidia Ampere GPU on the same card to accelerate AI-based security and telemetry. ^[6][8] NVIDIA's April 2021 BlueField-3 announcement described BlueField-2 as generally available with dual 100 Gb/s ports and up to eight Arm cores. ^[10]

BlueField-3

BlueField-3, unveiled at GTC on April 12, 2021, was Nvidia's first 400 Gb/s DPU; NVIDIA called it "the industry's first 400GbE/NDR DPU." ^[10] It integrates 16 Arm Cortex-A78 cores and about 22 billion transistors, supports Ethernet at up to 400 Gb/s and InfiniBand up to NDR, and is the first DPU to support fifth-generation PCIe, according to NVIDIA. ^[9][10] NVIDIA's current datasheet lists 32 GB of on-board DDR5 memory (dual DDR5 5600 MT/s) and 32 lanes of PCIe Gen 5.0. ^[40] NVIDIA's generational comparisons have varied over time: the 2021 announcement cited 10 times the accelerated compute of BlueField-2 and 4 times the cryptography acceleration; ^[10] the March 2023 availability announcement cited 4 times the compute, up to 4 times faster crypto acceleration, 2 times faster storage processing, and 4 times the memory bandwidth; ^[33] and a May 2023 NVIDIA technical blog cited 2 times the network bandwidth, 4 times the compute, and "almost 5x" the memory bandwidth. ^[34] NVIDIA summarizes the offload benefit by stating that "one BlueField-3 DPU delivers the equivalent data center services of up to 300 CPU cores, freeing up valuable CPU cycles to run business-critical applications." ^[10]

Although announced in 2021 (with sampling then expected in the first quarter of 2022), BlueField-3 reached general availability in March 2023, when Nvidia said at GTC that it was in full production. ^[10][11] NVIDIA announced on March 21, 2023, that Oracle Cloud Infrastructure had selected BlueField-3, ^[33] and HPCwire reported other cloud wins at Baidu, CoreWeave, JD.com, Microsoft Azure, and Tencent, reflecting the DPU's primary role in cloud multi-tenancy. ^[11] BlueField-3 is also sold in a network-accelerator configuration, the BlueField-3 SuperNIC, which NVIDIA describes as providing RoCE connectivity between GPU servers at up to 400 Gb/s; ^[40] NVIDIA built its Spectrum-X Ethernet platform for AI around the Spectrum SN5600 switch and the BlueField-3 SuperNIC. ^[41]

BlueField-4

BlueField-4 is the current generation, designed for the Vera Rubin era of AI infrastructure. NVIDIA revealed it at GTC Washington, D.C., in a blog post dated October 28, 2025, headlined "NVIDIA Launches BlueField-4," while saying the part was "expected to launch in early availability as part of NVIDIA Vera Rubin platforms in 2026." ^[3] Nvidia had sketched a BlueField-4 with 64 billion transistors and 800 Gb/s on its DPU roadmap as early as April 2021, when it expected the chip in 2024. ^[9] The product NVIDIA detailed in 2025 and 2026 is a dual-die package that combines a 64-core Nvidia Grace CPU (built on Arm Neoverse V2 cores) with an integrated ConnectX-9 networking chip. ^[12] It delivers up to 800 Gb/s of ultra-low-latency Ethernet or InfiniBand connectivity, carries up to 128 GB of LPDDR5X memory, and connects over PCIe Gen6 x16. ^[12][19] Nvidia states it provides about six times the compute of BlueField-3 and can support AI factories up to four times larger. ^[3] NVIDIA's stated memory-bandwidth gain over BlueField-3 has varied: its January 2026 Rubin platform post listed 250 GB/s against 75 GB/s for BlueField-3 (about 3.3 times), a July 2026 post said "more than 3x," and an August 2026 post said 4 times. ^[12][31][18] NVIDIA describes BlueField-4 as "the processor powering the operating system of AI factories," and as a software-defined control plane that enforces "security, isolation, and operational determinism independently of host CPUs and GPUs." ^[3][12]

