Apple is reportedly preparing its most aggressive Mac silicon reset yet, skipping the M6 Pro, M6 Max, and M6 Ultra while accelerating an AI-focused M7 family. The flagship M7 Ultra, expected in 2028, is being engineered to support as much as 1.5TB of unified memory, although Apple has not announced the chip or committed to selling a Mac with that capacity.
Bloomberg’s Mark Gurman detailed the roadmap in Sunday’s Power On newsletter, with subsequent reporting from VideoCardz, MacRumors, and Tech Times. The standard M6 is still expected to appear, but higher-end Mac buyers would move from today’s M5 Pro and M5 Max systems directly to M7-class replacements rather than receiving the usual M6 derivatives.
That distinction matters. This is a supply-chain report about products potentially arriving between 2027 and 2028, not an Apple launch announcement, and specifications can change long before manufacturing begins.
Since the M1 generation, Apple has generally expanded each silicon family through base, Pro, Max, and Ultra configurations. Some individual variants have been omitted—the M4 Ultra was a notable absence—but abandoning an entire generation of high-end chips would represent a broader break from that pattern.
According to Gurman, Apple began the M7’s tape-out process only six months after reaching the same engineering milestone with the M6. The reported reason is a substantial Neural Engine upgrade intended to improve on-device AI performance enough that shipping interim M6 Pro and Max parts no longer made strategic sense.
The provisional schedule now puts the base M7 in the first half of 2027, followed by M7 Pro and M7 Max near the end of that year. M7 Ultra would arrive in 2028, probably in a Mac Studio-class desktop, while a server processor derived from the same architecture is reportedly being considered for Apple’s infrastructure around 2029.
For professional MacBook buyers, the practical consequence is a longer gap between premium chip generations. Apple introduced M5 Pro and M5 Max MacBook Pro models in March 2026, and those machines may remain the current high-end options until the reported M7 replacements arrive late in 2027.
Apple has not publicly confirmed that there will be no M6 Pro or M6 Max. Buyers should therefore treat “canceled” as the present state of the reported internal roadmap rather than an irrevocable promise about future products.
The comparison is not straightforward, however. The Intel Mac Pro used replaceable ECC DIMMs, allowing memory to be upgraded after purchase. Apple’s unified memory is installed in the processor package and must be selected when the computer is ordered.
In return, the CPU, GPU, and Neural Engine can access the same memory pool without moving large datasets between conventional system RAM and a discrete graphics card. Apple’s M3 Ultra provides more than 800GB/s of memory bandwidth, and its 512GB configuration can accommodate models and datasets that will not fit inside the VRAM of a single conventional workstation GPU.
Scaling that architecture to 1.5TB could make an M7 Ultra Mac attractive for local inference, scientific visualization, large media projects, and development involving exceptionally large AI models. It would not automatically make the machine faster than an NVIDIA Blackwell accelerator.
NVIDIA’s B200 has 180GB of HBM3e and is built for vastly higher bandwidth, multi-GPU scaling, and sustained data-center operation. Apple’s potential advantage would be the amount of memory visible to a single application on one compact workstation; NVIDIA’s advantages would remain throughput, software maturity, interconnects, and deployment at cluster scale.
“Approaching Blackwell” should consequently be read as a broad design ambition, not a benchmark result. No M7 Ultra core counts, bandwidth figures, power limits, test results, or supported AI data types have been disclosed. Without those details, a direct performance comparison is premature.
SK Hynix chief executive Kwak Noh-jung told Reuters on July 10 that 2027 could bring the memory industry’s worst-ever supply shortage. He expects demand to continue exceeding the company’s production capacity beyond 2030 despite investment in new facilities.
That forecast overlaps directly with Apple’s reported M7 production window. Even if suitable high-density memory is technically available, Apple would need enough supply at a defensible price to build commercial configurations rather than a handful of engineering systems.
The same uncertainty applies to the reported M5 Ultra, which is expected later in 2026 with support for as much as 768GB. Support in the memory controller does not guarantee that every theoretical capacity will appear in Apple’s online configurator.
Pricing would be another barrier. High-memory Mac configurations already carry steep premiums, and a fully populated 1.5TB system would almost certainly sit far outside conventional desktop budgets. For IT departments, it would compete less with premium PCs than with specialized workstations and entry-level AI servers.
Apple is pursuing a tightly integrated machine with an enormous shared memory pool, predictable drivers, and relatively modest power consumption. Windows and Linux workstation vendors can instead combine replaceable system memory with one or more discrete accelerators, offering greater flexibility in GPU choice, PCIe expansion, networking, storage, and post-purchase upgrades.
