July 22, 2026 — Apple's next flagship chip may not be just another performance upgrade. According to a new report from Bloomberg's Mark Gurman, the M7 Ultra is being designed with one clear goal: dramatically boosting AI performance. Expected to arrive in 2028, the processor is reportedly being built to handle workloads on a scale that brings it closer to dedicated AI accelerators like Nvidia's Blackwell than traditional desktop processors [citation:1][citation:4].

This is not just a faster Mac chip. This is Apple's most aggressive move into enterprise AI computing yet.

1.5TB Unified Memory: Why It Matters

The biggest headline is the memory. According to Gurman, the M7 Ultra is being designed to support up to 1.5TB of unified memory — roughly double the capacity planned for the M5 Ultra (768GB) and three times the M3 Ultra's 512GB ceiling [citation:1][citation:5].

For context, most flagship laptops in 2026 ship with 16GB or 32GB of RAM. Apple is designing a chip that can hold nearly 50 times that amount [citation:5].

The reason is AI. Large language models are memory gluttons. A trillion-parameter model in FP16 requires roughly 2TB of memory just to load the weights. With 1.5TB of unified memory, a Mac could theoretically run models of that scale locally — something currently only possible on multi-GPU server clusters [citation:1][citation:2].

Apple's unified memory architecture is key here. Unlike traditional workstations with separate CPU and GPU memory pools, Apple Silicon allows CPU, GPU, and Neural Engine to access the same shared memory pool. This eliminates the data-copy bottleneck that plagues conventional systems when running large models [citation:3][citation:8].

However, there is one significant catch. Gurman explicitly notes that whether Apple can actually ship a 1.5TB configuration "will depend on the state of the memory market" [citation:1][citation:6]. Ongoing DRAM shortages have made high-capacity memory both scarce and expensive. SK Hynix has warned that 2027 could be the worst memory shortage year on record [citation:1].

At current enterprise memory pricing of roughly $25 per GB, a 1.5TB configuration alone would cost over $37,500 just for the memory components [citation:1]. Apple has already pulled the 512GB Mac Studio configuration this year due to supply constraints [citation:5][citation:6].

Blackwell-Class AI Performance

Gurman reports that the M7 Ultra will "dramatically upgrade AI performance, bringing it closer to the class of dedicated AI accelerators such as Nvidia Corp.'s Blackwell" [citation:4][citation:6].

This is a significant claim. Nvidia's Blackwell architecture is the current gold standard for enterprise AI compute, powering data center accelerators used for training and inference at scale. Apple is targeting that level of performance with an integrated processor designed for a desktop chassis [citation:2][citation:3].

However, the comparison should be taken carefully. Nvidia Blackwell is not just a chip — it's an entire ecosystem of GPUs, NVLink interconnect, HBM memory, CUDA software, and data center infrastructure [citation:3]. Apple's unified memory architecture may give it advantages in certain AI inference scenarios, but it does not automatically replace Nvidia's data center dominance [citation:3][citation:8].

The M7 Ultra's memory bandwidth is also expected to increase significantly. The M3 Ultra already reaches 819GB/s, and the M7 Ultra is expected to surpass 1TB/s [citation:1][citation:8]. While still well below Nvidia's HBM3 bandwidth (3.35TB/s on H100), the capacity advantage — 1.5TB vs 80-192GB — makes Apple's approach viable for memory-bound workloads where model size, not raw bandwidth, is the bottleneck [citation:8].

The Roadmap: M6 Skips to M7

Apple's chip roadmap has taken an unusual turn. The company is reportedly skipping the M6 Pro, Max, and Ultra tiers entirely. Only the base M6 is expected to ship in late 2026 for entry-level Macs [citation:6][citation:9].

Instead, Apple has fast-tracked the M7 family. According to Gurman, Apple taped out the M7 just six months after beginning the M6 tape-out — an unusually fast cadence that signals Apple's urgency around AI [citation:2][citation:6].

This decision is pure AI. Gurman reports that Apple had planned major Neural Engine upgrades for the M7 family and ultimately determined those improvements were significant enough to skip the rest of the M6 generation [citation:7]. The company is no longer building chips that happen to support AI — it's building chips around AI [citation:4][citation:5].

The 2028 generation also includes the M8 chip, codenamed Soko, built on TSMC's 1.4nm process (A14 node). Another high-end Mac chip, codenamed Cardinal, is also in development for the same generation [citation:6][citation:7].

Beyond Macs: Apple's AI Server Strategy

The M7 Ultra is not just being built for future Macs. According to Gurman, Apple also plans to use the chip as the backbone of its next-generation AI servers [citation:4][citation:5].

While an M5 Ultra-based server platform (codenamed J246) is expected to arrive first, engineers are already developing a more powerful M7 Ultra-powered architecture targeted for deployment around 2029. These servers would power Apple Intelligence both on-device and in the cloud [citation:3][citation:6].

This is a significant strategic shift. Apple has historically relied on third-party cloud providers, but it is now building its own AI infrastructure. The M7 Ultra server chip would allow Apple to run its most demanding AI workloads — including training and inference for its largest models — entirely on its own silicon [citation:4][citation:5].

In a way, this is a full-circle moment. The last Apple computer to support 1.5TB of RAM was the 2019 Intel Mac Pro, the famous "cheese grater" tower. That machine used traditional slotted memory. When Apple switched to its own silicon, unified memory arrived with blazing bandwidth but a hard ceiling. The M7 Ultra would finally match that memory record while keeping the enormous bandwidth advantage that unified memory brings [citation:5].

But at a price. Given the cost of memory alone, a fully specced M7 Ultra Mac or server would likely be priced well into six-figure territory, positioning it squarely in the workstation and enterprise segment — not a casual creator purchase [citation:2][citation:5].

Key Takeaways