Pixel 11’s 2nm Tensor G6 Runs On-Device AI 3.5× Faster

Google opened Made by Google 2026 in New York on Wednesday, and the headline spec of the Pixel 11 is a manufacturing node. The Tensor G6 is built on TSMC's N2 process, the commercial 2nm generation, and Google is spending almost all of the gain on inference rather than on raw application speed. The chip carries 50% more TPU compute and double the memory bandwidth of last year's G5, which Google says lets it process on-device AI tasks up to 3.5 times faster while drawing up to 3.5 times less energy.

Those two numbers do more work than the CPU figures printed next to them, which are modest by comparison: 15% quicker app launches and 25% faster web browsing. The energy half of the claim is the interesting one. On-device models have never been bottlenecked purely by how fast a phone can run them, but by how long it can run them before the battery and the thermal envelope force a hand-off to a data centre. Cutting the energy per task is what makes an assistant that watches context all day plausible on a handset. The G6 is paired with a new Gemini Nano generation, identified as Nano 4 in 9to5Google's roundup of the event, along with a custom accelerator for 4K bokeh video and a Titan M3 security chip Google positions against future quantum attacks.

The software layer is Gemini Intelligence, the Android agent platform Google first unveiled in May, now shipping on hardware designed around it. Task automation reaches more apps for orders and bookings, and Gemini can place phone calls to businesses to finish a job. Gboard picks up Rambler, a dictation mode that strips filler words and run-on structure and returns the intent as clean text. Magic Cue has been renamed Proactive Assistance and given a real interface: a bottom sheet that surfaces flight details, Maps listings, Calendar previews and comparisons in context. At a Glance becomes location-aware, pulling up loyalty cards inside a store and transit stops at a platform. Live Translate expands past conversation into podcasts and video, with real-time dubbing handled on-device by the G6.

Photography is where the extra TPU headroom is most visible. Magic Capture analyses roughly 500 frames to pick the best shot and can crop, straighten and unblur it without being asked. Camera Looks moves in the opposite direction, merging fewer frames for a deliberately less-processed result and rewriting colour lookup tables at the sensor level to imitate film stocks — Google ships six of them, including Digi, Editorial and Velvet. Instant Night Sight on the Pro models is up to 4.5 times faster than the previous generation, and Pro Stable Video kicks in for shaky handheld footage in bright light.

The lineup itself runs from the $899 Pixel 11 through the $1,099 Pixel 11 Pro and $1,299 Pixel 11 Pro XL to the $1,899 Pixel 11 Pro Fold, which is 10% lighter at 238g and 0.7mm thinner than its predecessor. Pro buyers get six months of AI Pro access bundled. The base Pixel 11 gains a 48MP wide sensor with 56% better light sensitivity and 25W Qi2.2 wireless charging; the Pro models add a 50MP wide sensor, 120x Pro Zoom and a 3,600-nit display. Pre-orders opened on Wednesday, with retail availability on 20 August.

Rounding out the event: a $399 Pixel Watch 5 on a Snapdragon W5 Gen 2 Accelerated with 3GB of RAM and 64GB of storage, generative AI watch faces, and blood pressure and insulin resistance trends. Google also described offline Gemini commands on the watch for timers, alarms and workouts, though early coverage disagrees on how much of that is live on day one, so it is worth checking against the shipping firmware rather than the keynote. A $29 UWB Pixel Tag arrives 11 November, and Pixel Buds Pro 2 pick up a Gemini voice control layer and 70-plus-language Live Translate in September.

Taken together the launch reads as Google betting that the next round of phone competition is fought on energy per inference rather than on benchmark scores. A 2nm node is an expensive way to buy that, and it only pays off if enough of the Gemini Intelligence surface actually stays on the device instead of quietly round-tripping to a server — which is the one thing a spec sheet cannot tell you.