MediaTek Claims the 2nm Crown: Dimensity 9600 Pro Runs 30B-Parameter Models On Your Phone

The race to 2 nanometers has a new leader, and for once it is not Apple, Qualcomm or Samsung. On September 15, 2026, MediaTek — the Taiwanese chipmaker better known for powering mid-range phones than winning spec-sheet wars — announced the Dimensity 9600 Pro, the industry’s first flagship smartphone system-on-chip built on TSMC’s 2nm-class N2P process. Alongside it came a sibling, the Dimensity 9600M, built on a 3nm node to carry flagship features into cheaper devices.

The move matters far beyond process-node bragging rights. It is the clearest signal yet that the smartphone industry’s center of gravity has shifted from cameras and displays to something else entirely: running serious AI models locally, around the clock, without draining the battery or shipping your data to a data center.

First to 2nm — and What That Actually Buys

MediaTek describes itself as the world’s number one mobile chipset provider by volume, and it says it is the first company to announce reaching the 2nm threshold in a phone chip. The claim rests on a partnership sealed a year earlier: on September 16, 2025, MediaTek said it was among the first to work with TSMC on the foundry’s enhanced N2P process and had already taped out its flagship SoC, with volume production expected in late 2026.

TSMC’s 2nm technology is the first in the foundry’s history to adopt a nanosheet transistor structure, abandoning the FinFET that has carried the industry since 2011. On paper, N2P delivers up to 18% higher performance at the same power, roughly 36% lower power at the same speed, and 1.2x the logic density of the previous N3E generation. Those are the abstract numbers. The concrete ones MediaTek disclosed for the 9600 Pro are more striking: up to 17% higher single-core performance and 15% higher multi-core performance versus the previous generation — alongside a 61% reduction in multi-core power consumption. (A footnote notes these figures come from demo device testing in MediaTek’s labs.)

That 61% figure is the headline for a simple reason: on-device AI is a sustained workload, not a burst. An agent that watches your calendar, your messages and your photos all day cannot rely on a performance spike; it needs efficiency at steady state. That is precisely what a 2nm node plus a redesigned power architecture buys.

All Big Cores, Again — but Faster

The CPU follows the now-familiar “all big core” philosophy that MediaTek helped popularize, in a 2+3+3 arrangement:

• 2x C2-Ultra cores at up to 4.55 GHz
• 3x C2-Pro cores at up to 4.35 GHz
• 3x C2-Pro cores at up to 3.1 GHz

There is no little efficiency cluster at all — every core is a big core, and the scheduling software decides which tier handles which task. Feeding them is a 34.5 MB system cache, and for the first time in a MediaTek flagship, support for LPDDR6 memory and UFS 5.0 storage. The memory jump matters for AI workloads specifically: LLM inference is bandwidth-hungry, and LPDDR6’s higher throughput directly translates into faster token generation on device.

An upgraded Dimensity Scheduling Engine coordinates hardware and software to allocate compute, power and memory dynamically — an approach MediaTek says lets manufacturers keep multiple apps alive in the background without the battery penalty that traditionally implies.

The Real Story: A Dual-NPU Architecture Built for Agents

If the process node is the foundation, the AI subsystem is the point. The Dimensity 9600 Pro carries a dual-NPU architecture paired with what MediaTek calls an Agentic AI Engine — hardware and software designed together for on-device AI “that perceives, reasons, and acts.”

The numbers MediaTek disclosed:

• A second-generation “Super Efficient NPU” cuts power consumption by 40% for always-on AI workloads
• The NPU 1090 delivers 51% higher LLM prefill performance (the phase that reads and digests your prompt) and 55% higher token generation per watt
• The platform supports on-device models of up to 30 billion parameters

Thirty billion parameters is a meaningful threshold. A year ago, flagship phones were showing off 7B-class models running locally. Crossing into 30B territory means a modern compact MoE model — the class of model that handles tool use, multi-step reasoning and code — can plausibly live entirely on a phone. MediaTek’s marketing frames the result as a “digital twin”: an agent that learns your preferences, analyzes on-device data and proactively assists, without your personal context ever leaving the device.

An “AI Computing Fusion Architecture” links the CPU and NPU so workloads flow between them, rather than treating the NPU as an isolated accelerator for camera tricks. That integration is what separates this generation from the “AI-powered” chips of two years ago, where the NPU mostly handled background blur.

GPU and Imaging: Still Pushing, Just No Longer the Pitch

The G2-Ultra NX GPU is no afterthought: MediaTek claims up to 27% higher peak performance, 24% lower power at peak, 18% faster ray tracing and on-device gaming up to 185fps. A new Dimensity Neural Graphics Architecture fuses GPU and AI compute, with HyperEngine tuning the experience.

Imaging gets a similar treatment. The Imagiq 1290 ISP pairs with a new AI Imaging Fusion Architecture that merges ISP and NPU capabilities: NPU-powered 60fps motion capture, native 4K240 ultra-slow-motion direct output (a first for the Dimensity line), 4K60 AI True Color Tone, 4K120 cinematic log recording and 17EV dynamic range. It is a competitive spec sheet — but tellingly, it reads as table stakes now. The order of the announcement tells the story: AI first, node second, GPU and camera after.

The 9600M and the Market Chessboard

The 3nm Dimensity 9600M brings the Agentic AI Engine, the Scheduling Engine, always-on perception and 185fps gaming to a lower tier, with first devices expected in the same window. MediaTek said the first smartphones powered by both chips are expected to launch this quarter — meaning flagships and upper-midrange devices with genuinely local 30B-class AI could be in consumers’ hands before the end of 2026.

The competitive context is brutal. Qualcomm is fighting on two fronts — its long-standing Snapdragon flagship franchise and a newly expanded data center business, including a multi-generation Amazon deal worth up to $60 billion announced September 8. Apple is expected to bring its own 2nm silicon, with TSMC capacity tight. Samsung’s foundry struggles have pushed even its own Exynos plans toward TSMC. In that landscape, MediaTek being first to announce a 2nm phone SoC is both an engineering milestone and a positioning coup: it tells OEMs that the company best known for value silicon now intends to define the flagship AI experience.

Why On-Device AI Is the Whole Ballgame

There is a reason every number in this announcement bends toward efficiency-per-watt and memory bandwidth. The industry’s AI narrative in 2026 has been dominated by frontier labs, data centers and hundred-billion-dollar compute deals. But the economics run both ways: inference in a data center costs the provider money every single time you use it, while inference on your phone costs the provider nothing and works on a plane, in a tunnel, and — crucially — with your private data never leaving the device.

Agentic AI sharpens the argument. An agent that reads your messages, watches your screen and acts on your behalf is exactly the kind of workload users may refuse to route through someone else’s cloud. A phone that can host that agent locally, at 30B parameters, with 40% lower always-on power draw, converts a privacy argument into a hardware purchase decision.

MediaTek’s bet is that this is the spec consumers will actually feel — not the fourth camera lens. If the first 2nm phones ship this quarter and the on-device agent experience holds up in reviews, September 15 may be remembered as the day the smartphone AI race stopped being about cloud chatbots and started being about silicon.

Either way, the message to Qualcomm and Apple is unambiguous: the 2nm era opens with MediaTek holding the pole position.