As reported by Bloomberg’s Mark Gurman (Power On).
Apple is reorganizing its entire silicon roadmap around AI neural performance — and the skipped generation reveals a strategic ambition that extends well beyond the Mac.
What Happened
Bloomberg’s Mark Gurman, reporting in Power On, reveals that Apple has made a move with no precedent in its silicon history: it is skipping the M6 Pro, M6 Max, and M6 Ultra entirely. The base M6 ships this fall, then Apple jumps a full tier and goes straight to M7. The reason, per Gurman’s reporting, is AI — planned neural-processing upgrades were deemed significant enough to collapse the roadmap and accelerate to a new generation rather than iterate within the existing one.
The pace is striking. Apple reportedly began taping out the M7 just six months after the M6, putting the M7 in the first half of 2027, M7 Pro and Max by end of 2027, and an M7 Ultra around 2028. That Ultra is the headline chip: Gurman reports its AI performance is described internally as moving closer to dedicated accelerators — with Nvidia’s Blackwell cited as the reference point. That is a relative claim, not a benchmark, and the roadmap is internal and multi-year; ship dates, configurations, and node targets (including an expected 1.4nm process for M8, code-named ‘Soko,’ around 2028) can all move.
The server dimension is where this stops being a Mac story. Apple already has an M5-Ultra-based AI server (code-named J246) coming soon. Behind it sits a planned M7-Ultra-based server targeting up to 1.5 terabytes of unified memory — roughly double the M5 Ultra — designed to power Apple Intelligence infrastructure at scale. Gurman notes the 1.5TB configuration is contingent on the memory market: the ongoing DRAM shortage has made high-bandwidth memory scarce and expensive, and even Apple’s data-center roadmap is gated by supply it does not control.
The key insight: Apple is not adding AI to its chip roadmap — AI is now the organizing principle that determines which chips get built, which get cancelled, and how fast the next generation ships. Skipping an entire high-end tier is the clearest signal a hardware company can send that the old priority stack (clock speed, GPU, battery, thinness) has been demoted beneath neural performance.
The Origin: A $10 Billion Failure That Built the Foundation
The AI-hardware stack Apple is now accelerating traces directly to its cancelled car project — a roughly decade-long, ~$10 billion effort aimed at Level-5 autonomous driving that Apple quietly wound down. The Level-5 autonomy work required custom silicon capable of real-time environmental inference at scale. That silicon became the Neural Engine, which shipped in the iPhone X in 2017 and has been in every Mac since the M1 in 2020.
The car never arrived. The Neural Engine is now the architectural foundation for Apple’s entire AI strategy — the thing that makes on-device Apple Intelligence possible, that differentiates its silicon from any x86 alternative, and that is now being extended into the data center. A $10 billion abandoned moonshot became the substrate of the next platform.
Bloomberg / Mark Gurman — Power On
“Apple began taping out the M7 just six months after the M6 — and the M7 Ultra is designed to bring AI performance closer to dedicated accelerators like Nvidia’s Blackwell.” [Reported, paraphrased per Gurman’s Power On newsletter, July 12 2026]
The Structural Read
On the Map of AI — the nine-layer stack from raw compute up through applications — Apple has historically been strongest at layers six through nine: devices, operating systems, developer platforms, and consumer applications. Its silicon gave it a durable moat at the device layer. What the M7 roadmap signals is a deliberate push downward into layers one through three: raw compute, accelerator silicon, and AI infrastructure. Apple is not just making chips for iPhones and Macs anymore. It is building the accelerator stack for its own AI cloud.
That repositioning has three structural dimensions worth unpacking.
Map of AI — Apple’s Layer Descent
From Device Maker to Accelerator Company
Apple has always owned layers 6-9 (devices, OS, platform, apps). The M7-Ultra-class server push is a deliberate expansion into layers 1-3 — compute, accelerators, infrastructure. The same vertical integration playbook that routed around Intel now routes around Nvidia, at least for Apple’s own workloads. This is not a feature. It is a structural repositioning that compounds over every subsequent chip generation.
Three Implications
IMPLICATION 1 — VERTICAL INTEGRATION, EXTENDED INTO THE DATA CENTER
Apple’s custom-silicon strategy — which routed around Intel for the Mac — is now routing around Nvidia for AI infrastructure, at least for Apple’s own workloads. An M7-Ultra-class server approaching Blackwell-class AI performance means Apple Intelligence runs on Apple accelerators from the device to the data center. The same pattern is playing out at Meta (MTIA) and Google (TPU). The Nvidia toll booth remains dominant for third-party AI training at scale, but every hyperscaler that builds its own inference accelerator is a workload that never touches an H100. Apple’s server roadmap makes that moat one step narrower. See also: Apple’s Baltra server chip and Nvidia’s toll-booth economics.
IMPLICATION 2 — THE MEMORY CHOKEPOINT IS UNIVERSAL, EVEN FOR APPLE
The 1.5TB server configuration is explicitly contingent on DRAM availability. That single sentence in Gurman’s reporting is worth sitting with: the world’s most valuable hardware company, with its own custom silicon, cannot guarantee the memory configuration of its own AI servers because of a commodity shortage it does not control. The DRAM scarcity taxing iPhone memory upgrades is the same scarcity compressing Apple’s data-center ambitions — and minting SK Hynix and Micron in the process. The chokepoint is not a supply-chain inconvenience; it is a structural ceiling on every AI roadmap simultaneously. See: The AI Memory Chokepoint.
IMPLICATION 3 — FAILURE AS SUBSTRATE: THE $10B CAR PROJECT IS EVERYWHERE
The cancelled Apple Car is the most expensive R&D write-off in consumer tech history that nobody talks about as a write-off. It produced the Neural Engine, which is now the architectural core of every Apple device and the competitive differentiator driving the M7 acceleration. The strategic lesson generalizes: abandoned moonshots do not disappear — they deposit capabilities into the organization that surface in the next platform cycle. The question for every company running speculative R&D is not just “did the product ship?” but “what does the failed attempt leave behind?” Apple’s answer, worth roughly a trillion dollars of differentiated silicon, is the Neural Engine.
