First 2nm phone chips from Qualcomm, Apple and MediaTek focus on AI

What happened

TrendForce, a Taiwanese market research firm, published a comparison of the three new chip families on 6 October 2026. All three are made by Taiwan Semiconductor Manufacturing Company (TSMC) on its 2 nanometre (2nm) process. TrendForce says this marks the industry's move to gate-all-around (GAA) nanosheet transistors. Each company puts AI first, but each takes a different approach to running it on the phone.

Qualcomm launched the Snapdragon 8 Elite Gen 6 and the higher-end Snapdragon 8 Elite Extreme Gen 6 at its Snapdragon Summit in Maui, Hawaii, which opened on 22 September, according to The Next Web. Qualcomm says the Extreme chip's processor is the first mobile CPU to reach 5 gigahertz. The company says its CPU is 13 percent faster than the last generation, its graphics processor 44 percent faster, and its neural processing unit (NPU) 35 percent faster.

Apple's A20 Pro, which powers the iPhone 18 Pro phones and the iPhone Duo, has a 32-core neural engine. Apple says it gives twice the AI processing power of the A19 Pro, with 50 percent more memory bandwidth. MediaTek's Dimensity 9600 Pro uses two NPUs. MediaTek says its efficient NPU cuts power use by 40 percent for always-on AI tasks, and the chip supports on-device models of up to 30 billion parameters.

The engineering behind it

A transistor is a switch that controls current flowing through a channel. In older planar transistors, the gate that controls the switch sits only on top of the channel. In FinFET transistors, used for about a decade, the channel stands up like a fin and the gate wraps three sides. In gate-all-around nanosheet transistors, the channel is made of thin flat sheets stacked on top of each other, and the gate surrounds each sheet on all four sides.

Wrapping the gate fully around the channel gives better control over the current. This is general device physics: the better the gate controls the channel, the less current leaks when the switch is off. Less leakage means less wasted power and heat. That matters for a phone, which has a small battery and no fan. The name 2nm is a label for a generation of manufacturing process, not the exact size of any one part.

An NPU is a part of the chip designed for the maths of neural networks, mainly large numbers of multiplications and additions. It does this work with far less energy than a general CPU. Running a large AI model also needs memory bandwidth, which is how fast data moves between memory and the processor. Apple's design places memory beside the chip, out of its heat path, and connects the chip to a vapour chamber to spread heat, the company says.

Qualcomm says the Extreme chip can run mixture-of-experts models with more than 30 billion parameters, and handle up to 32,000 tokens of context. In a mixture-of-experts model, only part of the network runs for each piece of text, which saves memory and power. To fit a model on a phone, developers also use quantisation, which stores each number with fewer bits. These are general techniques and not specific to these chips.

Demand and early tests

TrendForce, citing Economic Daily News, says Apple, Nvidia, AMD, Qualcomm and MediaTek have raised their bookings for TSMC's 2nm production by 10 to 20 percent. Industry sources expect TSMC's 2nm capacity to reach about 120,000 wafers a month by the end of 2026, above earlier estimates of 90,000 to 100,000. A wafer is the thin round slice of silicon on which many chips are made at once.

The Next Web reports that a Gizmodo reviewer tested Qualcomm's reference phones at the summit. They beat Apple's A20 Pro by 8 percent on a multi-core benchmark and by about 20 percent on a graphics test. Apple kept a lead of about 10 percent on single-core scores, and the test phones ran hot. The Next Web also notes that Counterpoint Research expects smartphone shipments to fall 14 percent in 2026, as memory shortages push up prices.

What it means in Nepal

The sources do not discuss Nepal, so this section is about skills rather than the local market. Running AI on the phone instead of on a remote server changes how apps are built. Data can stay on the device, and the app can work without a fast internet connection. But the developer must fit the model into limited memory, power and heat. Skills such as model compression, quantisation and measuring energy use become part of app development.

The chips also show why device physics and computer architecture are useful even for engineers who will never work in a chip factory. A phone chip is a set of choices about transistors, cores, memory paths and heat. Engineers who design embedded products, write performance-critical software or choose hardware for a project make better decisions when they understand those trade-offs. Reading launch claims carefully, and waiting for independent tests, is part of that skill.