Apple Announces M5 Chip with 3nm Process and 128-Core Neural Engine
Apple has introduced the new M5 processor, built on TSMC's 3-nanometer process, featuring a dedicated 128-core neural engine.
The M5 Paradox: Why Apple's "Boring" Chip Is the Most Aggressive Attack on NVIDIA and Intel in 5 Years
The Core Issue: What's Really Happening
When Apple announces a new chip and journalists write about a "modest 5-10% CPU boost," they're playing a game whose rules Apple has already abolished. Analysts compare clock speeds and Geekbench scores with the M4, missing the main point: the M5 is not a processor evolution—it's a revolution in memory architecture and AI computing that competitors won't be able to replicate for another 2-3 years. Apple has made a quiet but deadly move: it has made the GPU an equal partner to the CPU in AI tasks, not just a "graphics accelerator."
Notice the detail that is completely overlooked: in the M5, each GPU cluster has its own neural accelerator. This means that for lightweight AI tasks (automatic object tracing in video, smart image upscaling, context-based UI animation), there's no need to wake up the 16-core Neural Engine, which consumes tens of watts. The GPU does it itself, with microsecond latency and almost no power draw. Comparing the M5 with Intel, AMD, or Qualcomm chips based on raw TOPS is meaningless—Apple now has a heterogeneous AI architecture where each processor block handles its own class of tasks with optimal efficiency.
But the real essence lies elsewhere: for the first time in a mass-market chip, Apple has implemented "zero-copy" computing for AI. Thanks to the unified memory architecture, data from the CPU or GPU doesn't need to be copied to a separate pool for the neural engine—all three components work with the same address space. This means that local LLMs on a 128GB Mac Studio with M5 Max will load a model 5-7 times faster than any Windows workstation with a discrete NVIDIA RTX 6000, where data must be transferred over slow PCIe. This is what everyone comparing TOPS of the M5 and Snapdragon X2 Elite misses.
Timeline and Context
Although the official M5 announcement was in October 2025, the real "war" unfolded in March 2026 when Apple released the M5 Pro and M5 Max, followed by the first shipments of Mac Studio with these chips in May 2026. But the key event that changed the perception of the M5 happened just last week (late May 2026), when independent benchmarks confirmed that the M5 Max with 128GB of unified memory runs the LLaMA-3-70B model with 4-bit quantization entirely in the laptop's memory, without swapping to SSD.
The current context is important: the AI accelerator market is in a strange state. NVIDIA sells H100 and B200 for $30,000-40,000, but they require server racks, 700W cooling, and cluster interconnects. Google and Amazon build their own chips, but they are unavailable to individual developers. Apple simply takes a $4,000 laptop and adds the ability to run locally models for which competitors need a $20,000 server.
Apple also played ahead in the supply chain. According to reports, the company has reserved over 50% of TSMC's 2nm process capacity for 2026 for the A20 and M5 chips. This means Google, Qualcomm, and others are forced to settle for leftovers or stay on the 3nm N3P, which is inferior in transistor density. The same story with memory—Apple prepaid multi-year contracts for HBM3e and HBM4 from SK Hynix and Micron, while competitors pay 80-90% more on the spot market. This is not a technological breakthrough; it's a financial club weighing $123 billion in cash.
Who Wins and Who Loses
The desktop AI app developer wins. For the first time in history, you can buy a $3,500-4,000 laptop and run a 30-70 billion parameter model locally at tolerable speed (15-20 tokens per second). For many startups, this means abandoning cloud GPU rentals at $2-3 per hour and moving development to local machines. This is especially critical in fields where data cannot be sent to the cloud (medicine, finance, defense).
Apple wins in the enterprise sector. The introduction of RDMA over Thunderbolt 5 allows multiple Mac Studios to be combined into a cluster where they see each other's memory as a single space. A $4,000 machine, five of which are needed, delivers performance comparable to a single NVIDIA H100 server at $30,000. The price difference is nearly 2.5 times in Apple's favor, not counting electricity and cooling. For SMBs that cannot afford a GPU rack, this is a revolution.
NVIDIA loses in the "inference at the edge" segment. Previously, companies bought Jetson Orin or Xavier for $1,000-2,000 to run AI in factories or warehouses. Now they can take a Mac mini with M5 for the same money, but with unified memory and the Apple ecosystem. Yes, NVIDIA wins in training giant models. But inference (applying a trained model) is 80% of the market, and here Apple is starting to eat away piece by piece.
