Apple’s most powerful Mac chip is now six weeks into real-world use, and the picture of what it actually changes is getting clearer. The company introduced the M5 Ultra on August 25, 2026, packing a 36-core CPU, an 80-core GPU and up to 512GB of unified memory into a refreshed Mac Studio. Apple called it its “most powerful chip ever,” and on paper the spec sheet backs that up.

The more interesting story is what the chip signals about Apple’s AI strategy. For the first time on an Ultra-class chip, the GPU carries dedicated Neural Accelerators, and memory bandwidth climbs to 1.2TB/s. Apple is betting that Mac Studio buyers want to run large language models entirely on their desks, with no API bill and no data leaving the building. That pitch lands differently in October 2026 than it would have a year earlier, now that developers routinely weigh the cost of hosted inference against buying the hardware once and running models locally.

Apple Confirms the M5 Ultra’s Final Specs

Apple’s August 25 announcement locked in the specs that ship in the current Mac Studio lineup, which became available on September 22, 2026. The top configuration pairs a 36-core CPU, split into 12 “super cores” and 24 performance cores, with an 80-core GPU and a 32-core Neural Engine. Apple’s own technical specifications page lists unified memory bandwidth at 1.2TB/s, with configurations scaling up to 512GB of unified memory.

Apple left little doubt about who this chip targets. In its newsroom announcement, the company said that with the M5 Ultra, “Mac Studio scales up to a 36-core CPU, up to an 80-core GPU, and a staggering 512GB of unified memory, enabling users to run enormous LLMs entirely on device.” That line, taken directly from Apple’s own release, reads like a message to one specific audience: developers and researchers who want generative AI workloads running locally instead of through a hosted API.

A hands-on review of the M5 Ultra Mac Studio already measured sizable generational gains in real workloads, which lines up with the multipliers Apple published at launch. Six weeks on, the gap between Apple’s marketing claims and independent testing has mostly closed.

36 Cores, Two Kinds: What “Super Cores” Actually Means

The core count itself is new territory for Apple. Previous Ultra chips split cores into performance and efficiency tiers. The M5 Ultra instead combines 12 super cores with 24 performance cores for a 36-core total, a naming change Apple has not fully explained at the architectural level in the material it has published so far. What the company has disclosed is the performance outcome: according to Apple’s announcement, “the new Mac Studio with M5 Ultra features an up-to-36-core CPU with 12 super cores and 24 performance cores, delivering up to 1.3x higher multithreaded performance than M3 Ultra.”

A 1.3x multithreaded jump is a solid but not dramatic gain by chip-generation standards, and it tracks with the core count increase from the M3 Ultra’s 32 cores to 36. The headline number here isn’t the CPU. It’s what Apple did to the GPU.

The 80-Core GPU Gets Neural Accelerators for the First Time

Here’s where the M5 Ultra story actually gets interesting. The GPU still tops out at 80 cores, matching the M3 Ultra’s highest GPU configuration. Raw core count didn’t move. What changed is what’s inside each core. Apple’s announcement states that the chip’s “up-to-80-core GPU, the most powerful Apple silicon GPU ever, brings Neural Accelerators to the Ultra chip for the first time, enabling up to 4.3x the peak AI compute performance when compared to M3 Ultra.”

A 4.3x AI compute jump on identical GPU core counts means the gain comes almost entirely from dedicated matrix math hardware built into each core, not from scaling the chip up. That mirrors a pattern playing out across the GPU industry generally, where vendors increasingly embed tensor-style accelerators directly into shader cores rather than relying on general-purpose compute for AI workloads. Nvidia still dominates the broader discrete GPU market by a wide margin, a point underscored by a recent shipment tally putting Nvidia’s desktop GPU share near 90%, but Apple isn’t trying to compete on raw shipment volume. It’s competing on what a single machine can do without a network connection.

512GB of Unified Memory and the 1.2TB/s Bandwidth Number

Memory is the other half of the pitch, and it’s arguably the more consequential one for the AI crowd. The M5 Ultra supports up to 512GB of unified memory, shared across CPU, GPU and Neural Engine, moving at up to 1.2TB/s. Apple’s Mac Studio specs page confirms both figures as the ceiling for the top configuration.