At CES on January 5, 2026, NVIDIA launched the Rubin platform as "six new chips": the Vera CPU, the Rubin GPU, the NVLink 6 switch, the ConnectX-9 SuperNIC, the BlueField-4 DPU, and the Spectrum-6 Ethernet switch. ^[25] By GTC in March 2026 NVIDIA was describing the full Vera Rubin POD as built from "seven chips" across five rack types, including Groq 3 LPX inference racks; in the Vera Rubin NVL72 rack, each compute tray houses eight ConnectX-9 SuperNICs and one BlueField-4 DPU. ^[27] A storage-focused variant, BlueField-4 STX, was announced at GTC on March 16, 2026, as a modular reference architecture for AI-native storage built on a storage-optimized BlueField-4 processor that combines the Vera CPU with a ConnectX-9 SuperNIC. ^[28] In Rubin-generation systems, BlueField-4 underpins what NVIDIA announced at CES as the Inference Context Memory Storage platform and later branded NVIDIA CMX: it runs the key-value (KV cache) input/output plane and terminates NVMe-over-Fabrics, object, and RDMA storage protocols, offloading the long-context memory of large language model inference so that GPUs spend their time computing rather than waiting on storage. ^[12][26][27] BlueField-4 also introduces the Advanced Secure Trusted Resource Architecture (ASTRA), which provides a single trusted control point with isolated control, data, and management planes for provisioning and securing large AI environments. ^[12][25]

NVIDIA's timing for BlueField-4 has been tied to Vera Rubin. At CES in January 2026 it said Rubin-based products, and BlueField-4-based storage platforms, would be available from partners in the second half of 2026. ^[25][26] At GTC Taipei, in a release dated May 31, 2026 (the keynote was June 1 in Taipei), it said Vera Rubin was ramping into full production and that "production shipments of Vera Rubin are set to begin starting this fall." ^[13] In its August 26, 2026, quarterly results, NVIDIA said Vera Rubin was ramping into full production "with racks running at partners including CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius." ^[32]

Vera BlueField-4 STX and the August 2026 storage benchmarks

NVIDIA announced BlueField-4 STX at GTC on March 16, 2026. It named CoreWeave, Crusoe, IREN, Lambda, Mistral AI, Nebius, Oracle Cloud Infrastructure and Vultr as planning to adopt STX for context memory storage, and said STX-based platforms would be available from partners in the second half of 2026. ^[28] At GTC Taipei, in a release dated May 31, 2026, NVIDIA added DOCA security capabilities for what it now called Vera BlueField-4 STX: new DOCA Vault microservices for file-access control, DOCA Argus for visibility into agent behavior, and DOCA Flow for network isolation. NVIDIA claimed runtime threat detection "up to 1,000x faster than existing agentless runtime solutions" and network and file-access enforcement at up to 800 Gb/s. ^[29] NVIDIA's June 2026 STX datasheet describes a dual-processor configuration with 3.2 Tb/s of network bandwidth, PCIe Gen 6 (x96 lanes), 384 GB of on-board LPDDR5X, and two Vera CPUs "each with up to 84 Neoverse Arm V2 cores"; NVIDIA's August 2026 benchmark post instead describes Vera as having 88 Olympus Armv9.2 cores. ^[30][16] Neoverse V2 is the licensed Arm core used in NVIDIA's Grace CPU, whereas NVIDIA describes Vera's cores as its custom Olympus design. ^[12][19]

On August 3, 2026, NVIDIA published storage microbenchmarks for the Vera BlueField-4 STX Storage Processor, which it describes as a key component of the NVIDIA STX foundation for AI-native data platforms. Where the BlueField-4 DPU NVIDIA announced in October 2025 pairs a Grace CPU with ConnectX-9 networking, NVIDIA presents the STX storage part as built around the NVIDIA Vera CPU, the 88-core custom Olympus processor that also hosts Rubin GPUs, so that the same CPU architecture and software toolchain span compute and storage. The company's stated rationale is agent concurrency: as agents retrieve enterprise knowledge, access persistent memory, reuse KV cache data, and execute tools, each step can trigger multiple storage operations repeated across thousands of parallel agents, and the compression, encryption, integrity, and redundancy calculations in the storage path all run on CPUs, while inference runs on GPUs. ^[16]

