Samsung and Arm Reportedly Start a Custom On-Device AI Chip Project

Samsung Electronics and Arm reportedly started a custom on-device AI chip project after Arm approved initial engineering funding in late August. The claim reached Chinese readers through an rsshub 36kr news item on September 4. It originated with a report from South Korea and has not been independently confirmed by either company.

The reported arrangement goes further than Samsung simply licensing another Arm CPU design. Arm would provide accelerator architecture and register-transfer-level technology, while Samsung would design, manufacture, and deliver the completed system-on-chip to an unnamed customer.

That division of labor creates the real tension. Arm has traditionally supplied intellectual property to chip designers, while Samsung has competed as both a chip designer and contract manufacturer. A joint custom program would bring those roles closer together, potentially pressuring Qualcomm, MediaTek, and other suppliers competing for device AI workloads.

It would also test whether Samsung can turn vertical integration into dependable silicon. Designing a capable AI processor is only one part of that challenge. Manufacturing it economically on an advanced process, controlling heat, supporting software, and shipping it on schedule remain separate hurdles.

What the Samsung and Arm Report Actually Claims

The reported project combines Arm technology with Samsung's design and manufacturing operations, but its customer and delivery schedule remain undisclosed.

The original Korean chip report says Arm approved non-recurring engineering spending near the end of August. NRE covers one-time development work required before a chip reaches mass production.

According to the report, Arm commissioned Samsung to help build a customized system-on-chip, commonly shortened to SoC. An SoC integrates several computing functions onto one piece of silicon, reducing space and communication overhead inside a device.

Samsung's System LSI division would reportedly handle the chip design. Samsung Foundry would then manufacture it using a 2-nanometer process, according to the same account.

Arm would supply an AI accelerator architecture and RTL, which describes a digital circuit's behavior before engineers convert it into a physical layout. It would also communicate the customer's requirements and oversee development progress.

The unnamed end customer would reportedly receive the chip directly from Samsung. Samsung would collect the associated design and manufacturing revenue, while Arm would act as the technology provider and development partner.

Those details matter because they describe a complete commercial path, not a research demonstration. The project reportedly starts with customer requirements and ends with production silicon delivered by Samsung.

However, the account leaves major blanks. It does not identify the customer, targeted device category, production date, chip performance, model capacity, power envelope, or NRE amount.

It also uses industry sources rather than an announcement from Samsung or Arm. Neither company had published a release specifically confirming this project when the report appeared.

The rsshub 36kr summary preserved the central claims, but it was not the originating report. Readers should therefore distinguish the reported project from the companies' previously announced partnerships.

That distinction prevents an unverified development program from being mistaken for a finished processor. An approved engineering budget signals commitment, but it does not guarantee a successful tape-out or commercial launch.

A tape-out occurs when a completed chip design is sent for fabrication. Reaching that milestone requires architecture work, verification, physical design, software preparation, and validation against the customer's requirements.

The most defensible conclusion is narrow. Samsung and Arm reportedly began a customer-driven development effort, with Samsung positioned to design and manufacture the resulting chip.

Why This Is More Than Another Arm License

The important change is Arm's reported involvement in delivering a customer-specific chip through Samsung, not merely licensing reusable processor cores.

Arm built its business by licensing instruction sets and processor designs to companies that create their own chips. That model lets many vendors use compatible software while differentiating memory, graphics, connectivity, and specialized accelerators.

A customized AI SoC requires tighter coordination. The accelerator must communicate efficiently with CPUs, memory, security components, and the software that schedules machine-learning workloads.

Arm has already moved toward more integrated foundations through Compute Subsystems, or CSS. These packages combine verified CPUs, interconnects, system IP, software, and implementation guidance.

Arm says its compute subsystems can reduce integration work and help partners reach silicon sooner. The company cites partner-reported delivery improvements of up to 12 months and substantial NRE savings.

Those are general Arm claims, not measured outcomes for the reported Samsung project. Still, they explain why a customer might prefer a coordinated platform over assembling every component separately.