Why Local LLMs Need That Much Memory

Running a large language model locally means fitting every parameter, plus the active context window, into addressable memory. A 70-billion-parameter model stored at 16-bit precision alone needs well over 100GB before accounting for context and overhead. A 512GB pool gives a single Mac Studio enough headroom to hold several large models at once, or one model with a very long context window, without swapping to disk. That’s the practical reason Apple keeps repeating the phrase “enormous LLMs entirely on device” rather than talking about gaming or video rendering first.

None of that memory comes cheap right now. Memory pricing across the industry has been climbing for months, and a recent report tracked RAM eating up to 60% of some device bills of materials as supply stays tight. A 512GB configuration sits well above what most consumer machines ship with, which puts Apple’s top Mac Studio tier squarely in a market segment defined as much by memory economics as by raw chip performance.

Six Weeks Later: What Changed Since the August 25 Unveiling

By October 4, the M5 Ultra has been shipping in customers’ hands for close to two weeks, and sitting in the news cycle for six. That’s enough time for the initial announcement excitement to fade and for harder questions to surface. Pricing for every configuration tier is one of them. Apple has not published a full, itemized price breakdown across every M5 Ultra memory and storage tier in the sources reviewed for this piece, so anyone shopping for a specific build should check Apple’s own store listing directly rather than rely on secondhand figures.

What has emerged instead is a clearer sense of who’s actually buying. Early coverage from outlets tracking the launch, including an analysis from Macworld, points to the same buyer profile Apple targeted at launch: developers running local models, studios doing heavy render and compute work, and researchers who want a desk-sized machine instead of a cloud bill.

How the M5 Ultra Compares With Three Generations of Apple Silicon

Apple’s Ultra tier has a short, slightly uneven history. The company skipped an “M4 Ultra” entirely, going straight from the M3 Ultra to the M5 Ultra. Lined up against its three predecessors, the generational jump looks like this:

ChipAnnouncedMax CPU CoresMax GPU CoresMax Unified MemoryMemory Bandwidth
M1 UltraMarch 8, 202220 (16+4)64128GB819.2GB/s
M2 UltraJune 5, 202324 (16+8)76192GB819.2GB/s
M3 UltraMarch 12, 202532 (24+8)80512GB819.3GB/s
M5 UltraAugust 25, 202636 (12 super + 24 perf)80512GB1.2TB/s

Two things stand out. GPU core count has plateaued at 80 since the M3 Ultra, meaning the M5 Ultra’s AI compute gains come from architecture, not brute-force scaling. And memory bandwidth jumped from roughly 819GB/s, where it sat for three straight generations, to 1.2TB/s. That’s the single biggest architectural change in the table, and it’s the number that matters most for anyone loading large models into memory.

The Competitive Picture: Nvidia, AMD and the Workstation Market

Apple isn’t selling the M5 Ultra as a head-to-head GPU competitor to Nvidia’s workstation and data center cards, and it shouldn’t try. The desktop GPU market remains lopsided: Nvidia holds close to 90% of shipments, per the shipment data cited above, and its stock buyback program alone recently outpaced Apple’s buyback spending, a sign of how much cash the AI chip boom has generated for Nvidia specifically. Consumer GPU pricing has gotten ugly too, with flagship cards like the RTX 5090 vanishing from US retail shelves and reselling near $9,500 amid tight supply.

Apple’s angle is different: a single purchase, a single power outlet, and a unified memory pool big enough to hold models that would otherwise require multiple discrete GPUs linked together. That’s a narrower pitch than “biggest GPU wins,” but it’s a real one for teams that don’t want to manage a GPU cluster. Apple briefly overtook Nvidia on a notable AI-related market milestone earlier this year too, though that lead lasted only two months before reversing, a reminder that the AI hardware race rewards whoever ships the next thing fastest, not whoever led last quarter.

