Apple's silicon ambitions just got a whole lot bigger. According to Bloomberg's Mark Gurman, writing in his Power On newsletter, Apple is designing its upcoming M7 Ultra chip to support as much as 1.5 terabytes of unified memory — a figure that would more than triple the 512GB ceiling of the current M3 Ultra and roughly double the capacity planned for the M5 Ultra. If Apple delivers on this spec, the M7 Ultra won't just be another Mac chip upgrade. It could be the foundation of a new era of on-device AI computing.
What Is 1.5TB of Unified Memory, and Why Does It Matter?
Unified memory sits at the heart of what makes Apple Silicon different. Rather than keeping CPU and GPU memory separate, Apple fuses everything onto a single die — giving every core lightning-fast access to the same pool of memory with no data copying between chips. The trade-off has always been capacity: soldering memory onto the processor limits how much you can pack in.
The M7 Ultra appears set to break that constraint decisively. At 1.5TB, the chip would match the maximum RAM configuration Apple offered on the 2019 Intel Mac Pro — a machine famous for its upgrade ceiling — while delivering the bandwidth advantages of Apple's unified architecture. For context, that's 1,536GB of memory, accessible simultaneously by every processor core on the chip.
Why does raw memory capacity matter so much for AI? Because the limiting factor for running large language models locally isn't CPU speed — it's how much model data can fit in memory at once. Today's frontier models can require hundreds of gigabytes just to load their parameters. A machine with 1.5TB of unified memory could run models that currently require dedicated cloud infrastructure, entirely on-device, with no latency and no data leaving the user's machine.
Developers already exploring creative ways to push large models onto consumer hardware will recognise what this unlocks. We recently covered how colibri runs a 744B-parameter MoE model on just 25GB of RAM — an impressive feat of compression engineering. The M7 Ultra would make feats like that entirely mainstream.
Apple's Accelerated Chip Roadmap: Skipping M6 Pro, Max, and Ultra
The M7 Ultra announcement is embedded in a broader — and remarkable — shift in Apple's silicon strategy. According to Gurman, Apple is skipping the M6 Pro, M6 Max, and M6 Ultra entirely, something that has no precedent in Apple Silicon history. The company will ship a base M6 chip later in 2026, then jump directly to the full M7 family.
The reason, per Gurman's reporting, is AI. The Neural Engine upgrades planned for the M7 generation were deemed significant enough to accelerate the entire roadmap rather than iterate within the existing M6 cycle. Apple reportedly began taping out the M7 just six months after completing the M6 — an unusually compressed development window.
Here's how the reported timeline looks:
- M6 (base only) — Late 2026, expected in a redesigned MacBook Pro
- M5 Ultra — Late 2026, up to 768GB unified memory, for a refreshed Mac Studio
- M7 (base) — First half of 2027, ~240GB/s memory bandwidth
- M7 Pro & M7 Max — Late 2027
- M7 Ultra — 2028, up to 1.5TB unified memory, AI server-class performance
- M8 (codenamed "Soko") — 2028, targeting a 1.4nm manufacturing process
The AI Ambition: Approaching Nvidia Blackwell Territory
Gurman's report is direct: the M7 Ultra is being internally described as bringing AI performance closer to dedicated accelerators like Nvidia's Blackwell. That's a relative benchmark, not a hard specification, and Apple hasn't published any comparison figures. But the direction of travel is unmistakable.
Apple is clearly positioning the M7 Ultra not just as a Mac chip, but as an AI accelerator capable of competing in workloads traditionally reserved for discrete GPU clusters. The memory bandwidth on the base M7 is already expected to hit ~240GB/s, a 56% increase over the M5 and a critical metric for how quickly large models can process tokens.
This ambition is inseparable from the AI chip supply story playing out across the industry. TSMC — the foundry manufacturing Apple Silicon — recently reported a 36% revenue surge, driven almost entirely by AI chip demand. We analysed what that means in our piece on TSMC's AI chip demand momentum. Apple's aggressive roadmap is a direct response to the same pressures driving hyperscaler spending.