The project described by the rsshub 36kr item appears to extend that logic. Arm would contribute more of the technical foundation, while Samsung would perform the customization and physical implementation needed for production.

This model occupies a middle ground between standard chip licensing and a fully custom processor program. The customer gets more differentiation than a generic component offers without building an entire semiconductor organization.

Samsung also gains several possible revenue layers. System LSI can earn design work, Samsung Foundry can fabricate the chip, and Samsung's broader semiconductor operations can compete for memory or packaging business.

That potential does not mean every Samsung division will participate. The report specifically identifies System LSI and the foundry operation, while providing no confirmed memory or packaging details.

For Arm, the attraction is broader strategic control. More complete designs can shorten customers' development cycles and make Arm technology harder to replace within the resulting platform.

The approach also puts Arm closer to the final chip transaction. That direction deserves attention because Arm has been expanding beyond individual core licenses toward validated subsystems and more complete computing platforms.

The companies already have a foundation for this work. In February 2024, Samsung announced an optimized Cortex-X collaboration using its gate-all-around process technology.

Gate-all-around, or GAA, wraps the transistor gate around the channel. The structure is intended to improve electrical control as transistor dimensions shrink.

That earlier program used design-technology co-optimization. DTCO means chip architecture and manufacturing rules are tuned together instead of being treated as separate stages.

The newly reported effort has a different emphasis. It describes an AI accelerator, a custom customer design, and Samsung's responsibility for both design and fabrication.

If accurate, the project represents a deeper operational relationship. Arm supplies the computing foundation, and Samsung becomes the organization responsible for turning that foundation into a deliverable product.

On-Device AI Makes Integration the Main Contest

On-device AI rewards the supplier that can balance compute, memory, thermals, software, and manufacturing inside a strict power budget.

Running AI locally can reduce network dependence and latency. It can also keep selected inputs on the device, although actual privacy still depends on the application and its data handling.

The hardware challenge is unforgiving. A smartphone cannot cool a processor like a server, and it cannot devote unlimited battery capacity to continuous AI inference.

Inference is the process of running a trained model to produce an output. On a phone, it may support text generation, image editing, transcription, translation, search, or context-aware assistance.

These workloads do not depend on one accelerator alone. CPUs handle general instructions, GPUs process parallel operations, and neural processors execute machine-learning calculations efficiently.

Memory bandwidth can become another constraint. The chip must move model weights and intermediate data fast enough without consuming excessive energy or creating thermal pressure.

Arm's 2025 Lumex platform illustrates this system-level approach. It combines CPU, GPU, software, physical implementation support, and technologies designed to accelerate local AI.

Samsung publicly supported that platform. Nak Hee Seong, a Samsung semiconductor executive, said the company planned to use Arm's subsystem platform in future flagship mobile products.

That statement verifies an established direction, but it does not confirm the September 2026 report. The unnamed custom chip might use related technology, a separate accelerator architecture, or another implementation path.

The report's reference to an Arm AI accelerator architecture is particularly important. Arm offers several relevant components, yet the article does not provide a product name or specification.

Without that identity, it is impossible to determine how the chip divides work among its CPU, GPU, and dedicated AI hardware. It is also impossible to compare its performance with shipping processors.

Software may become the decisive factor. Developers need supported operators, quantization tools, model converters, runtimes, debuggers, and predictable behavior across applications.

Quantization reduces the numerical precision used by a model. It can lower memory use and accelerate inference, but aggressive compression can reduce output quality.

A custom chip must support these workflows before its theoretical compute becomes useful. Weak tooling can leave accelerator capacity idle or force developers to spend months rewriting models.

This is where Arm's broad software compatibility helps. Its CPU architecture already anchors most smartphones, and its AI libraries connect to widely used machine-learning frameworks.

Samsung can add device-specific optimization through Exynos software and Galaxy products. An external customer, however, would require documentation and support that extend beyond Samsung's own hardware.

The integration contest therefore has two layers. Samsung and Arm must first produce viable silicon, then make its AI capabilities accessible to the customer's developers.

That second layer often receives less attention than process nodes or benchmark scores. It determines whether a processor supports one demonstration or a sustained pipeline of useful applications.