Memory Economics: Why 512GB of RAM Costs What It Costs

The DRAM Shortage Context

Unified memory isn’t free to manufacture, and it’s getting more expensive industry-wide, not just at Apple. A separate piece of reporting found that DRAM now costs 54% more than leading-edge TSMC 2nm logic silicon, a striking reversal from the usual relationship between memory and compute pricing. When the memory chips inside a machine cost more per square millimeter than the processor itself, high-memory configurations stop being a minor upsell and start driving the bulk of a machine’s price.

That dynamic puts Apple in an unusual position. Unified memory is central to the M5 Ultra’s entire pitch, since the chip’s value proposition for AI work depends on fitting huge models into a single memory pool. But if DRAM supply stays tight into 2027, the 512GB configuration could become the hardest tier to keep in stock, not the easiest to sell at a premium.

Mac Studio’s Place in the Local AI Hardware Race

The broader context here is a hardware category that barely existed three years ago: desktop machines built explicitly to run frontier-scale AI models without a cloud connection. Cloud GPU rental still wins on raw throughput for training, but for inference, especially for small teams running fine-tuned models, the economics increasingly favor owning hardware outright. A Mac Studio with 512GB of unified memory removes an entire category of decisions, like how to shard a model across multiple cards, that cloud GPU users have to make constantly.

Apple is not alone in chasing this market, and it won’t be the cheapest option in every case. What it has, that most PC-based competitors don’t, is a single chip design that scales from a laptop to a workstation without switching architectures. That consistency is worth something to developers who want their local testing environment to resemble their production environment.

Checking Your Own Mac Studio’s Core and Memory Configuration

For anyone who already bought an M5 Ultra Mac Studio and wants to confirm exactly which configuration shipped, macOS exposes the core counts and chip details directly through the terminal, without needing third-party software.

sysctl -n hw.physicalcpu hw.logicalcpu
sysctl -n machdep.cpu.brand_string
system_profiler SPHardwareDataType | grep -i -E "chip|memory|cores"

Running those three commands in Terminal returns the physical core count, the chip name as macOS reports it, and the installed unified memory size, which is the fastest way to verify a configuration matches what was ordered.

Market Impact: What the M5 Ultra Means for Apple’s AI Positioning

Apple has spent the past two years fielding criticism that it lagged behind in the generative AI race compared with Microsoft, Google and Nvidia’s ecosystem of partners. The M5 Ultra doesn’t change that criticism directly, since Apple still doesn’t operate a frontier cloud model to rival GPT or Gemini at scale. What it does is stake out a different battlefield entirely: the hardware layer underneath everyone else’s models. If a developer runs an open-weight model locally on a Mac Studio instead of renting cloud GPU time, Apple wins that transaction regardless of which AI lab built the model.

That’s a quietly aggressive strategy. It sidesteps the extremely expensive, extremely competitive race to build a leading foundation model and instead sells the infrastructure that any model can run on. Whether that’s enough to matter at Apple’s scale depends on how many developers are willing to pay a premium for a desktop machine instead of a monthly cloud bill, which is exactly the tradeoff the 512GB memory ceiling is designed to tip in Apple’s favor.

Historical Context: From the M1 Ultra to Today

UltraFusion’s Four-Year Run

The Ultra tier has existed for four and a half years, starting with the M1 Ultra in March 2022. Apple built every Ultra chip since by fusing two Max-tier dies together with its UltraFusion interconnect rather than designing a standalone Ultra die from scratch. That approach explains the oddly flat GPU core ceiling between the M3 Ultra and M5 Ultra, both capped at 80 cores: fusing two dies caps the GPU core count at whatever two Max chips can provide, and Apple apparently hit a practical ceiling there two generations ago.

The gap in the lineup is just as telling. Apple shipped the M3 Ultra in March 2025 and then skipped straight to the M5 Ultra in August 2026, never releasing an Ultra version of the M4 generation at all. That’s the longest stretch yet between Ultra-tier releases, and it suggests Apple is now treating the Ultra chip less as an annual refresh and more as an occasional, memory-and-bandwidth-driven event tied to specific workloads rather than a yearly calendar.

