On Monday, Apple did something it rarely does: it raised the price of a product without pretending it hadn’t. The new Mac Studio with the M5 Ultra chip starts at $5,499 — $200 more than the M3 Ultra it replaces — while the M5 Max version holds steady at $2,499. The company announced the M6, its first 2-nanometer chip, in the same breath, and the tech press dutifully filed its usual “Apple wins the silicon race again” stories.
But the chip isn’t the story. The price is.
The $200 That Says Everything
Apple doesn’t raise prices on pro hardware casually. The Mac Studio has been the quiet workhorse of the Apple lineup since 2022 — no redesigns, no keynote theatrics, just a steady cadence of spec bumps. When the company raises the price on the Ultra tier while holding the Max tier flat, it’s making a statement about who it thinks is buying these machines.
The M5 Max at $2,499 is still a creative pro’s machine. Video editors, 3D artists, the people who need eight external displays and Thunderbolt 5. The M5 Ultra at $5,499 is something else. Configure it with the full 512GB of unified memory — which, notably, doesn’t ship until late October, a full month after the base models — and you’re looking at a machine that costs more than a used Honda Civic and exists for exactly one reason: running large AI models locally.
512GB of Unified Memory Is Not for Photoshop
Here’s the spec that should make Nvidia’s workstation division sit up straight: 512GB of unified memory. That’s not a typo. The M5 Ultra can address half a terabyte of memory as a single pool, shared between CPU and GPU. No discrete GPU on the market offers anything close to that at this price point — and Apple is selling an entire computer, not just a card.
Apple’s own press release says the Mac Studio delivers “up to 4.3x faster performance” for on-device AI. The marketing images show LM Studio and MATLAB running local inference. This is not subtle. Apple is selling a machine for people who want to run large language models without renting time on someone else’s data center.
One machine learning engineer who pre-ordered the 512GB config within an hour of the announcement put it bluntly: “I’ve been renting A100s on AWS for two years. This machine pays for itself in eight months, and I don’t have to deal with spot instance pricing or data egress fees.” He asked not to be named because his employer is currently negotiating a cloud contract.
The Cloud Providers Should Be Nervous
The conventional wisdom in AI infrastructure has been simple: training happens in data centers, inference happens in data centers, and the only question is which hyperscaler gets the contract. Nvidia’s entire valuation rests on that assumption. Apple just made a different bet.
If a $5,499 desktop can run a 70-billion-parameter model locally — and with 512GB of unified memory, it can — then a meaningful slice of the inference market doesn’t need the cloud at all. Developers who fine-tune models, researchers who run experiments, companies that don’t want their proprietary data leaving the building: these are exactly the customers AWS, Azure, and Google Cloud have been counting on to justify their AI infrastructure buildout.
Apple isn’t going to kill the data center. Training frontier models still requires thousands of GPUs and more power than any desktop can draw. But inference — the thing everyone actually pays for on a recurring basis — is a different story. Every developer who buys a Mac Studio instead of renting cloud GPUs is a small leak in the hyperscaler revenue model. Small leaks add up.
The M6, for all its 2-nanometer novelty, is a laptop and consumer chip. The M5 Ultra is the strategic move. Apple has spent five years building the best consumer silicon in the world, and now it’s quietly pointing that silicon at the most lucrative market in tech: AI infrastructure. The $200 price hike isn’t greed. It’s Apple telling the market what it thinks this machine is worth — and what it thinks the cloud is not.