We unfolded the iPhone Duo

What It Covers

The Vergecast covers Apple's first event under new CEO John Ternus, examining the iPhone Duo foldable ($1,999–$3,199), iPhone 18 Pro camera upgrades, Apple Watch Series 12 health features including ambient audio recording, and a deep dive into whether Apple Silicon Macs genuinely outperform NVIDIA GPUs for running local AI models at home.

Key Questions Answered

• •iPhone Duo software strategy: Apple limits the Duo to two simultaneous apps maximum, deliberately avoiding the complexity of Samsung's four-app multitasking. This "foldable on training wheels" approach prioritizes seamless continuity between folded and unfolded states over raw capability, targeting iPad-familiar users rather than power users. Developers should expect multitasking complexity to expand incrementally across future versions, mirroring the iPad's gradual multitasking evolution over several years.
• •iPhone 18 Pro variable aperture trade-off: The new main camera shifts from a fixed aperture to a variable f/1.48–f/4 range. The brighter f/1.48 maximum benefits low-light photography, but stopping down to f/4 on a physically tiny sensor produces full-frame equivalent depth-of-field around f/7, making the variable aperture most meaningful when paired with the new manual pro controls for shutter speed, white balance, and ISO rather than casual everyday shooting.
• •Apple Watch heart rate monitoring upgrade: The Series 12 increases heart rate sampling from once every five minutes during daily wear to once every five seconds — a 60x frequency increase — enabled by a redesigned sensor array with 32 electrode spokes replacing the previous two. This unlocks recovery heart rate variability (HRV) tracking and a readiness score, features Oura Ring, Whoop, and Garmin have offered for years, representing catch-up functionality with Apple's characteristically conservative, holistic presentation.
• •Apple Watch ambient audio privacy architecture: The new Siri Recap feature uses a dedicated secure exclave chip to process ambient audio locally, auto-deleting raw audio immediately after processing and purging summaries after seven days. It requires manual opt-in, produces speaker-attribution-free summaries, and triggers an audible chime plus visual indicator when active. Understanding this architecture matters because bystanders cannot verify a user's configuration, creating a social trust gap regardless of Apple's technical privacy guarantees.
• •Local AI: Mac vs. NVIDIA GPU decision framework: For models under 24GB, an RTX 6000 Pro runs significantly faster than a Mac Studio. For models exceeding 24GB, NVIDIA cards force inference into slow system RAM, making high-unified-memory Macs (up to 512GB) the only practical consumer option. A used NVIDIA P40 card at roughly $150 delivers competitive inference performance for models that fit within its 24GB VRAM, making it the cost-efficient entry point before justifying Mac Studio pricing.
• •Local AI economics for individuals vs. companies: Cloud inference offers lower per-token unit costs due to subsidized infrastructure, making it rational for startups. For individuals running models continuously at home, a one-time hardware purchase plus electricity costs produces dramatically lower long-term costs than $20/month subscription services. Open-source models on platforms like OpenRouter have flipped from 40% to over 60% of market token usage in six months, signaling that frontier lab economics remain structurally unsustainable against commoditized inference alternatives.