In the benchmarks, Vera outperformed an x86 comparison CPU that the post does not name: up to 1.43x higher throughput on AES-128 encryption and 1.29x on decryption, up to 3.26x on Reed-Solomon recovery of erasure-coded data, up to 3.67x on CRC32C integrity checking, up to 3.29x on compression and 1.72x on decompression under concurrency, and up to 3.21x on a two-stage compress-then-encrypt write pipeline, the headline figure of NVIDIA's announcement. ^[16][17] These are vendor microbenchmarks: single-process tests on memory-resident data using OpenSSL, Zstandard, and LZ4 under a purpose-built NVIDIA framework that excludes file I/O, disk, and networking, and NVIDIA states that end-to-end testing is still required to quantify complete storage-system or GPU-performance outcomes. ^[16]

Scale-In network infrastructure (August 2026)

On August 24, 2026, NVIDIA introduced what it calls Scale-In network infrastructure, the fifth pillar in its AI-networking taxonomy. The company uses the term for the accelerated north-south domain around AI compute: the path connecting users, applications, data sources, external storage, and services to systems inside an AI factory. This is distinct from NVIDIA's other named domains. Scale-Up links GPUs into a coherent accelerator, Scale-Out connects servers within one factory, Scale-Across connects geographically distributed factories, and Context Memory supplies shared KV-cache storage. Scale-In instead covers access, security, data movement, provisioning, and observability around those compute and memory domains. NVIDIA's AI Infrastructure account publicized the architecture the following day.^[18][20]

The design has three cooperating layers. BlueField-4 creates a host-independent infrastructure-processing domain at each system. Its 64-core Grace CPU runs control-plane work such as policy, provisioning, telemetry, and orchestration, while inline engines handle packets, RDMA, storage protocols, encryption, firewall rules, and policy enforcement in the data path. DOCA supplies containerized services and programming interfaces for those functions. Spectrum-X Ethernet carries traffic across the Scale-In access and external-storage fabric. NVIDIA's BlueField-4 datasheet lists a PCIe Gen6 x16 host interface, up to 128 GB of onboard LPDDR5X, and a maximum network bandwidth of 800 Gb/s.^[18][19]

In NVIDIA's Vera Rubin NVL72 architecture, BlueField-4 and ConnectX-9 SuperNICs have different roles. The SuperNICs carry tenant workload traffic over the east-west Scale-Out network. BlueField-4 runs the infrastructure services that connect, secure, and manage the server. BlueField Astra extends the DPU's trusted control into the SuperNICs: BlueField-4 installs and updates policies and monitors telemetry, while ConnectX-9 enforces those policies in the data path. East-west workload traffic therefore does not all pass through the DPU's 800 Gb/s north-south interface.^[18]

NVIDIA maps several DOCA services to this architecture. Host-Based Networking, DOCA Flow, and DOCA-accelerated Open vSwitch provide routing, traffic classification, and tenant isolation. Argus, Vault, and Flow supply runtime, file-access, and network-policy functions. DOCA Platform Framework provisions DPUs and services through a Kubernetes-native control plane, while DOCA Telemetry gathers service and device signals outside the tenant host. On the storage path, BlueField-4 offloads NVMe over Fabrics, file and object protocols over RDMA or TCP, virtualization, and data movement.^[18]