Qualcomm and MediaTek Are the Immediate Pressure Point

A successful custom program would challenge the merchant-chip model led by Qualcomm and MediaTek, but Samsung still depends on those companies in important products.

Qualcomm packages CPU, GPU, neural processing, connectivity, camera processing, and software into Snapdragon platforms. Device manufacturers can adopt those platforms without sponsoring a custom SoC program.

MediaTek follows a similar model across premium and high-volume device segments. Both suppliers benefit from spreading development costs across many customers and product generations.

A custom Samsung-Arm chip offers a different proposition. It can target one customer's priorities, including model types, cameras, sensors, security requirements, or power limits.

The tradeoff is higher project risk. Custom silicon requires early commitments, detailed requirements, engineering funding, software coordination, and enough shipment volume to justify the effort.

Qualcomm remains especially important to Samsung's own mobile business. In July 2026, the companies announced an expanded Galaxy chip partnership spanning phones, watches, and intelligent eyewear.

Qualcomm said its Snapdragon platforms would support local, context-aware AI across those devices. Samsung described the products as the result of close engineering collaboration between the companies.

That relationship prevents a simple Samsung-versus-Qualcomm interpretation. Samsung can use Snapdragon in Galaxy products while its semiconductor divisions pursue other customers and custom designs.

The reported Arm project would pressure Qualcomm at the supplier level, not necessarily terminate Samsung's purchasing relationship. It gives device makers another possible route to differentiated Arm-based silicon.

Arm also works with multiple foundries and chip companies. Its value comes partly from serving a broad ecosystem rather than selecting one exclusive manufacturing partner.

Samsung must therefore compete for Arm-based implementations against other foundries. Arm technology alone does not guarantee that the reported chip will favor Samsung on cost, yield, or performance.

Yield measures the share of manufactured chips that meet required specifications. Poor yield can raise effective costs even when a process offers strong laboratory performance.

This is where the 2-nanometer claim becomes strategically significant. A successful customer chip would provide Samsung Foundry with design experience, production volume, and a reference for future business.

A delayed or inefficient implementation would create the opposite signal. Customers would have another reason to choose a competing foundry or an established merchant processor.

Apple and Google provide additional context because both pursue differentiated device silicon. Apple tightly coordinates its processors, operating systems, and hardware, while Google uses Tensor chips to support Pixel-specific experiences.

Their strategies show why device companies value control over silicon. Custom chips can prioritize features that a broadly sold processor cannot optimize for every customer.

They also show the scale of the undertaking. Custom silicon needs long-term software investment, repeated product generations, and enough devices to amortize development costs.

The reported Samsung-Arm model tries to lower that entry barrier. Arm supplies more of the foundation, and Samsung supplies design execution plus manufacturing.

If it works, more device companies could consider custom processors without copying Apple's organization. If it fails, merchant platforms will retain their advantage in predictability and shared development costs.

The 2-Nanometer Plan Is Also the Biggest Risk

The project's strongest strategic feature, Samsung's control from design through fabrication, also concentrates its execution risk.

The original report assigns mass production to Samsung Foundry's 2-nanometer process. It provides no manufacturing date, process variant, target yield, or expected volume.

A process label alone does not establish performance. Final results depend on transistor characteristics, design libraries, memory choices, physical layout, packaging, cooling, and software behavior.

Samsung began producing chips with its first-generation 3-nanometer GAA technology in 2022. Its earlier Arm collaboration sought to optimize future Cortex processors for newer GAA nodes.

That history makes the reported 2-nanometer plan technically plausible. It does not verify that a customer design has passed validation or reached a production-ready stage.

Advanced chip projects routinely move through architecture, RTL development, verification, physical implementation, tape-out, fabrication, and silicon validation. Problems at any stage can force revisions.

The NRE approval cited by the report belongs near the beginning of that sequence. It indicates that someone has authorized engineering work, assuming the sourcing is correct.

It does not reveal whether the parties have frozen the specification. It also does not show whether the customer has committed to production volume after initial samples arrive.