Project Titan's Second Act: The Neural Engine's Origin Story
There's a fascinating lineage behind the M7 Ultra's Neural Engine. Apple's cancelled self-driving car project — Project Titan — consumed over $10 billion across roughly a decade before being shut down in 2024. But the machine learning and custom silicon work developed for an autonomous vehicle didn't disappear. It became the foundation for the Neural Engine.
A car that drives itself needs to process camera and sensor data, identify hazards, and make split-second decisions entirely on-device, with no cloud dependency. Those exact requirements — local inference, low latency, massive throughput — are what Apple is now building into its consumer and server chips. The abandoned moonshot became the substrate of Apple's AI strategy.
Apple's Server Play: Beyond the Mac
One of the most significant details in Gurman's report is that the M7 Ultra isn't just destined for Mac Studios. Apple is reportedly developing a dedicated AI server product built around the chip, targeting a 2029 launch. Before that, an M5 Ultra-based server (codename: J246) is already in the pipeline to power Apple Intelligence infrastructure.
This is Apple moving into territory historically dominated by Nvidia, AMD, and Intel. A 1.5TB unified memory ceiling makes the M7 Ultra server-competitive in ways previous Apple Silicon chips simply weren't — able to hold entire large model weights in-memory without offloading to slower storage.
The parallel with distributed inference approaches is worth noting. Tools like Mesh LLM, which pools GPU memory across machines, exist precisely because no single device has had enough memory to run frontier models alone. Apple's bet is that its silicon can eventually make that workaround unnecessary — at least for a certain class of workload.
The Caveat: Memory Shortages Could Limit the Top Tier
Gurman adds an important qualifier: whether Apple ships a 1.5TB configuration will depend on the state of the memory market. High-bandwidth DRAM is currently scarce and expensive — SK Hynix's CEO has warned that 2027 could see the worst memory shortage on record. Apple itself has already been forced to limit the current Mac Studio to a 96GB ceiling, down sharply from the 512GB that was available during the M3 era.
Even if Apple engineers design for 1.5TB, supply chain constraints could force a lower ceiling at launch. And the pricing would be steep either way — based on Apple's current rate of roughly $25 per additional gigabyte, upgrading from a base 128GB configuration to 1.5TB would cost well over $35,000. This is workstation and server territory, not consumer hardware.
The broader AI infrastructure investment thesis remains strong regardless. As we noted in our coverage of AI compute demand showing no signs of slowing, the appetite for memory-heavy hardware among enterprises and researchers continues to grow even as chip stocks face volatility.
What This Means for AI Developers and Pro Users
The practical implications of 1.5TB of unified memory are significant for several audiences:
- AI researchers who want to run frontier-class models locally, without cloud API costs or data privacy concerns
- Professional video editors and 3D artists working with assets that overwhelm current Mac memory limits
- Enterprise teams looking for private, on-premises inference that doesn't require a rack of Nvidia GPUs
- Developers building Apple Intelligence features, who will gain access to more capable on-device model infrastructure
The M7 Ultra is still roughly two years away. Between now and 2028, the M5 Ultra with its 768GB ceiling and M7 Pro and Max chips will keep Apple's high-end lineup moving forward. But the direction is clear: Apple is making a deliberate, long-term hardware bet that the future of AI runs locally — and that it intends to own that hardware layer.
Conclusion
The Apple M7 Ultra with 1.5TB of unified memory isn't just a spec bump — it's a strategic statement. Apple is reorganising its entire chip roadmap around neural performance, accelerating development timelines it previously treated as fixed, and positioning its silicon to compete in the data center. Whether the memory shortage allows the full configuration to ship, and whether Apple's AI software can match its hardware ambitions, remain open questions. But the M7 Ultra signals that the era of on-device AI running models at scale is no longer theoretical. Apple is building the hardware for it right now.