M5 Ultra’s Generational Gains at a Glance

Apple published specific multipliers comparing the M5 Ultra against the M3 Ultra, its immediate Ultra-tier predecessor. Laid out together, the gains are uneven across categories, which tells its own story about where Apple focused its engineering effort this generation.

MetricM3 UltraM5 UltraApple’s Stated Gain
CPU Cores32361.3x multithreaded performance
GPU Cores8080No core increase
Peak AI ComputeBaselineHigherUp to 4.3x
Memory Bandwidth819.3GB/s1.2TB/sRoughly 1.46x
Max Unified Memory512GB512GBUnchanged

The pattern is clear once it’s laid out this way. Apple held memory capacity and GPU core count flat while pushing bandwidth up by roughly 46% and AI compute up by more than 4x. That’s an architecture generation built around moving data faster and processing AI workloads more efficiently with the same core budget, not a generation built around brute-force scaling.

What Comes Next: Predictions for Apple Silicon Through 2027

  • Apple will likely carry Neural Accelerators down into the standard M6 GPU lineup rather than reserving the feature for Ultra-tier chips, following its usual pattern of pushing Ultra-tier features down a generation later.
  • The Mac Studio with M5 Ultra becomes the reference machine cited in local-LLM benchmarking discussions through 2027, simply because few competing desktop systems offer a comparable unified memory ceiling.
  • If DRAM pricing keeps climbing the way the TSMC 2nm comparison suggests, expect the 512GB configuration to face stock constraints before Apple cuts its price.
  • Apple is unlikely to release another Ultra-tier chip on an annual cycle. Expect the gap pattern from M3 Ultra to M5 Ultra, built around specific memory and bandwidth milestones rather than a calendar, to repeat.
  • Competing PC vendors will respond with higher unified or shared-memory ceilings of their own rather than competing purely on GPU core count, following the same bandwidth-first logic Apple just demonstrated.

Frequently Asked Questions

What is the Apple M5 Ultra?
The M5 Ultra is Apple’s highest-end Mac chip, announced August 25, 2026, available in the Mac Studio since September 22, 2026. It scales up to a 36-core CPU, an 80-core GPU, a 32-core Neural Engine, and up to 512GB of unified memory.

How much faster is the M5 Ultra than the M3 Ultra?
Apple states up to 1.3x higher multithreaded CPU performance and up to 4.3x higher peak AI compute performance compared with the M3 Ultra, its direct predecessor.

What is a “super core” on the M5 Ultra?
Apple’s own materials describe the 36-core CPU as a combination of 12 super cores and 24 performance cores. Apple has not published a detailed architectural breakdown of how super cores differ internally from performance cores.

How much does the M5 Ultra Mac Studio cost?
Apple has not published a complete price breakdown across every memory and storage configuration in the sources reviewed for this article. Check Apple’s official Mac Studio store page for current pricing on specific configurations.

Why does unified memory matter for running AI models?
Large language models need their full parameter set loaded into memory to run. A 512GB unified memory pool, shared across CPU, GPU and Neural Engine, lets a single Mac Studio hold very large models or long context windows without swapping data to slower storage.

Did Apple ever release an M4 Ultra?
No. Apple went directly from the M3 Ultra, announced in March 2025, to the M5 Ultra in August 2026, skipping an Ultra-tier version of the M4 generation entirely.

How does the M5 Ultra’s memory bandwidth compare with earlier Ultra chips?
Memory bandwidth sat at roughly 819GB/s across the M1 Ultra, M2 Ultra and M3 Ultra. The M5 Ultra raises that to 1.2TB/s, a jump of around 46%, the largest bandwidth increase in the Ultra chip’s history.

Can the M5 Ultra compete with Nvidia GPUs for AI workloads?
Not directly on raw throughput. Nvidia holds close to 90% of desktop GPU shipments and dominates large-scale AI training. Apple’s pitch is different: a single machine with enough unified memory to run large models locally without assembling a multi-GPU cluster.