Arm expands physical AI and cloud scale with SME2 and cobalt

The Tensor RoboCar integrates 433 Arm-based cores to support autonomous mobility, while Elastic Cloud on Microsoft Azure reports a 44% price-performance boost using Arm-based Azure Cobalt processors. These October 2026 innovations establish Arm as the primary compute foundation for both physical AI and cloud scaling. If you are deploying edge models or managing cloud search, these hardware shifts change your baseline for performance-per-watt and cost-per-query.

Arm establishes a common foundation for physical and agentic AI

Arm’s hardware strategy focuses on creating a unified compute foundation that allows partners to move beyond software-only AI toward “physical AI”—intelligence that interacts with the real world. By providing consistent architectures across smartphones, PCs, and cloud infrastructure, Arm enables developers to build agentic (self-acting) systems that can make independent decisions in real time. This convergence means that the same architectural logic used to power a cloud search engine can now power a humanoid robot or a smart wearable.

The Tensor RoboCar demonstrates this distributed intelligence by utilizing 433 Arm-based cores

The Tensor RoboCar demonstrates this distributed intelligence by utilizing 433 Arm-based cores, including Arm Neoverse AE (Arm’s high-performance computing architecture—a server chip) and Cortex processors. (Neoverse: a high-performance computing architecture from Arm used for server chips and cloud infrastructure.) This massive core count allows the vehicle to coordinate everything from road planning to steering and braking. By distributing the workload across the chassis, Arm solves the bottleneck of centralized processing for safety-critical autonomous systems.

Robotic learning also relies on this hardware consistency.

Robotic learning also relies on this hardware consistency. The PSYONIC Ability Hand uses 12 Arm-based microcontrollers to coordinate touch-sensitive feedback and motor control. This same hardware architecture allows human-generated grip data to be used to train machines, creating a direct link between prosthetic technology and industrial robotics.

SME2 provides critical throughput gains for on-device AI

Arm Scalable Matrix Extension 2 (SME2)—which provides AI math speed—acts as a performance multiplier for compact, quantized (compressed model) AI models. Liquid AI’s LFM3-1.7B model demonstrates this by achieving 3.1x faster prefill and 1.5x faster decoding when run on Arm CPUs with SME2 enabled. These gains allow developers to run more capable models locally, reducing the need for high-latency cloud round-trips.

Tencent Media Lab further leverages SME2 to make neural image compression practical on mobile

Tencent Media Lab further leverages SME2 to make neural image compression practical on mobile devices. By using matrix acceleration and reduced data movement, Tencent reduced the latency of a 2K image decoder from 61 ms to 30 ms on a reference smartphone using two CPU threads. This allows for high-quality image sharing without the computational overhead that typically drains mobile batteries.

How do I optimize AI models for Raspberry Pi 5?

Optimizing AI models for Raspberry Pi 5 involves using the Arm AI Model Optimizer to achieve up to 5.9x faster inference and 74% lower memory use for image-classification models. (inference: the process of an artificial intelligence model providing an output or prediction based on input data.) The tool, built on Nota AI’s NetsPresso, allows developers to evaluate how changes in precision and model structure affect performance on specific Arm hardware. This makes it possible to deploy models that were previously too heavy for low-power edge devices.

How much can I save on cloud cost using Azure Cobalt?

Saving on cloud costs using Azure Cobalt is achieved through the use of Arm-based Azure Cobalt processors, which Elastic reports provide a 44% price-performance boost for Elastic Cloud search compared to Fsv2 instances. These instances sit alongside traditional x86-based cloud infrastructure, offering a cost-effective way to scale search and retrieval workloads. For high-volume search services, this performance jump directly lowers the annual procurement cost of cloud resources.

The limits of current Arm-based innovations

Current Arm-based innovations are limited by the fact that Liquid AI’s prefill and decoding figures are provisional and were tested using only two CPU threads. While SME2 provides significant gains, the source does not specify how these improvements scale as model sizes increase beyond the 1.7B parameter mark. Furthermore, the specific requirements for the Robotics Capability Framework remain unstated, leaving the exact hardware targets for next-generation robots unconfirmed.

Verification of these claims depends on independent benchmarks of the LFM3 family

The claim that Arm provides a universal foundation holds as long as SME2 and Neoverse architectures can maintain efficiency gains as AI models grow in complexity. If the power overhead of larger models outpaces the efficiency of the Arm-based cores, developers may still be forced toward dedicated high-end GPU clusters for non-edge tasks. Verification of these claims depends on independent benchmarks of the LFM3 family on production-grade Arm silicon.

Who this is for: Robotics hardware engineers

Robotics hardware engineers who integrate Neoverse V3AE CPUs into humanoid platforms can now use the Agile ONE SN as a roadmap for distributed physical AI compute. These engineers should look for the Neoverse V3AE architecture to reduce the integration work that typically slows down real-world validation.

Edge AI developers who deploy models like Liquid AI’s LFM3 family can now achieve 3.1x faster prefill by helping ensure their target hardware supports SME2. These developers should verify that their hardware supports these specific matrix extensions before choosing a model to deploy locally.

Cloud infrastructure architects who manage Elastic Cloud search can now provision Azure Cobalt instances to realize a 44% price-performance improvement over Fsv2 instances. These architects should review the cost savings of Arm-based instances during their next annual procurement cycle.

Who this is not for: General consumer electronics shoppers

General consumer electronics shoppers who want to know which smartphone has the best camera should skip this analysis, as it focuses on architecture and developer tools. Those readers should instead look for consumer reviews of the Alexa Tablet 12 Pro or other retail-focused hardware guides.

Software-only mobile app developers who do not have control over the underlying hardware or compiler optimizations should ignore the SME2 and LFM3 benchmarks. These developers should instead consult the standard Android or iOS development documentation for high-level API usage.

A scenario for evaluating Arm hardware: Henrik’s deployment

Henrik runs a robotics startup and needs to reduce the latency of a real-time image recognition system on a mobile robot. He identifies that his current model is too slow and checks the Liquid AI LFM3-1.7B model for its efficiency on Arm hardware. Because his target device supports SME2, he implements the model and achieves 3.1x faster prefill, allowing his robot to react to obstacles in real time without a cloud connection.

If Henrik’s target device did not support SME2, he would be forced to rely on cloud infrastructure for inference, which would increase his system latency and significantly raise his monthly operational costs.

What these innovations can be used for

These hardware advancements enable several high-impact applications across different industries:

Guessing wrong on SME2 support leads to significantly higher latency or the forced adoption

Guessing wrong on SME2 support leads to significantly higher latency or the forced adoption of expensive cloud infrastructure to handle local tasks. Conversely, guessing wrong on the 44% price-performance gain of Azure Cobalt results in higher-than-necessary cloud spend during the annual procurement cycle. Choosing the wrong compute foundation now may result in a complete architectural redesign if the chosen hardware cannot scale with the next generation of agentic AI models.

Original reporting: Top Arm-based innovations advancing on-device AI, cloud infrastructure, and physical AI in October 2026 (Arm), on 2 November 2026.