A future 1.5TB Mac could run models that exceed the memory of any single discrete GPU, but capacity alone does not settle the platform question. Windows workstations can distribute workloads across multiple professional GPUs, while enterprise AI stacks remain heavily optimized around NVIDIA’s CUDA software and data-center hardware.
Apple will also have to persuade developers that its software environment can use such a large memory pool efficiently. Model compatibility, quantization support, Metal acceleration, framework performance, and sustained thermal behavior will matter more than the headline capacity once systems reach customers.
The strongest immediate signal is not that Apple has built a “Blackwell killer.” It is that Apple reportedly sees local AI as important enough to disrupt its established product cadence and leave professional buyers without M6-class Pro, Max, or Ultra upgrades.
Anyone who needs a Mac workstation now should evaluate current M5 Pro, M5 Max, and M3 Ultra hardware against actual workloads rather than waiting solely for an unannounced 2028 configuration. The next meaningful checkpoint will be Apple’s expected M5 Ultra launch later in 2026, which should reveal how much high-capacity unified memory the company can source—and how aggressively it is willing to price it—before the far more ambitious M7 Ultra reaches production.
Bloomberg’s Mark Gurman detailed the roadmap in Sunday’s Power On newsletter, with subsequent reporting from VideoCardz, MacRumors, and Tech Times. The standard M6 is still expected to appear, but higher-end Mac buyers would move from today’s M5 Pro and M5 Max systems directly to M7-class replacements rather than receiving the usual M6 derivatives.
That distinction matters. This is a supply-chain report about products potentially arriving between 2027 and 2028, not an Apple launch announcement, and specifications can change long before manufacturing begins.
Apple Trades a Predictable Cadence for an AI Leap
Since the M1 generation, Apple has generally expanded each silicon family through base, Pro, Max, and Ultra configurations. Some individual variants have been omitted—the M4 Ultra was a notable absence—but abandoning an entire generation of high-end chips would represent a broader break from that pattern.According to Gurman, Apple began the M7’s tape-out process only six months after reaching the same engineering milestone with the M6. The reported reason is a substantial Neural Engine upgrade intended to improve on-device AI performance enough that shipping interim M6 Pro and Max parts no longer made strategic sense.
The provisional schedule now puts the base M7 in the first half of 2027, followed by M7 Pro and M7 Max near the end of that year. M7 Ultra would arrive in 2028, probably in a Mac Studio-class desktop, while a server processor derived from the same architecture is reportedly being considered for Apple’s infrastructure around 2029.
For professional MacBook buyers, the practical consequence is a longer gap between premium chip generations. Apple introduced M5 Pro and M5 Max MacBook Pro models in March 2026, and those machines may remain the current high-end options until the reported M7 replacements arrive late in 2027.
Apple has not publicly confirmed that there will be no M6 Pro or M6 Max. Buyers should therefore treat “canceled” as the present state of the reported internal roadmap rather than an irrevocable promise about future products.
The 1.5TB Figure Is About Capacity, Not Confirmed Performance
A 1.5TB unified-memory ceiling would triple the 512GB maximum supported by the M3 Ultra Mac Studio introduced in 2025. It would also match the maximum RAM capacity of the 2019 Intel Mac Pro, finally closing one of the few conspicuous specification gaps between Apple silicon and Apple’s former Xeon workstation.The comparison is not straightforward, however. The Intel Mac Pro used replaceable ECC DIMMs, allowing memory to be upgraded after purchase. Apple’s unified memory is installed in the processor package and must be selected when the computer is ordered.
In return, the CPU, GPU, and Neural Engine can access the same memory pool without moving large datasets between conventional system RAM and a discrete graphics card. Apple’s M3 Ultra provides more than 800GB/s of memory bandwidth, and its 512GB configuration can accommodate models and datasets that will not fit inside the VRAM of a single conventional workstation GPU.
Scaling that architecture to 1.5TB could make an M7 Ultra Mac attractive for local inference, scientific visualization, large media projects, and development involving exceptionally large AI models. It would not automatically make the machine faster than an NVIDIA Blackwell accelerator.
NVIDIA’s B200 has 180GB of HBM3e and is built for vastly higher bandwidth, multi-GPU scaling, and sustained data-center operation. Apple’s potential advantage would be the amount of memory visible to a single application on one compact workstation; NVIDIA’s advantages would remain throughput, software maturity, interconnects, and deployment at cluster scale.
