Apple is reportedly skipping three M6-series chips to deliver an M7 Ultra processor with up to 1.5 terabytes of unified memory by 2028, a move that would eliminate the last meaningful capacity gap between Apple silicon and the 2019 Intel Mac Pro. For Windows workstation buyers, the reported roadmap exposes a fundamental fork in professional AI hardware: a tightly integrated Mac with a colossal shared memory pool versus the modular, multi-GPU architecture of the PC ecosystem.

What actually changed

In his July 13 Power On newsletter, Bloomberg’s Mark Gurman reported that Apple has canceled the M6 Pro, M6 Max, and M6 Ultra entirely. The standard M6 will still ship, but professional Mac users will jump from today’s M5 Pro and M5 Max directly to M7-class replacements. The provisional timeline places the base M7 in the first half of 2027, M7 Pro and M7 Max in late 2027, and the M7 Ultra in 2028—likely inside a Mac Studio-class desktop. A server processor derived from the same architecture may follow around 2029.

At the heart of the decision is a substantially upgraded Neural Engine. Apple wants the M7 Ultra to approach the performance class of dedicated AI accelerators like NVIDIA’s Blackwell architecture, and concluded that shipping interim M6 Pro and Max parts made little strategic sense. The company began the M7 tape-out only six months after reaching the same milestone with the M6, an unusually compressed schedule.

The headline number is the 1.5 TB memory ceiling, but it comes with critical caveats. This is an engineering target, not a confirmed retail configuration. Gurman explicitly states that whether Apple offers a Mac with that much memory will depend on availability and cost—two factors that are growing more unpredictable as AI infrastructure consumes advanced DRAM. SK Hynix CEO Kwak Noh-jung warned on July 10 that 2027 could bring the industry’s worst-ever supply shortage, with demand exceeding production capacity beyond 2030. Apple has already been forced to discontinue high-memory Mac Studio configurations due to the current shortage.

What it means for Windows users

The immediate impact on everyday Windows PC users is negligible. The drama plays out in the professional workstation arena, where the two platforms are taking sharply different roads to address AI workloads that demand enormous memory.

For Windows professionals running large language models locally, Apple’s unified memory architecture offers a singular advantage: every byte of that 1.5 TB is accessible to the CPU, GPU, and Neural Engine without the copying overhead that plagues discrete GPU setups. A Mac with 192 GB of unified memory already provides roughly 192 GB of effective GPU memory—something no consumer NVIDIA card can touch, since even top-tier GeForce GPUs top out at 32 GB of VRAM. Scaling that to 1.5 TB would allow a single Mac to hold massive AI models that simply will not fit in any single discrete GPU.

But Windows and Linux workstations answer with modularity. You can pair system RAM with one or more professional GPUs—NVIDIA RTX 6000 Ada cards with 48 GB each, for instance—and scale across multiple accelerators using NVLink and PCIe expansion. For AI training and high-throughput inference where many users issue concurrent requests, NVIDIA’s memory bandwidth (roughly 8 TB/s on the B200) and CUDA software maturity remain in a different league. Apple’s current ultra-class bandwidth hovers around 800 GB/s, a tenfold gap that translates directly into tokens-per-second speed.

The decision tree splits along use cases: if your workflow demands loading a 100-billion-parameter model for local experimentation or creative work on a single quiet, low-power machine, Apple’s architecture is uniquely suited. If you need to train models, serve inference at scale, or rely on CUDA-optimized libraries, a multi-GPU Windows or Linux workstation remains the better bet.

IT departments face a starker calculus. A fully configured 1.5 TB Mac would almost certainly cost well over $35,000 based on Apple’s current unified memory pricing, pushing it into competition with entry-level AI servers rather than traditional desktops. And unlike a Dell Precision or HP Z workstation, the memory is soldered and non-upgradeable. You commit to your capacity at purchase.

How we got here

Apple’s memory arc has been the slow-burning subplot of the Apple silicon transition. The M1 Max in 2021 supported 64 GB. The M2 Ultra doubled that to 192 GB. The M3 Ultra Mac Studio, launched in March 2025, pushed the ceiling to 512 GB with over 800 GB/s of bandwidth. But the 2019 Intel Mac Pro, with its user-replaceable DDR4 ECC DIMMs, still held the capacity crown at 1.5 TB—though its per-channel bandwidth was anemic by modern standards. The M7 Ultra’s target would close that gap with the bandwidth advantage inverted: unified memory bandwidth that the Xeon workstation never approached.

The AI catalyst pulls from Apple’s canceled Project Titan self-driving car program. The machine learning and custom silicon work done for that decade-long, $10 billion effort directly shaped the Neural Engine that debuted in the iPhone X in 2017. Tim Cook described the project as “the mother of all AI projects,” and the redirected research is now driving the M7’s neural processing ambitions. Without that earlier investment, Apple would likely be even further behind in on-device AI.

Meanwhile, the memory market has become the wild card. Apple already raised MacBook Pro prices by $300 on June 25 and pulled 256 GB and 512 GB Mac Studio configurations from sale. SK Hynix’s dire forecast for 2027 lands squarely over the period when Apple’s engineers must finalize what the M7 Ultra actually ships with. Samsung and Micron have issued broadly aligned warnings, though Bloomberg Intelligence suggests the shortage may peak in mid-2026 and ease by 2028—a timeline that would serendipitously align with the M7 Ultra’s arrival.

What to do now

If you are evaluating a professional workstation today, do not anchor your purchasing decision to an unannounced product three years away. The most reliable approach is to evaluate current hardware against real workloads.

For Windows and Linux users:
- If your software stack depends on CUDA or requires multi-GPU scaling, your path is clear: stick with the PC ecosystem. Look at current NVIDIA RTX professional cards or the next-generation “Blackwell” workstation GPUs expected in the coming months.
- For large-model inference that exceeds 48 GB of VRAM, consider a dual-GPU configuration or a system with ample system RAM and CPU offloading. Tools like llama.cpp and MLC-LLM can make large models usable on PC hardware without requiring a single giant memory pool.

For those considering a Mac:
- The M5 Ultra Mac Studio, expected later in 2026 with up to 768 GB of unified memory, is the near-term high-memory option. It will likely be the most capable Mac workstation until at least late 2027.
- If you do decide to wait for the M7 Ultra, understand that the 1.5 TB configuration may never ship at retail, and if it does, the price will be substantial. The safer assumption is that 768 GB or 1 TB will be the practical ceiling, with the chip’s AI performance being the larger differentiator.
- Professionals who need a MacBook Pro with Pro or Max silicon should know that the M5 Pro and M5 Max models released in March 2026 will remain the current high-end options until the M7 Pro and M7 Max appear in late 2027. There will be no M6 Pro or M6 Max MacBooks.

The strongest near-term signal will come later this year when Apple launches the M5 Ultra. How much memory Apple can source and how aggressively it prices that memory will reveal how realistic the M7 Ultra’s 1.5 TB ambition truly is.

Outlook

Apple’s gambit is not really about one big memory number. It is about reordering priorities: AI inference throughput now sits above CPU benchmarks, battery life, and thinness as the primary driver of Apple’s chip roadmap. For Windows workstation buyers, the M7 Ultra’s emergence won’t change the fundamental calculus—expandability, software ecosystem, and modular GPU scaling still tilt in the PC’s favor—but it will raise the bar for what a single, integrated machine can do with local AI. Watch the memory market, and watch what NVIDIA does next.