Mac Studio M5 Ultra Brings Trillion-Parameter AI Models to Desks

A desktop pitched as a replacement for cloud AI

Apple's new Mac Studio is sold as a workstation, but the way Apple markets it says something about how the company plans to compete in AI. The refreshed desktop comes with either an M5 Max or an M5 Ultra chip. Apple's own product page says the Ultra version can run "frontier-class models on device" without paying for cloud tokens.11 Several outlets report that Apple announced the machine on August 25, 2026, and that most configurations started shipping on September 22. The top 512GB memory configuration is expected in late October, and some reports say those orders are slipping toward late January.1417

The main news is the chips. The M5 Ultra is essentially two M5 Max chips joined into one, giving it a 36-core CPU, an 80-core GPU and up to 512GB of unified memory, the shared memory pool that the CPU and GPU both draw from. Apple claims memory bandwidth of about 1.2TB per second.12 Both new chips put a Neural Accelerator, a block dedicated to AI math, inside every GPU core. Apple says that gives the M5 Ultra up to 4.3 times the AI compute of the M3 Ultra it replaces.14 One review points out that this is Apple's own comparison, not an independent benchmark.18 The M5 Max model starts at $2,499 and the M5 Ultra at $5,499.1215

The launch matters less as a spec bump than for the economic case Apple is making to buyers. That case lines up with the strategy investors have been rewarding all year: let other companies build huge data centers, rent the frontier model, and push AI work onto hardware customers pay for themselves.

Trillion-parameter models on one power outlet

Apple's launch demo made its intentions clear. According to one account of the event, Apple linked four Mac Studios that ran a trillion-parameter model to find and fix bugs in graphics code, all from a single wall outlet.13 Apple chip chief Johny Srouji reportedly told buyers the machine is a one-time purchase with no per-token charges afterward. Apple pitched enterprises directly on the idea that owning hardware is cheaper than renting cloud compute from OpenAI or Anthropic.13

The clustering feature is what makes this possible. The Mac Studio can now use Thunderbolt 5 with RDMA, a way for linked machines to read each other's memory directly, so several units act as one shared memory pool. Apple says four machines together deliver up to three times the inference speed of one.1418 That is a much smaller gain than four times the hardware might suggest. Still, the larger benefit for big models is likely the combined memory, since that decides whether a model fits at all.

Apple did not invent this use case. Ars Technica reports that hobbyists and researchers had already been chaining Mac minis and Mac Studios to run very large language models. The practice took off after macOS 26.2 added low-latency Thunderbolt 5 links for distributed inference with MLX, Apple's open-source machine learning framework.12 The new desktops build that grassroots use into the product. Apple's launch images even showed LM Studio, a third-party app for running local models.12

The reviews mostly back the hardware claims. PCMag gave the M5 Ultra model an Editors' Choice award, calling it a top machine for local AI inference and saying it is ready to run the most capable open-source models available. The reviewer also noted that its tested configuration cost $12,299.16 Explainers for developers make the same practical point: for local AI, memory capacity sets the largest model you can load and bandwidth sets how fast it answers. That is why the M5 Ultra still beats the cheaper M6 Mac mini despite the smaller generation number.20

The timing against cloud model launches

One report puts the push in a more competitive light. It says John Ternus, who it identifies as Apple's CEO after Tim Cook stepped down this year, posted a new AI-focused pitch for the machines on September 22. That was the same week Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Sol and Luna launched, which the outlet describes as price cuts meant to make frontier AI cheaper to use remotely.20 That timing comes from a single source. If it is accurate, the strategy is clear: just as cloud model prices fall, Apple is telling buyers they can skip the meter entirely.