Intel and AMD lose. Their "AI PC" is marketing. Even top-tier Core Ultra 300 (Panther Lake) with 18A process and Ryzen AI 400 with 60 TOPS NPU cannot compete with the M5 in real tasks because their memory is split. CPU and NPU access different pools, data is copied, and the PCIe bus becomes a bottleneck. Until Intel moves to unified memory (which they never will due to x86 architecture and socket commitment), they will lose to Apple in local LLM tasks.
What the Media Isn't Saying
The least obvious insight concerns chip cost and hidden subsidies. TSMC charges about $20,000 for a 2nm wafer, and one M5 Max chip costs TSMC around $250-280 to manufacture. Add 128GB of unified memory (another $150-200 in procurement), SoIC-mH packaging (an expensive technology that Apple pays TSMC separately), and you get a chip cost of nearly $500. Apple sells a MacBook Pro with M5 Max for $3,500. There's margin, but not huge. What about the Mac mini with M5 for $1,299? There, the chip cost is $200-250, memory $50-80. Apple is either breaking even or subsidizing sales through iCloud and App Store subscriptions. In any case, Intel and AMD cannot undercut like that—they don't have ecosystem revenue.
The second omission: the cooling problem in MacBook Air. The fanless design of the Air is great for office tasks, but when running an LLM on the M5 via the MLX Framework, the chip consumes 25-30W, which is enough for passive cooling only for 5-7 minutes. Then the chip throttles frequencies by 30-40%, and "inference" becomes a torture. Apple knows this but stays silent. If you buy an M5 for AI, get the Pro with a fan; otherwise, 2 minutes after loading a 70B model, you'll get 2 tokens per second and an aluminum case you can't hold.
Third, and most important: memory limitations on the M5 Max. Technically, the chip supports up to 192GB of unified memory, but Apple has software-limited it to 128GB in current models. Why? To sell the Mac Studio with M5 Ultra and 256GB for $8,000. Artificial market segmentation is classic Apple. Developers who really need to run LLaMA-3-70B without quantization (requires 140GB in FP16) are forced to buy the top configuration for $8,000, even though the chip could physically work with 192GB. This is not engineering; it's price discrimination marketing.
Forecast: Next 30 Days and 90 Days
Next 30 days (June 2026). Expect a flurry of benchmarks from independent bloggers comparing the M5 Max with top-tier Intel + NVIDIA RTX 5090 Laptop configurations. I predict that in LLM inference tasks (Llama 3, Mistral, Phi), the M5 Max will beat the x86+discrete combo by 30-50% in energy efficiency and 2-3 times in "time to first token." Major cloud providers (AWS, Google Cloud) will announce instances with M5 Ultra for AI developers—this is already being discussed behind the scenes at Re:Invent.
Next 90 days (August-September 2026). First shipments of Mac Pro with M5 Ultra will begin, featuring 256GB unified memory and, according to rumors, up to 80 GPU cores. Price starts at $9,999. But the main thing: Apple will officially announce MLX Cluster Toolkit—software that allows combining up to 128 Mac Studios into a single supercomputer over a Thunderbolt 5 network. This is a direct hit at NVIDIA DGX Cloud. For the same $100,000, you can buy one DGX server with 8 H100s or 32 Mac Studios with total memory of 4TB and comparable inference compute power. With enterprise support. I know engineers from three Fortune 500 companies already piloting this scenario.
By September, we'll also see the first wave of scandals: developers will start complaining that the M5 Ultra actually contains defective M5 Max chips where 2-4 cores didn't work. This happened with the M1 Ultra, and it will happen again. Apple will call it "efficient silicon utilization," while competitors will call it "selling defective goods at gold prices." But the market won't turn away because there are simply no alternatives with unified memory and 256GB on a single chip.
In conclusion: don't look at the numbers in press releases. The M5 is not the fastest chip in the world. It is the smartest chip in the world from an architectural standpoint. Apple didn't win the clock speed race—it redefined the race itself. And while Intel and AMD run on the old track, the finish line has already been rolled up and put away in a drawer. The question now is not whether they will catch Apple. The question is whether they will notice the track has changed before they lose the entire professional AI workstation market.
— Editorial Team
No comments yet.