Table of Contents
Custom edge-AI SoC — but no confirmation from either company
In sum – what we know:
- 2nm edge-AI SoC – Reports say the two companies are co-developing a custom on-device AI chip with Samsung Foundry’s 2nm GAA process as the production technology.
- Arm as design partner – Arm’s reported role goes beyond licensing: accelerator architecture, RTL design, and joint program management under a Limited Use License.
- Unconfirmed by both companies – The reporting traces to Korean sources, with no release from Samsung or Arm and no confirmed end customer.
Samsung Electronics and Arm have reportedly started work on a custom on-device AI chip planned for Samsung Foundry’s 2nm process, according to a run of trade reports. The reported trigger was Arm’s approval of initial non-recurring engineering funding in late August. NRE is the upfront engineering spend that gets a bespoke chip program off the ground, and approving it is a meaningful commitment — but it is not a production order, a volume guarantee, or a promise that anything ever tapes out.
Neither Samsung nor Arm has published a release describing this specific program, and the reporting traces back to Korean and digital-industry sources rather than to either company. Reports have speculated about OpenAI as a possible customer, but Samsung, Arm and OpenAI had not publicly confirmed that connection.
Arm has spent the past several years pushing further up the stack from CPU designs into compute subsystems and reference platforms, and Samsung has spent the same period trying to convince the market that its foundry can hold advanced-node business against TSMC. A jointly developed edge-AI SoC on Samsung 2nm would sit at the intersection of both ambitions.
What is reportedly being developed
The project is described as a custom system-on-chip for edge AI — processing AI workloads locally on the device instead of sending every request to a data center. Reports frame the likely applications as smartphones and other terminal devices, which is broad enough to cover a lot of ground and specific enough to suggest the target isn’t a server part. Beyond that framing, no product category has been established.
Samsung Foundry’s 2nm Gate-All-Around process is the reported production technology. GAA is a transistor architecture in which the gate wraps the channel on all sides, rather than on three sides as in a FinFET. The point is tighter control over current leakage, which chipmakers use to chase either more performance at the same power or the same performance at less power. How much of either you actually get depends on the final design, the standard-cell libraries, operating voltage, yield and the workload — not on the architecture alone.
It’s also worth being careful with the node label. “2nm” is a process-generation name, not a measurement of anything on the die, and it doesn’t translate one-for-one between foundries. Samsung’s 2nm competes commercially with TSMC’s N2, which entered volume production in late 2025 and has already booked the industry’s marquee customers, and with Intel’s 18A. Being at the same numbered node is not the same as being at the same density, power curve or yield.
Division of work
Arm’s reported contribution is the AI accelerator architecture and the register-transfer-level design technology behind it. RTL is the level at which a chip’s logic is actually described in code before it becomes a physical layout, so supplying RTL is a substantially deeper involvement than licensing an instruction set or a CPU core. Reports also put Arm in charge of design support, relaying end-customer requirements, and joint program and schedule management — which reads less like an IP vendor and more like a design partner.
Samsung System LSI, the company’s chip-design arm, would handle overall SoC design and integration around that Arm IP. Samsung Foundry would manufacture on 2nm, with reports also referring to advanced packaging, commercialization, final manufacturing and delivery. If that holds, Samsung is supplying design, fabrication, packaging and logistics for a part whose accelerator architecture comes from Arm, which is a fuller version of the one-stop pitch Samsung has been making to custom-silicon customers for years.
The reports describe the licensing framework as a Limited Use License, an arrangement that restricts how the licensed IP can be used. What the LUL actually covers here isn’t public. The contract terms, the commercial value, IP ownership, whether it’s exclusive, the packaging configuration and the end customer are all absent from the source material. Reading exclusivity into it — or treating it as evidence of a new Arm product line — goes past what anyone has said.
Why it matters
For Samsung, the appeal is obvious. System LSI and the foundry business have historically been pitched separately, and getting both into a single customer-specific program at the leading edge is exactly the kind of win Samsung needs. A working 2nm AI chip would function as a reference design, and reference designs are how foundries convert curiosity into orders. Samsung’s own 2026 quarterly materials point to added major-customer design wins and 2nm high-performance-computing engagements, though none of those disclosures name this chip or attach a production timetable to it.
Arm’s stake is more strategic than financial. Approving NRE for someone else’s SoC is a fairly aggressive way to spend money for a company whose business model is collecting royalties, and it suggests Arm wants its architecture sitting at the center of endpoint AI designs rather than adjacent to them. That puts it in more direct proximity to Qualcomm and MediaTek, both of which build their own NPUs, and to the vertically integrated approach Apple takes with the Neural Engine and Google takes with Tensor.
For device makers, the case for local inference is somewhat modest. Running a model on-device can cut latency, keep working when connectivity is poor, and keep some data off remote servers.
This is not the first Samsung–Arm work
In 2024, the two companies announced that Arm’s then-next-generation Cortex-X CPU had been optimized for Samsung Foundry’s latest GAA technology through design-technology co-optimization, a process where the chip design and the manufacturing process get tuned against each other rather than in sequence. That announcement also gestured at 2nm GAA plans for data-center custom silicon and an AI chiplet aimed at generative AI on mobile. Useful background, and not the same thing as the bespoke edge-AI SoC now being reported.
The relationship goes back considerably further than that. By Arm’s own account, its physical-IP and process work with Samsung Foundry started at 65nm and continued through 11LPP, 7LPP and 5LPE. That history establishes precedent — these two know how to work together at the process level — but it doesn’t validate any particular detail in the current reports.
