{"slug": "top-arm-based-innovations-advancing-ai-infrastructure-on-device-intelligence-and", "title": "Top Arm-based innovations advancing AI infrastructure, on-device intelligence and physical AI in July 2026", "summary": "NVIDIA's Vera Rubin platform, featuring the Arm-based Vera CPU, is moving into production with support from 300 global partners, and early results from CoreWeave show it delivering 10 times more throughput per megawatt than Grace Blackwell NVL72 on a DeepSeek-R1 benchmark. Additionally, Databricks and Microsoft extended their partnership into the 2030s, with Databricks increasing use of Arm-based Azure Cobalt processors, planning to adopt Cobalt 200. PSYONIC's Ability Hand, using Arm-based computing, advances bionic dexterity for prosthetics and robotics.", "body_md": "# Top Arm-based innovations advancing AI infrastructure, on-device intelligence and physical AI in July 2026\n\nJuly’s roundup of Arm-based innovations shows how the Arm ecosystem is pushing intelligence from the cloud to the edge, turning breakthrough research into real-world systems that are already reshaping AI infrastructure, devices, robotics and even space exploration.\n\nThis month’s highlights range from new infrastructure for agentic AI and evidence-based optimization for AI coding agents to intelligent mobile applications, browser-based AI and tools that make complex graphics workloads easier to understand. Beyond software, Arm-based technology is also helping robots move from demonstrations into commercial environments, giving bionic hands more responsive control and taking autonomous machines to the Moon.\n\n## NVIDIA Vera Rubin reaches a major AI infrastructure milestone\n\nNVIDIA Vera Rubin is moving into production with support from 300 global partners, including cloud and infrastructure providers deploying the platform around the world.\n\nAt the center of the rack-scale platform is the Arm-based NVIDIA Vera CPU, designed to coordinate the data movement, control flow and orchestration work surrounding increasingly complex AI models and agents. NVIDIA says Vera Rubin combines seven chips and five rack trays through system-level codesign rather than assembling independent, off-the-shelf components.\n\nEarly results from CoreWeave show Vera Rubin NVL72 delivering 10 times more throughput per megawatt than Grace Blackwell NVL72 on a DeepSeek-R1 benchmark, illustrating how performance per watt is becoming one of the defining measures for power-constrained AI infrastructure.\n\n## PSYONIC advances bionic dexterity for people and robots\n\nPSYONIC is developing bionic hardware designed to bring greater dexterity to both prosthetic users and robotic systems. Dermot O’Driscoll speaks with PSYONIC founder and CEO Dr. Aadeel Akhtar about the company’s Ability Hand. Muscle signals from a user’s residual limb are translated into movement, while sensors in the fingertips detect pressure and provide touch feedback.\n\n[Arm-based computing helps coordinate sensing, motor control and feedback in real time](https://www.arm.com/company/success-library/psyonic), allowing the hand to respond quickly and helping users control their grip when handling objects ranging from everyday tools to delicate items.\n\nThe same technology can also be used as a robotic hand. As intelligent machines move into real-world environments, responsive and power-efficient control will be important for systems that must sense, decide and act reliably.\n\n## Databricks expands its use of Arm-based Microsoft Azure Cobalt\n\nDatabricks and Microsoft have extended their strategic partnership into the 2030s, with Databricks deepening its use of Microsoft Azure infrastructure for its own business operations, analytics and enterprise AI services.\n\nAs part of the agreement, Databricks is increasing its use of Arm-based Azure Cobalt processors. It currently uses [Cobalt 100](https://newsroom.arm.com/blog/microsoft-azure-cobalt-100-vms-arm-neoverse) and plans to adopt [Cobalt 200](https://newsroom.arm.com/blog/microsoft-azure-cobalt-200-arm-neoverse-css-v3) to improve performance and efficiency across agentic AI and data-intensive workloads.\n\nThe announcement provides further evidence of Arm-based CPUs being adopted for large-scale cloud and AI platforms where compute density, efficiency and predictable performance are increasingly important.\n\n## CPUs provide the control layer for persistent local AI agents\n\nAs AI applications move beyond isolated prompts and responses, developers need to consider how agents maintain context, coordinate memory, retrieve information and manage workflows over time.\n\nOdin Shen, Principal Physical AI BU Champion, explores [why CPUs play a central role in persistent local AI systems](https://developer.arm.com/community/arm-community-blogs/b/ai-blog/posts/rethinking-the-role-of-cpus-in-ai-agent-orchestration-on-dgx-spark). Rather than focusing only on model execution, the CPU can act as the system’s control layer, coordinating state, memory, retrieval, events and control flow across multiple stages of an agent workflow.\n\nUsing DGX Spark as a practical environment, the article shows how CPU-side orchestration can connect local inference with semantic memory and contextual retrieval. This system-level approach helps turn an impressive one-off AI demonstration into a stable runtime that can continue to respond, remember and operate over time.