“Approaching Blackwell” should consequently be read as a broad design ambition, not a benchmark result. No M7 Ultra core counts, bandwidth figures, power limits, test results, or supported AI data types have been disclosed. Without those details, a direct performance comparison is premature.
Memory Suppliers May Decide What Apple Can Sell
Apple may design support for 1.5TB without offering that configuration at launch. Gurman’s reporting explicitly makes the retail ceiling dependent on memory availability and cost, two factors that are becoming less predictable as AI infrastructure consumes a growing share of advanced memory production.SK Hynix chief executive Kwak Noh-jung told Reuters on July 10 that 2027 could bring the memory industry’s worst-ever supply shortage. He expects demand to continue exceeding the company’s production capacity beyond 2030 despite investment in new facilities.
That forecast overlaps directly with Apple’s reported M7 production window. Even if suitable high-density memory is technically available, Apple would need enough supply at a defensible price to build commercial configurations rather than a handful of engineering systems.
The same uncertainty applies to the reported M5 Ultra, which is expected later in 2026 with support for as much as 768GB. Support in the memory controller does not guarantee that every theoretical capacity will appear in Apple’s online configurator.
Pricing would be another barrier. High-memory Mac configurations already carry steep premiums, and a fully populated 1.5TB system would almost certainly sit far outside conventional desktop budgets. For IT departments, it would compete less with premium PCs than with specialized workstations and entry-level AI servers.
Windows Workstations Still Own the Expansion Story
For Windows professionals, Apple’s roadmap highlights two increasingly different approaches to local AI hardware.Apple is pursuing a tightly integrated machine with an enormous shared memory pool, predictable drivers, and relatively modest power consumption. Windows and Linux workstation vendors can instead combine replaceable system memory with one or more discrete accelerators, offering greater flexibility in GPU choice, PCIe expansion, networking, storage, and post-purchase upgrades.
A future 1.5TB Mac could run models that exceed the memory of any single discrete GPU, but capacity alone does not settle the platform question. Windows workstations can distribute workloads across multiple professional GPUs, while enterprise AI stacks remain heavily optimized around NVIDIA’s CUDA software and data-center hardware.
Apple will also have to persuade developers that its software environment can use such a large memory pool efficiently. Model compatibility, quantization support, Metal acceleration, framework performance, and sustained thermal behavior will matter more than the headline capacity once systems reach customers.
The strongest immediate signal is not that Apple has built a “Blackwell killer.” It is that Apple reportedly sees local AI as important enough to disrupt its established product cadence and leave professional buyers without M6-class Pro, Max, or Ultra upgrades.
Anyone who needs a Mac workstation now should evaluate current M5 Pro, M5 Max, and M3 Ultra hardware against actual workloads rather than waiting solely for an unannounced 2028 configuration. The next meaningful checkpoint will be Apple’s expected M5 Ultra launch later in 2026, which should reveal how much high-capacity unified memory the company can source—and how aggressively it is willing to price it—before the far more ambitious M7 Ultra reaches production.
References
- Primary source: Tech Times
Published: 2026-07-13T19:03:59+00:00
Apple Rewires Chip Roadmap Around AI: M7 Ultra Targets 1.5TB, Eyes NVIDIA Blackwell
Apple M7 Ultra chip is engineered to support 1.5TB of unified memory and approach NVIDIA Blackwell AI performance levels, Bloomberg’s Mark Gurman reports. Apple has canceled M6 Pro and M6 Maxwww.techtimes.com - Independent coverage: videocardz.com
Published: 2026-07-13T06:43:04+00:00
- Related coverage: 9to5mac.com
M7 Ultra to potentially feature up to 1.5TB of RAM, finally matching 2019 Mac Pro: report - 9to5Mac
According to the latest Power On newsletter from Mark Gurman, Apple is preparing to include up to 1.5TB of RAM...9to5mac.com - Related coverage: macrumors.com
Apple's M7 Ultra Chip Designed to Match a 2019 Mac Pro Feat
Apple's M7 Ultra chip coming in 2028 is designed to support up to 1.5TB of unified memory, according to Bloomberg's Mark Gurman. However, whether such a configuration is offered may depend on the state of the ongoing memory chip shortage. However, whether such a configuration is...www.macrumors.com - Official source: support.apple.com
Mac Studio power consumption and thermal output (BTU) information - Apple Support
Learn about the power consumption and thermal output of Mac Studio computers.support.apple.com - Official source: apple.com
Apple reveals M3 Ultra, taking Apple silicon to a new extreme - Apple (MY)
Apple today announced M3 Ultra, offering the most powerful CPU and GPU in a Mac.www.apple.com