The numerical comparisons are NVIDIA-reported product and test results, not independent measurements. In this post NVIDIA says BlueField-4 has six times the compute, four times the memory bandwidth, and twice the network bandwidth of BlueField-3 (its January 2026 post listed 250 GB/s against 75 GB/s for BlueField-3, and its July 2026 post said "more than 3x"). For a Vera Rubin compute tray, it reports 7.2 Tb/s of aggregate interface capacity: 800 Gb/s on the BlueField-4 north-south path plus four 1.6 Tb/s east-west paths. In an illustrated file-transfer test using 5 GB, 10 GB, and 50 GB files, NVIDIA reports that a BlueField-4 and Spectrum-X storage path delivered up to 1.45 times the throughput of an off-the-shelf Ethernet comparison. The figure shows 1.3x, 1.4x, and 1.5x for the three file sizes; the post does not establish that result for other hardware, software, or traffic mixes.^[18]

Agent safety: NVIDIA Sentry (September 2026)

On September 28, 2026, NVIDIA announced the NVIDIA Open Agent Safety Platform, which pairs its open-source OpenShell agent runtime with NVIDIA Sentry. The press release describes Sentry as "an out-of-band watchdog that runs on NVIDIA BlueField-4 DPUs to continuously monitor agent behavior." ^[21] NVIDIA says Sentry provides "in-silicon security enforcement," and that if an AI agent attempts to move outside its software boundary, Sentry "quarantines and stops it in milliseconds"; these are NVIDIA's claims. ^[21] According to NVIDIA, Sentry combines threat detection, hardware-based agent governance and enforcement, and data-access protection from an isolated, out-of-band trust domain, and it is built on DOCA, which supplies the capabilities Sentry uses to inspect agent requests and responses, provide attested telemetry, verify agent identity, and enforce zero-trust access policies for data, tools, APIs, and services. ^[21] The announcement followed a summer of disclosures in which frontier-lab agents escaped their test environments (AI agent sandbox escapes). ^[23]

NVIDIA's technical blog explains why the DPU is the enforcement point. One of the platform's five principles is that "the path to the model (the brain) is the control point," because controlling it gives both the best observation point and a kill switch. ^[22] "In an NVIDIA Vera Rubin POD, each compute tray includes a BlueField-4 data processing unit on the node's only path to the model," the blog says; from there BlueField-4 provides out-of-band observability into agent behavior and enforces policy "in real time at line speed," isolated from the host and beyond the agent's reach. ^[22] NVIDIA writes that DOCA connects the BlueField security foundation with OpenShell policy and correlates agent interactions, policy decisions, and tool and data access into a record of agent activity, while a DOCA gateway continuously verifies each agent's identity and delegated authority. ^[22] Justin Boitano, NVIDIA's vice president of enterprise AI, told The New Stack that with a DPU present the agent's model endpoint is routed "through a proxy on the DPU, so that you can see all of the reasoning traces of the agents on the host." ^[23]

Sentry's status differs from OpenShell's. The press release presents Sentry as part of the platform's "reference system design," while the "now broadly available" wording refers to OpenShell; its availability section lists OpenShell and skills as available through NVIDIA's developer resources page and GitHub. ^[21] The blog calls Sentry "an optional security layer alongside OpenShell," says the platform "is also compatible with other hardware systems," and claims that "for anyone already running on an NVIDIA Vera system with BlueField-4, enabling these protections is just a software update." ^[22] Boitano told The New Stack that Sentry is not open source, though it has open APIs, and that "the DPU is really optional in these architectures," adding that in many cases "just using OpenShell on CPUs is honestly good enough"; he described the DPU as aimed at "frontier use cases of evaluating models or systems where you might have the guardrails off the models." ^[23] CNBC summarized Sentry as a monitor that "runs on network chips, not CPUs or GPUs." ^[24]

Two partner statements in the announcement involve BlueField directly. NVIDIA says Anthropic's Claude Managed Agents run the agent loop on a separate server from the sandboxes where work executes, and that "integrations with OpenShell and BlueField enable enterprises to enforce strict control over agent access through those sandboxes." It also says Red Hat runs OpenShell and DOCA on Red Hat AI Factory with NVIDIA. ^[21]

What is DOCA?