\n\n## Arm Performix gives AI coding agents evidence for optimization\n\nAI coding agents are increasingly being asked not only to write software, but also to find and resolve performance problems. Doing this reliably requires measurement and architecture-aware profiling rather than assumptions based on general coding knowledge.\n\nHenry Wang, Staff Software Engineer, explains [how an Arm Performix agent skill gives tools such as Claude Code and GitHub Copilot](https://developer.arm.com/community/arm-community-blogs/b/servers-and-cloud-computing-blog/posts/teaching-ai-coding-agents-to-optimize-arm-workloads-with-arm-performix) the knowledge needed to select profiling recipes, run an investigation and interpret the results. In the example, an agent used Arm Performix to identify a memory-bound loop in a C workload. After changing the loop order, runtime fell from 26.03 seconds to 4.67 seconds, producing a 5.6-times speedup with an unchanged checksum. The agent then profiled the application again to validate the improvement rather than assuming the change had worked.\n\nThe demonstration shows how profiling data can make AI-assisted optimization more measurable, repeatable and architecture-aware.\n\n## ASRock Rack brings the Arm AGI CPU to high-density server platforms\n\nThe launch of the [Arm AGI CPU ](https://newsroom.arm.com/blog/introducing-arm-agi-cpu)introduced a production silicon platform designed for the compute density, memory bandwidth and power efficiency required by agentic AI infrastructure.\n\nASRock Rack was among the first manufacturers to offer systems powered by the Arm AGI CPU. At the OCP EMEA Summit earlier this year, that collaboration could be seen at rack level through the company’s 2OU2N-ARM server platform.\n\nThe system combines a 136-core Arm AGI CPU, based on Arm Neoverse V3, with 12 DDR5 memory channels and PCIe 6.0 connectivity within a 300-watt CPU power envelope. It shows how silicon, server and rack-level design can come together to support dense cloud services and AI inference workloads more efficiently.\n\n## Robotics moves from demonstrations to real-world deployment\n\nAs robotics matures, the industry conversation is shifting from what an individual machine can demonstrate to how autonomous systems can deliver value reliably and at scale. John Kourentis, Director of Market Development for Physical AI at Arm, discusses the opportunities emerging across transportation, warehousing, logistics, healthcare and hospitality [at the AGIBOT UK launch 2026.](https://newsroom.arm.com/blog/agibot-physical-ai-robotics-arm)\n\nDeploying robots in these environments requires more than increasingly capable AI models. Systems must combine real-time control with efficient AI processing, connect cloud-based development and simulation to edge deployment, and provide a consistent architecture across sensors, microcontrollers, onboard computing and cloud infrastructure.\n\nPower efficiency is also critical. Robots have limited battery capacity, cooling and physical space, but must still process sensor data, execute AI workloads, plan movement and respond to their surroundings in real time. These system-level requirements will play an increasingly important role as physical AI moves into commercial environments.\n\n## AMBA CHI C2C helps chiplets interoperate across generations\n\nChiplets give silicon designers more flexibility to combine specialized components, but scalable adoption depends on open standards that allow independently developed chiplets to communicate and work together. Francisco Socal, Director of Product Management, Architecture and Technology Group, [introduces a new property-negotiation guide](https://developer.arm.com/community/arm-community-blogs/b/soc-design-and-simulation-blog/posts/from-complexity-to-simplicity-scaling-and-future-proofing-chiplets-with-amba-chi-c2c-property-negotiation) and Issue B of the AMBA CHI C2C specification.\n\nProperty negotiation allows connected chiplets to discover one another’s capabilities and select the features needed for a particular system. This supports forward and backward compatibility while allowing anything from a simple peripheral to a fully coherent accelerator to use an appropriate subset of the protocol.\n\nIssue B also adds capabilities including multi-request operations, software-visible registers and multi-hop routing. These improvements are especially relevant as larger AI models increase demand for memory and interconnect bandwidth across increasingly complex heterogeneous systems.\n\nAny re-use permitted for informational and non-commercial or personal use only.", "url": "https://wpnews.pro/news/top-arm-based-innovations-advancing-ai-infrastructure-on-device-intelligence-and", "canonical_source": "https://newsroom.arm.com/blog/top-arm-based-innovations-in-july-2026", "published_at": "2026-08-05 12:18:51+00:00", "updated_at": "2026-08-05 14:46:40.171988+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-infrastructure", "ai-chips", "robotics"], "entities": ["NVIDIA", "Arm", "CoreWeave", "Databricks", "Microsoft", "PSYONIC", "Dermot O'Driscoll", "Aadeel Akhtar"], "alternates": {"html": "https://wpnews.pro/news/top-arm-based-innovations-advancing-ai-infrastructure-on-device-intelligence-and", "markdown": "https://wpnews.pro/news/top-arm-based-innovations-advancing-ai-infrastructure-on-device-intelligence-and.md", "text": "https://wpnews.pro/news/top-arm-based-innovations-advancing-ai-infrastructure-on-device-intelligence-and.txt", "jsonld": "https://wpnews.pro/news/top-arm-based-innovations-advancing-ai-infrastructure-on-device-intelligence-and.jsonld"}}