Apple.com

In Sunday's Power On newsletter, Bloomberg's Mark Gurman reported that Apple is engineering the M7 Ultra to support as much as 1.5 terabytes of unified memory — roughly double the ceiling planned for the upcoming M5 Ultra — and that the company wants the chip to approach the performance class of dedicated AI accelerators like NVIDIA's Blackwell architecture. That goal has shaped a decision with no precedent in the Apple Silicon era: Apple is canceling the M6 Pro, M6 Max, and M6 Ultra entirely, jumping straight to M7 in an effort to ship a generational leap in AI performance rather than an incremental refinement.

The 1.5 TB figure, if it ships, would erase the last meaningful memory capacity advantage the 2019 Intel Mac Pro held over Apple Silicon — a gap that has persisted across six years and four M-series chip generations. Whether it actually ships is genuinely uncertain: the same global DRAM shortage that forced Apple to raise MacBook Pro prices by $300 on June 25 and pull 512 GB Mac Studio configurations from sale has also made the memory components the M7 Ultra would require harder to source and more expensive to include. SK Hynix CEO Kwak Noh-jung told Reuters on July 10 that 2027 will be "the worst year in the industry's history from the supply perspective" — a forecast that lands squarely over the period when Apple's engineers will need to finalize what the M7 Ultra actually ships with.

What every Mac buyer needs to know right now: the M6 Pro and M6 Max that many professionals were waiting for are not delayed — they are canceled. Anyone considering a high-end Mac upgrade faces a clearly defined window. The M5 Ultra Mac Studio is expected later this year; the next Pro and Max chips for the MacBook line are now the M7 Pro and M7 Max, arriving no earlier than late 2027. The M7 Ultra follows in 2028. The 1.5 TB memory ceiling, if it clears the memory market, arrives with that 2028 machine.

Why Apple Skipped an Entire Generation for the First Time

Apple's chip roadmap has followed the same pattern since M1 arrived in November 2020: a base chip, followed by Pro and Max variants with more CPU cores, GPU cores, and memory, then an Ultra that fuses two Max dies together. The company previously skipped the M4 Ultra, but it had never canceled all three high-end variants of a generation until now.

Gurman's reporting explains the decision as AI-driven. Apple had been designing major upgrades to the M7's Neural Engine — the dedicated on-device AI processor embedded in every M-series chip — and judged those improvements significant enough to compress the timeline rather than ship an intermediate M6 Pro generation that would have been outpaced months later. Apple began taping out the M7 just six months after completing the same engineering stage for the M6, an unusually compressed schedule that reflects how strongly the company has prioritized getting the AI silicon out.

The Neural Engine itself has an origin story that adds context to the urgency. Apple's canceled Project Titan self-driving car program — shut down in February 2024 after more than a decade of development and an estimated $10 billion in spending — never produced a vehicle. But the machine learning and custom silicon work done for that project became the direct foundation for the Neural Engine, which debuted in the A11 Bionic chip inside the iPhone X in 2017. Apple CEO Tim Cook described the self-driving car program as "the mother of all AI projects" that same year. The engineers and research Apple redirected from Project Titan into its silicon division are now driving the M7's neural processing ambitions. Without that earlier investment, Gurman reported, Apple would likely be even further behind in AI than it already is.

Apple's 1.5 TB Memory Ceiling: What It Actually Is — and What It Isn't

The memory specification requires some unpacking, because the 1.5 TB figure describes a capability Apple is engineering the chip to support, not a product Apple has committed to selling.

Apple's unified memory architecture integrates LPDDR5X DRAM chips directly onto the same package as the processor's compute cores, connecting them with a very wide internal memory bus. This physical proximity and bus width are what give Apple Silicon its high memory bandwidth — the M-series flagship chips have reached approximately 800 GB/s on the M2 Ultra, well above what any conventional laptop RAM configuration can deliver. Because CPU cores, GPU cores, and the Neural Engine all share the same memory pool without copying data between separate chips, large AI models can use every byte of that memory simultaneously. A Mac with 192 GB of unified memory provides roughly 192 GB of effective GPU memory — a capability no consumer graphics card matches, since even the largest discrete GPUs top out at 32 GB of VRAM before requiring expensive workstation hardware.