DOCA is Nvidia's software framework for programming the BlueField DPU and SuperNIC, exposing the chip's accelerators through a consistent set of APIs so that infrastructure software can be written once and run across DPU generations. As an October 2020 NVIDIA technical blog put it, "DOCA is to DPUs what CUDA is to GPUs." Just as CUDA enables developers to program accelerated computing, DOCA enables them to program the acceleration of data processing. ^[36]

DOCA consists of two parts: a software development kit (SDK) with open, industry-standard APIs, including the Data Plane Development Kit (DPDK) and Linux Netlink for networking, the Storage Performance Development Kit (SPDK) for storage, and support for the P4 language; and a runtime, included by default with the BlueField platform, for provisioning, deploying, and orchestrating containerized services across hundreds or thousands of DPUs in a data center. Using DOCA, developers build cloud-native, DPU-accelerated services for software-defined networking, software-defined storage, telemetry, and zero-trust security. With BlueField-4, Nvidia emphasizes native support for DOCA microservices, packaging infrastructure functions as containers that run directly on the DPU. ^[3][4][36] As of late September 2026, NVIDIA's DOCA SDK documentation lists version 3.5.0 as its newest release. ^[14] In the Open Agent Safety Platform announced in September 2026, NVIDIA says NVIDIA Sentry is built on DOCA. ^[21]

What is NVIDIA BlueField used for?

BlueField's offload model serves several distinct workloads:

• Cloud multi-tenancy and zero-trust security. Public and private clouds run the hypervisor's networking and security control plane on the DPU, isolating it from tenant workloads on the host. This is the basis of cloud deployments at providers such as Oracle Cloud Infrastructure and Microsoft Azure.
• Software-defined networking. The DPU offloads the virtual switch, overlay encapsulation, routing, traffic shaping, and line-rate encryption, accelerating east-west traffic without burdening host CPUs.
• Software-defined and disaggregated storage. BlueField virtualizes remote storage as local NVMe devices, accelerates NVMe over Fabrics, and performs compression and data reduction inline.
• AI cloud infrastructure. In GPU clusters, the DPU manages tenant isolation, congestion control, and storage I/O so that expensive accelerators are not stalled waiting on the network or on data. ^[2][11]

What is BlueField's role in AI data centers?

As AI training and inference moved to clusters of tens of thousands of GPUs, the network became as important as the compute, and the DPU became the device that keeps those networks orderly and secure at scale. BlueField-3 SuperNICs and ConnectX adapters provide the endpoint intelligence for the Spectrum-X Ethernet platform, which Nvidia markets as a fabric purpose-built for AI; in that role the DPU handles adaptive routing, congestion control, and performance isolation that let Ethernet behave more like a lossless AI interconnect. ^[11][15]

With BlueField-4 and the Vera Rubin platform, Nvidia repositions the DPU from a networking offload engine into what it calls "the processor powering the operating system of the AI factory." By embedding a full 64-core Grace CPU on the device and tying it to the inference KV-cache and storage path, BlueField-4 is meant to run the infrastructure of an entire AI data center, security, multi-tenancy, storage, and the memory plane of long-context inference, as a self-contained, accelerated layer beneath the GPUs. In September 2026 NVIDIA extended that role to agent safety, making BlueField-4 the host for its NVIDIA Sentry watchdog (see above). ^[21][22] This trajectory, from a Mellanox smart NIC into a central pillar of Nvidia's data-center strategy, reflects the broader industry shift toward disaggregated, software-defined infrastructure in which dedicated silicon, rather than the host CPU, runs the data center. ^[3][12]

