Arm at the Edge – Apple’s AI Paradox
Arm Holdings (NASDAQ: ARM) has become indispensable to the AI economy.
It does not build chips: it defines the blueprints that everyone else follows. Its instructions set architecture (the language that defines which tasks the processor can undertake) underpins nearly every smartphone. Its Neoverse platform now powers a growing share of data-center central processing units (CPUs). The next frontier for AI, however, lies at the edge—where intelligence runs locally on the device itself.
It was to strengthen Arm’s foothold on the edge that ARM recently expanded access to its licensing program to include the Armv9 edge-AI platform. The move lowers barriers for startups and hardware developers. So far, more than 300 companies have joined the program, with over 400 chip designs completed and ready for manufacturing.
The stock leapt on the news as it renewed investor confidence that Arm’s royalty growth will accelerate as AI computation increasingly moves from cloud data centers to consumer devices.
Yet ravenous as consumers may be AI for at the edge, not all the usual suspects are delivering equally. Apple’s (AAPL) iPhone 17 is breaking sales records, yet its on-device AI has drawn tepid reviews compared with Samsung’s (SSNLF) Galaxy AI or Google’s (GOOGL) Pixel 9. How can Apple, whose processors use the same Arm instruction set as its competitors, deliver such uneven results? The answer reveals the distinction between architecture and implementation—and why Arm’s business remains insulated from Apple’s design trade-offs.
Arm and Apple
Apple’s reliance on Arm dates to 1990, when Apple, Acorn, and VLSI Technology co-founded Advanced RISC Machines to develop low-power processors. RISC stands for Reduced Instruction Set Architecture. Reducing the instruction set—simplifying the language—reduces the power required by the processor, just what’s needed for say iPhones or smart watches.
The first Arm-based iPhones arrived in 2007, with the A4 chip designed around Arm architecture. Over the next decade, Apple’s A-series chips (through A18 today) maintained that foundation while introducing custom CPU and Neural Engine designs. In 2020, Apple extended the partnership to its Mac line by replacing Intel’s x86 processors with Arm-based M1 silicon. In 2023, the companies renewed their licensing agreement through at least 2040, ensuring that every future iPhone, iPad, and Mac remains built on Arm IP.
Apple’s entire hardware stack is Arm-based, yet its AI results vary widely. That’s because Apple designs its own cores—the individual processing units that speed computation by working in parallel—and its own neural units—devoted to AI tasks. Apple then integrates these on top of Arm’s instruction set.
Apple’s Edge-AI Reality
Apple’s dominance in premium smartphones masks a technical paradox. The A18 chip inside the iPhone 17 is built on Arm architecture—the same foundation used by Qualcomm, Samsung, and Google—yet Apple’s on-device AI functions lag those rivals. The difference is intentional.
Arm defines the language of computation; Apple writes its own dialect through custom CPU, GPU, and Neural Engine designs. For the A18, Apple prioritized efficiency and thermal balance over raw neural throughput. The result is a Neural Engine delivering about 35 TOPS (trillion operations per second). That’s sufficient for photo classification and voice tasks but below the 45–60 TOPS range in Snapdragon 8 Gen 3 and Exynos 2400 chips. Those competitors designed their NPUs for massive transformer-scale workloads that can address Large Language Model tasks such as text generation and translation. Apple tuned its hardware for deterministic tasks within a tight power budget.
Software strategy compounds the gap. Apple’s Neural Engine remains largely closed to third-party developers. Samsung and Qualcomm, by contrast, use Arm’s open ecosystem to run generative models directly on the device. Apple’s emphasis on privacy routes complex inference through its Private Cloud Compute infrastructure built on M-series servers, trading speed for data security.
The same instruction set powers both Apple’s restraint and Samsung’s aggressiveness. The divergence comes from each company’s allocation of die area, thermal budget, and developer openness.
The generational gap is widening. In October 2025, Samsung reported its upcoming Exynos 2600 achieved more than six-times higher neural-processing performance than Apple’s in-house A19 Pro during internal tests. Samsung’s implementation demonstrates how aggressive NPU scaling and transformer-specific optimization can transform on-device AI into a differentiating feature.
The same underlying Arm architecture now powers two radically different outcomes: Apple’s restrained, privacy-oriented design and Samsung’s high-throughput, generative-AI-ready SoC. Arm captures royalties from both sides of this divergence—Apple’s volume and Samsung’s performance—as edge-AI workloads grow.
Implications for Arm and Edge AI
Apple’s conservative strategy underscores Arm’s strength. Even with modest neural performance, every iPhone 17 sold adds to Arm’s royalty base. Meanwhile, Android OEMs using Arm’s latest AI-optimized cores generate higher-margin royalties from performance expansion. The result is a balanced model where Arm benefits from both volume (Apple) and innovation (Android and others).
As AI migrates to edge devices through 2026-2027, all roads lead to Arm.
Financial Outlook
As shown in Table 1, Arm’s valuation reflects high expectations but still-developing realization of its AI-driven royalties. Earnings growth is steady yet still catching up to its backlog of new licenses. Margin expansion depends on 2026-era deployment cycles.
Investor Takeaway
Whether manufacturers optimize for efficiency like Apple or for speed like Samsung and Qualcomm, the royalty stream flows back to Arm. The latest move above the 200-day average confirms market recognition that Arm is emerging as the architectural standard for on-device AI. Its strength lies not in building chips, but in building the language that makes intelligence possible.