The architectural tradeoff is that capacity is bounded by how much memory can be physically integrated into the chip package. Apple's UltraFusion technology — which fuses two Max-class dies together to create Ultra chips — is what makes large unified memory capacities possible, but scaling to 1.5 TB requires either significantly denser memory dies, a larger package, or advanced packaging innovations that Apple has not detailed publicly. The M5 Ultra's expected ceiling of 768 GB is itself already a record for Apple Silicon; doubling it again for M7 Ultra is an engineering challenge, not just a purchasing decision.

The Intel Mac Pro comparison is headline-valid and historically meaningful, but it obscures a significant technical distinction. The 2019 Mac Pro's 1.5 TB used standard DDR4 ECC DIMM modules running at roughly 50 GB/s of bandwidth per memory channel — user-replaceable, expandable, but slow by Apple's standards. The M7 Ultra's 1.5 TB, if it ships, would operate at unified memory bandwidth that the 2019 Mac Pro never approached. The closure of the capacity gap also comes with a bandwidth advantage that reverses the historical picture entirely.

What "Approaching Blackwell Territory" Actually Means

Gurman described Apple's goal for the M7 Ultra as approaching the performance class of dedicated AI accelerators like NVIDIA's Blackwell architecture. That is a meaningful aspiration, and it comes with a meaningful qualifier that the framing tends to obscure.

NVIDIA's B200 — the standard Blackwell GPU for data center deployments — carries 180 GB of HBM3e memory running at approximately 8 terabytes per second of memory bandwidth. Apple's current M-series flagship chips run at around 800 GB/s. The gap is roughly tenfold. For AI workloads where memory bandwidth is the binding constraint — specifically, generating tokens from a large language model, where each token requires reading essentially the entire model's weight parameters from memory once — NVIDIA's bandwidth advantage translates directly into speed. A B200 can generate tokens approximately ten times faster than an Apple M-series chip, or serve roughly ten times as many concurrent users at equivalent speed.

Where Apple's architecture wins is capacity relative to what any single GPU can hold. No NVIDIA consumer GPU carries more than 32 GB of VRAM. Professional workstation cards reach 48 GB. Even the B200's 180 GB, while substantial, is less than one-eighth of what Apple is engineering the M7 Ultra to support. For workloads where the model itself is too large to fit in any NVIDIA GPU's memory — frontier language models with 100 billion or more parameters at full or near-full precision — Apple's unified memory architecture is the only single-node option that can hold the model at all. That class of workload, which is increasingly relevant to AI researchers, high-end creative professionals, and organizations running large models locally, is where the M7 Ultra's "approaching Blackwell territory" claim makes the most sense.

What it does not mean is that the M7 Ultra would displace Blackwell in data centers, training infrastructure, or production serving environments where hundreds of users issue concurrent requests. NVIDIA's bandwidth and NVLink-scale interconnect architecture remain in a different league for those use cases, and there is no indication Apple is targeting that market with the Mac workstation.

A Roadmap Shaped by the Memory Market

Apple's capacity ambitions depend on a memory market that is, by the assessment of its largest suppliers, heading into its worst years.

The shortage has already forced Apple's hand multiple times. The 512 GB and 256 GB configurations of the M3 Ultra Mac Studio were discontinued earlier this year, leaving buyers with 96 GB as the highest configuration currently available on that machine — despite the M3 Ultra chip being capable of supporting significantly more. The same DRAM demand from AI server buildouts that has driven Mac prices up has restricted the high-bandwidth memory Apple needs for its own workstation products.

SK Hynix CEO Kwak Noh-jung's forecast — 2027 as the worst year for supply, with demand exceeding capacity beyond 2030 — represents the most pessimistic major-manufacturer view, but Samsung and Micron have issued broadly aligned warnings. Bloomberg Intelligence analyst Shuli Ren offers a more optimistic reading: the shortage may have already peaked in the second quarter of 2026, and conditions could ease by 2028 as new manufacturing capacity comes online. That scenario would be well-timed for the M7 Ultra's expected arrival, though it remains a minority view against the major supplier consensus.

The base M7 chip is expected to arrive in the first half of 2027, appearing first in an entry-level MacBook Pro. M7 Pro and M7 Max variants are expected to follow in late 2027 — the Pro-tier chips that professionals waited for from the M6 generation and will not get. The M7 Ultra, with its potential 1.5 TB ceiling, is expected in 2028. An M7 Ultra-based AI server for Apple's own infrastructure is planned for around 2029. Apple is also reported to be working on M8 chips on a 1.4 nm manufacturing process for 2028, moving to a new node that promises better efficiency per computation than the 2 nm process expected for M7.