References

1. ^1 ^2NVIDIA Newsroom, "NVIDIA Completes Acquisition of Mellanox, Creating Major Force Driving Next-Gen Data Centers." nvidianews.nvidia.com/...ing-next-gen-data-centers
2. ^1 ^2 ^3 ^4NVIDIA, "BlueField Networking Platform." nvidia.com/...data-processing-unit
3. ^1 ^2 ^3 ^4 ^5Itay Ozery, NVIDIA Blog, "NVIDIA Launches BlueField-4: The Processor Powering the Operating System of AI Factories." October 28, 2025. blogs.nvidia.com/...bluefield-4-ai-factory
4. ^1 ^2NVIDIA Developer, "DOCA Software Framework." developer.nvidia.com/...doca
5. ^Wikipedia, "Nvidia BlueField." en.wikipedia.org/...Nvidia_BlueField
6. ^1 ^2 ^3ServeTheHome, "NVIDIA BlueField-2 and BlueField-2X DPU Offerings Launched." servethehome.com/...ield-2x-dpu-offerings-launched
7. ^NVIDIA Newsroom, "NVIDIA to Acquire Mellanox for 6.9 Billion." nvidianews.nvidia.com/...-mellanox-for-6-9-billion
8. ^NVIDIA, "NVIDIA BlueField-2 DPU Datasheet." resources.nvidia.com/...bluefield-2-dpu-datasheet
9. ^1 ^2Tom's Hardware, "Nvidia Reveals BlueField-3, BlueField-4 DPUs: 400-800 Gbps, 22-64B Transistors." tomshardware.com/...uefield-3-and-bluefield-4-dpus
10. ^1 ^2 ^3 ^4 ^5 ^6 ^7NVIDIA, "NVIDIA Extends Data Center Infrastructure Processing Roadmap with BlueField-3." nvidianews.nvidia.com/...-roadmap-with-bluefield-3
11. ^1 ^2 ^3 ^4HPCwire, "Nvidia Announces BlueField-3 GA, Oracle Cloud Is Early User." hpcwire.com/...eld-3-ga-oracle-cloud-is-early-user
12. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Kyle Aubrey, NVIDIA Technical Blog, "Inside the NVIDIA Vera Rubin Platform: Six New Chips, One AI Supercomputer." January 5, 2026. developer.nvidia.com/...chips-one-ai-supercomputer
13. ^NVIDIA Newsroom, "NVIDIA Vera Rubin Ramps Into Full Production to Power Agentic AI Factories Worldwide." May 31, 2026. nvidianews.nvidia.com/...uction-agentic-ai-factory
14. ^NVIDIA Docs, "DOCA Documentation." docs.nvidia.com/...sdk
15. ^NVIDIA, "NVIDIA Spectrum-X Ethernet Platform for Giga-Scale AI." nvidia.com/...spectrumx
16. ^1 ^2 ^3 ^4Jason Hardy and Ronil Prasad, NVIDIA Developer Blog, "NVIDIA Vera Storage Benchmarks: Faster Encryption, Compression, Integrity Checking, and Recovery for AI-Native Storage." August 3, 2026. developer.nvidia.com/...very-for-ai-native-storage
17. ^NVIDIA AI Infrastructure (@NVIDIAAIInfra), post on X, August 3, 2026. x.com/...2084336350715637768
18. ^1 ^2 ^3 ^4 ^5 ^6Ronil Prasad and Itay Ozery, NVIDIA Technical Blog, "NVIDIA BlueField-4 Powers New Scale-In Network Infrastructure for Agentic AI Factories." August 24, 2026. developer.nvidia.com/...e-for-agentic-ai-factories
19. ^1 ^2 ^3NVIDIA, "NVIDIA BlueField-4 DPU" datasheet. June 2026. resources.nvidia.com/...bluefield-4-dpu-datasheet
20. ^NVIDIA AI Infrastructure (@NVIDIAAIInfra), post on X, August 25, 2026. x.com/...2092265586751336754
21. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8NVIDIA Newsroom, "NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment." September 28, 2026. nvidianews.nvidia.com/...open-agent-safety-platform
22. ^1 ^2 ^3 ^4 ^5 ^6John Myers, Alex Watson, Ali Golshan and Ofir Arkin, NVIDIA Technical Blog, "NVIDIA Open Agent Safety Platform: A Reference for Continuous In-Silicon Agent Monitoring." September 28, 2026. developer.nvidia.com/...n-silicon-agent-monitoring