The unnamed customer increases uncertainty. A smartphone vendor, wearable maker, personal-computer company, automotive supplier, or another device business would impose different requirements.

The phrase "on-device AI" covers all those markets. Each has different power limits, safety requirements, software stacks, product cycles, and acceptable costs.

The absence of a named product also prevents meaningful competitive comparison. There is no defensible basis for declaring the chip faster than Snapdragon, Dimensity, Tensor, or Apple's processors.

The reporting chain requires equal care. The rsshub 36kr entry points to a news summary, which itself rests on a Korean report citing semiconductor industry sources.

Multiple republications do not create multiple confirmations. They can trace back to the same unidentified sources and should be treated as one underlying claim.

Samsung and Arm have issued detailed announcements about earlier collaborations. Their silence on this specific program might reflect customer confidentiality, an early development stage, or an inaccurate report.

Each explanation remains possible. A customer-specific chip program often includes contractual restrictions, so the absence of public confirmation does not prove the report false.

Readers should also resist interpreting "Arm funding" as a disclosed investment. The report describes approved engineering expenses, but it does not state the amount or publish contractual terms.

Commercial responsibility remains another open question. Samsung would reportedly receive payment from the final customer, yet the allocation of design risk, intellectual property rights, and warranty obligations is unknown.

The success test is therefore concrete. The parties need working silicon that meets performance, power, area, cost, and schedule requirements.

Until those results appear, the story is about a reported development model rather than a proven product advantage.

Three Signals Will Show Whether the Project Is Real

A named customer, a silicon milestone, and a software path will separate a serious product program from an early engineering experiment.

The first signal is an official acknowledgment from Samsung, Arm, or the customer. A confirmation would clarify the device category and show that the effort extends beyond preliminary technical evaluation.

The most useful announcement would identify the target market without relying on broad phrases such as edge intelligence. Smartphones, wearables, computers, and automotive systems demand different chip designs.

A named customer would also reveal the competitive stakes. An established device manufacturer could provide the volume needed to support an advanced custom process.

The second signal is a measurable silicon milestone. Watch for a tape-out, sample shipment, foundry production announcement, certification record, or device launch tied to Samsung's 2-nanometer process.

Such a milestone would strengthen the report because it requires architecture and design work to progress toward physical hardware. Continued silence would weaken near-term production expectations.

Benchmarks will matter only if they disclose workloads and testing conditions. Useful comparisons should include latency, energy consumption, sustained performance, model size, memory use, and thermal behavior.

Peak operations-per-second figures alone cannot describe a user's experience. A processor may reach a high burst score yet throttle during longer workloads.

The third signal is developer support. Arm or Samsung would need to expose compilers, runtimes, optimized operators, model-conversion tools, or documentation for the accelerator.

Software evidence can appear before a consumer product. Framework integrations, development boards, technical sessions, and partner demonstrations often reveal which workloads a new platform prioritizes.

Developers should watch for support across text, image, audio, camera, and multimodal models. They should also examine whether optimization applies broadly or only to selected demonstrations.

Enterprise buyers have a different concern. Local inference can reduce cloud dependence, but deployment teams still need security controls, model governance, update mechanisms, and reliable data handling.

Knowledge workers will notice the change only when local AI improves real tasks. Faster transcription, private document retrieval, offline assistance, and responsive media processing are more meaningful than architecture labels.

Teams assessing those workflows should record vendor claims alongside later measurements. A searchable technical knowledge base can help connect announcements, specifications, tests, and deployment findings.

For now, the cautious judgment is straightforward. The reported project fits the public direction of both companies, but its defining details remain unverified.

Samsung wants more valuable custom-silicon and foundry work. Arm wants its technology embedded deeper within complete computing platforms. Device companies want local AI without building every semiconductor capability internally.

Those incentives make the partnership credible. They do not guarantee that this particular chip will reach production, meet its targets, or win a major customer.

The next useful rsshub 36kr update would not be another anonymous confirmation. It would be a named product, working sample, software release, or manufacturing milestone.

Until then, readers should treat the Samsung-Arm chip as an active reported program with a plausible technical foundation and a substantial verification gap.