How Does This Change What Pro Mac Buyers Should Do?

The practical decision tree for anyone currently in the market has been simplified by the M6 cancellation — which is, depending on your perspective, either clarifying or disheartening.

If the primary need is the highest available unified memory and top AI inference throughput for a Mac workstation, the near-term option is the M5 Ultra Mac Studio when it arrives later this year, with up to 768 GB of unified memory — a record for Apple Silicon that will hold until the M7 Ultra appears in 2028. The long-term option is waiting for the M7 Ultra, with the understanding that the 1.5 TB ceiling may ship, may not, and will carry a price tag well above $35,000 for the maximum memory configuration based on Apple's current pricing of roughly $25 per additional gigabyte of unified memory.

If the need is a MacBook Pro with Pro or Max-class silicon, the relevant products are the current M5 Pro and M5 Max machines. The next MacBook Pro with Pro or Max silicon will carry the M7 Pro and M7 Max — expected late 2027. There is no M6 Pro or M6 Max MacBook coming at any point.

The broader signal in Apple's chip strategy is not the memory figure, though that is the most striking number in the announcement. The signal is that Apple has decided AI inference throughput now ranks above CPU performance, battery efficiency, and thinness as the primary driver of its chip roadmap. That reordering — from consumer benchmarks to AI accelerator comparison points — is what the M7 generation represents, and it shapes the product line for the next several years regardless of whether the 1.5 TB Mac arrives on schedule.

Frequently Asked Questions

Why did Apple skip the M6 Pro and M6 Max entirely?

Apple concluded that the Neural Engine upgrades planned for the M7 family were significant enough to justify collapsing the roadmap rather than shipping an intermediate M6 Pro generation. According to Gurman's reporting, the company began taping out the M7 just six months after completing the same engineering stage for the M6, compressing the schedule to get the AI-focused chip out faster. An M6 Pro would have been outpaced by M7 Pro months after release, making it an awkward product to build and sell. The decision is unprecedented — Apple has previously skipped individual chip variants but has never canceled all three high-end configurations of one generation before M7.

Will Apple actually ship a Mac with 1.5 TB of RAM?

It depends on the memory market. Apple is designing the M7 Ultra to support 1.5 TB, but Gurman explicitly noted that whether a configuration at that ceiling is offered will depend on the state of the global memory industry. SK Hynix, one of the three companies that produce nearly all of the world's DRAM, forecast in July 2026 that 2027 will be the worst year for supply shortages in the industry's history. Apple has already been forced to discontinue high-memory Mac Studio configurations due to the current shortage. Bloomberg Intelligence offers a more optimistic counterpoint — the shortage may ease by 2028, which would coincide with the M7 Ultra's planned arrival — but that remains a minority view.

How does Apple's unified memory actually compare to NVIDIA's GPUs for AI work?

The comparison depends entirely on the workload. Apple's unified memory architecture integrates LPDDR5X directly onto the chip package, giving every byte of RAM full-speed access to the GPU and Neural Engine without the data-copy overhead that slows down conventional CPU-plus-discrete-GPU setups. The advantage is capacity: a Mac with 192 GB of unified memory can hold AI models that no consumer NVIDIA GPU can accommodate, since even top-tier discrete cards cap out at 32 GB of VRAM. Where NVIDIA's Blackwell architecture wins is bandwidth — the B200's 8 terabytes per second of memory bandwidth is roughly ten times what Apple's current Ultra chips deliver, which translates directly into faster token generation speeds and the ability to serve many more simultaneous users. Apple's architecture is better for loading and running very large models on a single machine; NVIDIA's is better for generating output quickly from those models or serving many people at once.

What is the Neural Engine, and what does Project Titan have to do with it?

The Neural Engine is Apple's dedicated on-device AI processor, embedded in every M-series Mac chip and every modern iPhone and iPad chip. It handles machine learning inference tasks — things like image recognition, on-device language processing, and features in Apple Intelligence — without requiring the GPU or CPU, which keeps those compute resources available for other work. Its origins trace back to Apple's canceled self-driving car program, Project Titan, which required Apple engineers to solve real-time AI processing under strict power and latency constraints. The processor developed for the car was never finished, but the underlying research directly shaped the Neural Engine that debuted in the iPhone X in 2017. Apple has scaled and accelerated that technology across every generation of M-series chips since, and the Neural Engine upgrades planned for M7 are the stated reason Apple decided to skip the M6 Pro and M6 Max entirely rather than wait another full cycle.