MediaTek put the two-nanometre AI race in a phone

01

WHAT ACTUALLY CHANGED

MediaTek announced the Dimensity 9600 Pro in Taiwan on September 15. Reuters identifies it as the company's first smartphone chip manufactured with TSMC's two-nanometre process. MediaTek also announced the Dimensity 9600M, made on a three-nanometre process for a broader part of the flagship market. The company says the first phones using the two chips will arrive soon, but it has not named the complete device lineup, launch dates, prices, regions, or shipment volumes.

The Pro chip includes a dedicated neural processing unit for AI workloads. MediaTek describes a dual-NPU design and an Agentic AI Engine intended to run more complex generative and assistant tasks on the handset. The announcement establishes the presence and intended role of that hardware. It does not establish which models, apps, languages, context sizes, or agent actions will work locally in a retail phone.

MediaTek says the NPU is 51 percent faster than the previous generation at processing a user's prompt before a model begins producing its response. Reuters reports that the company gave no further detail for the comparison. The public materials do not identify the model, prompt set, precision, memory configuration, software version, power limit, temperature, run length, or whether the result represents a peak or sustained measurement.

Two nanometres is the name of a manufacturing generation, not a promise that every important feature in the chip literally measures two nanometres. Newer process technology can let designers pursue better performance, lower power use, greater density, or some negotiated combination. The phone maker still chooses clock speeds, memory, cooling, battery capacity, software, radios, cameras, and the performance profile that customers eventually feel.

The launch places local AI hardware closer to the centre of flagship-phone competition. Reuters says MediaTek is trying to gain ground against Qualcomm in premium handsets while component costs and AI features push prices upward. That market contest matters, but a chip launch is not a completed user experience. The evidence arrives when named phones ship and independent reviewers can repeat the same work across devices.

02

WHY THIS MATTERS

Local computation can remove a round trip to the cloud. That can shorten response time, preserve a feature when connectivity is weak, and avoid transmitting some raw audio, images, documents, or sensor data. The word can is doing important work. An app may still send telemetry, retrieve a larger model, synchronize history, call an outside service, or upload the original material. Hardware capability and product data policy are separate controls.

A phone is a tight thermal budget wearing a glass shirt. A benchmark may look excellent during a short burst while a longer transcription, image generation, or agent task heats the device and forces it to slow down. Useful evaluation should show both the first run and repeated runs, with battery drain, surface temperature, latency, output quality, memory pressure, and network activity measured together.

On-device AI can widen access when features remain useful without fast or affordable mobile data. Live captions, translation, reading support, photo description, note summarization, and voice control can become more dependable on a train, in a rural area, during travel, or inside a building with poor reception. Access still depends on which phones receive the chip, how much they cost, which languages are supported, and how long software updates continue.

More capable local hardware also increases the need for clear permissions. An assistant that can read the screen, hear a request, inspect files, and take actions should expose what it accessed, what left the device, which step used a cloud service, and what the user can undo. A tiny cloud icon in a settings menu is not an adequate record for a system acting across personal information.

The industry needs benchmarks that resemble a day with a phone, not a drag race between isolated components. A good comparison would use the same models and tasks, include offline and connected modes, measure quality as well as speed, repeat runs until temperature stabilizes, and publish the power and software settings. That makes the chip easier to understand and makes marketing much harder to hide behind.

03

WHERE IT COULD HELP

• Run live captions, transcription, translation, image description, photo cleanup, search, and short assistant tasks locally when the model and app disclose that the input remains on the phone
• Build an explicit local-first mode that works without a connection and asks before sending a task, file, recording, or conversation to a cloud model
• Show a per-task receipt identifying which model ran, whether it ran locally or remotely, what data was transferred, how long it was retained, and which action can be reversed
• Test retail phones with repeated real tasks while recording response quality, time to first output, total completion time, battery drain, surface temperature, memory use, and network traffic
• Support local models, accessibility features, and language packs through several years of operating-system and security updates instead of limiting the useful hardware to a launch-season demo