Apple's first 2nm phone chip doubles the Neural Engine after three flat years
Apple also widened memory bandwidth by half, moved the DRAM off the thermal path and tripled the vapor chamber, and every one of those decisions is about running models on the phone.
Apple spent its keynote on a Siri that is still in beta, in English, outside the European Union. The commitment is in the silicon.
The A20 Pro, announced on 9 September 2026 for the iPhone 18 Pro, the Pro Max and the iPhone Duo, is the first smartphone processor built on a two-nanometer process. Apple's first chip on that node was the M6, announced in August for a Mac mini. This is the first one in a phone.
The headline claims are ordinary for an Apple September: a six-core CPU with two performance "super cores" up to a fifth faster than last year's, and a seven-core graphics processor up to 40 percent faster. Those went on the big slide. The numbers that tell you what the chip is for were smaller.
Three changes, all of them about running a model locally
The Neural Engine doubled. Apple had held it at 16 cores for three consecutive generations. The A20 Pro carries a "Dual 16-core Neural Engine … 32 total cores for double the AI processing power of A19 Pro", and Tom's Hardware adds that the graphics processor's neural accelerators give twice-as-fast FP8 compute — the numeric precision most on-device language models actually run at.
Memory bandwidth is up by half, which Apple calls "the widest memory interface in an iPhone". Why does that matter more than the CPU figure? Because for a model running on your phone, bandwidth rather than compute is usually the ceiling on how fast words come out. Every generated token re-reads the weights.
And the packaging is now Mac-style. Apple placed the die beside the memory instead of stacking it, "removing the memory from the thermal path of the chip and allowing A20 Pro to attach directly to a next-generation vapor chamber", which has three times the surface area of the one in the 17 Pro. The claim that follows is up to 40 percent better sustained performance, and double that of the 16 Pro. Sustained, not peak, is what matters for a model that runs for the length of a conversation rather than the length of a benchmark.
There is also an in-house C2 modem replacing Qualcomm across the line, using 15 percent less energy than its predecessor and adding mmWave in the United States.
The workload Apple names, and the one it does not
Apple's release says the new Siri runs on "on-device processing and Private Cloud Compute". The on-device half is what the doubled Neural Engine buys. Decrypt reports the heavy half runs on Google's Gemini in the cloud, which Apple's release does not mention.
A company that doubles its neural processor after three flat generations, re-plumbs its packaging around a vapor chamber and spends a slide on memory bandwidth is building a phone for a model that does not yet exist in shipping form. The Siri demo is the placeholder.
The comparison to watch is not the next Snapdragon. It is Apple's own Mac silicon: the A20 Pro's packaging is "inspired by M-series Apple silicon", and Tom's Hardware reports Apple will skip its high-end M6 Macs to fast-track an AI-focused generation for 2027. We would expect Apple to quote a tokens-per-second figure for an on-device model on stage by next September's event, the way it quotes video-playback hours now.
Until it does, the doubled core count is a promise. And anyone selling a dedicated AI wearable on the premise that your phone cannot do this locally should read the spec sheet twice.