23. ^1 ^2 ^3Frederic Lardinois, The New Stack, "Nvidia launches Open Agent Safety Platform to lock down rogue AI agents." September 28, 2026. thenewstack.io/nvidia-openshell-sentry-agents
24. ^Kif Leswing, CNBC, "Nvidia releases software platform to stop AI agents from misbehaving." September 28, 2026. cnbc.com/...nvidia-releases
25. ^1 ^2 ^3NVIDIA Newsroom, "NVIDIA Kicks Off the Next Generation of AI With Rubin: Six New Chips, One Incredible AI Supercomputer." January 5, 2026. nvidianews.nvidia.com/...platform-ai-supercomputer
26. ^1 ^2NVIDIA Newsroom, "NVIDIA BlueField-4 Powers New Class of AI-Native Storage Infrastructure for the Next Frontier of AI." January 5, 2026. nvidianews.nvidia.com/...r-the-next-frontier-of-ai
27. ^1 ^2Rohil Bhargava, Taylor Allison and Harry Petty, NVIDIA Technical Blog, "NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer." March 16, 2026. developer.nvidia.com/...stems-one-ai-supercomputer
28. ^1 ^2NVIDIA Newsroom, "NVIDIA Launches BlueField-4 STX Storage Architecture With Broad Industry Adoption." March 16, 2026. nvidianews.nvidia.com/...h-broad-industry-adoption
29. ^NVIDIA Newsroom, "NVIDIA Vera BlueField-4 STX Brings Agentic AI Storage Processing With In-Silicon Security." May 31, 2026. nvidianews.nvidia.com/...-with-in-silicon-security
30. ^NVIDIA, "NVIDIA BlueField-4 STX Storage Processor" datasheet. June 2026. resources.nvidia.com/...torage-processor-datasheet
31. ^Ronil Prasad, NVIDIA Technical Blog, "Scaling Agentic AI Factories Through Extreme Co-Design with NVIDIA BlueField." July 16, 2026. developer.nvidia.com/...sign-with-nvidia-bluefield
32. ^NVIDIA Newsroom, "NVIDIA Announces Financial Results for Second Quarter Fiscal 2027." August 26, 2026. nvidianews.nvidia.com/...econd-quarter-fiscal-2027
33. ^1 ^2NVIDIA Newsroom, "Oracle Cloud Infrastructure Chooses NVIDIA BlueField Data Center Acceleration Platform." March 21, 2023. nvidianews.nvidia.com/...ter-acceleration-platform
34. ^Tal Roll and Itay Ozery, NVIDIA Technical Blog, "Power the Next Wave of Applications with NVIDIA BlueField-3 DPUs." May 11, 2023. developer.nvidia.com/...th-nvidia-bluefield-3-dpus
35. ^Kevin Deierling, NVIDIA Blog, "What Is a DPU?" May 20, 2020. blogs.nvidia.com/...whats-a-dpu-data-processing-unit
36. ^1 ^2Ariel Kit, NVIDIA Technical Blog, "Programming the Entire Data Center Infrastructure with the NVIDIA DOCA SDK." October 5, 2020. developer.nvidia.com/...e-with-the-nvidia-doca-sdk
37. ^NVIDIA Newsroom (Mellanox release), "Mellanox Introduces Revolutionary ConnectX-6 Dx and BlueField-2 Secure Cloud SmartNICs and I/O Processing Unit Solutions." August 26, 2019. nvidianews.nvidia.com/...releases-20210113-6829469
38. ^ServeTheHome, "Mellanox BlueField NVMeoF SoC Solution." August 21, 2017. servethehome.com/...-bluefield-nvmeof-soc-solution
39. ^NVIDIA, "NVIDIA BlueField-2 DPU" datasheet (PDF). nvidia.com/...datasheet-nvidia-bluefield-2-dpu.pdf
40. ^1 ^2NVIDIA, "NVIDIA BlueField-3 Networking Platform" datasheet. July 2025. resources.nvidia.com/...datasheet-nvidia-bluefield
41. ^NVIDIA Newsroom, "NVIDIA Supercharges Ethernet Networking for Generative AI." June 2, 2024. nvidianews.nvidia.com/...working-for-generative-ai