The two pitches don't fully conflict. Cloud labs compete on the most capable models. Apple is betting that open-weight models, ones anyone can download and run, are now good enough that memory and privacy matter more than raw capability for many business jobs. Local processing can appeal to teams handling sensitive data, though large models and some apps will still need the cloud.18

How the hardware fits Apple's model strategy

The Mac Studio is one part of a larger AI setup that Apple rents rather than builds. In January 2026, Apple and Google announced a multi-year deal under which Apple's next foundation models would draw on Gemini models and Google's cloud infrastructure.9 Reports widely put the cost at about $1 billion a year for a custom Gemini model of roughly 1.2 trillion parameters, and up to $5 billion over the contract.12

Here the reporting disagrees on something important. Some coverage says Google's AI is effectively Siri's new brain. One report says Apple got full access to Gemini in its own data centers and has distilled it into at least five smaller Apple models.1 But Apple software chief Craig Federighi reportedly said at WWDC 2026 that the new Siri does not run on Gemini's models or infrastructure. In his account, Apple's own models were trained on Apple's proprietary data and then refined using outputs from Gemini.9 Both accounts are probably true in part. The most likely explanation is that Gemini acts as a teacher model, a larger model whose outputs are used to train smaller ones, while Apple's distilled models run on devices and in Apple's Private Cloud Compute servers. That arrangement fits with a desktop built to run large models locally.

Apple's claim that its chips share the same design across the lineup supports this. The company says on-device models can run on a high-end Mac Studio and still work on entry-level iPhones and iPads.13 Investor coverage also notes that the base M5 chip runs 30-billion-parameter models locally in under three seconds. That coverage describes Apple's roughly 2.5 billion active devices as a spread-out network of inference machines that need little Apple-owned infrastructure.5

The investor case: rent the model, sell the hardware

The market sees this approach as discipline, not lagging behind. FinanceBuzz reports that Apple budgeted $14 billion for capital spending in fiscal 2026, roughly one-fiftieth of the $650 billion that Alphabet, Amazon, Meta and Microsoft plan to put into AI data centers. It adds that the Gemini license costs less than 1% of Apple's free cash flow and that Apple bought back $90.7 billion of stock in the fiscal year.5 FinanceFeeds compares Apple's capex of under $15 billion with Alphabet's guidance of $195 billion to $205 billion. It credits that gap with making Apple the best performer among the "Magnificent Seven" big tech stocks in 2026, up about 25% this year.2 Another analysis describes Apple as collecting rent on AI through upgrades to devices it has already sold. It notes that December-quarter capex was just $2.37 billion against $53.93 billion of operating cash flow.4

The Mac business shows early signs of the hardware side working. One analysis reports that Mac revenue grew 29% as businesses adopted Apple silicon for on-device inference.8 Wedbush's Dan Ives estimates Apple's AI strategy could add about $75 a share in value, and expects AI to eventually make up more than 15% of Services revenue.10

Where the bull case is weak

The risks are real, and the coverage names them clearly. Apple trades at roughly 36 times expected earnings, according to a figure attributed to the Financial Times. That premium is partly justified by the low-spending AI approach, which leaves little room for mistakes.6 A Seeking Alpha analysis estimates Apple would need to grow free cash flow about 16% a year for a decade to justify its share price. It also notes that the Gemini partnership makes Apple dependent on a direct competitor.7 Siri AI is still not available on iPhone, iPad or Apple Watch in the EU, and remains blocked in China.6 Cook has also warned about rising memory costs.8 That matters for a product line whose main selling point is very large amounts of memory.

Access is a problem too. The machine that makes Apple's "no cloud tokens" argument most convincingly starts at $5,499 and climbs quickly. Its flagship memory configuration ships late, and replies to Apple's launch posts asking for education discounts reportedly went unanswered.1620

The bottom line

The Mac Studio refresh is Apple's capital-light AI strategy in physical form. Apple licenses frontier-model knowledge from Google and distills it into its own models. It then sells the silicon that runs open models and its own models where customers work, and customers pay for that compute up front. The bet is that frontier models become a commodity faster than cloud labs can lock in customers. If open-weight models keep improving, a $5,499 desk machine with no metered billing gets more attractive every quarter. If they don't, Apple's light spending could turn into dependence on Google. For now the market is betting on the first outcome, and the new Mac Studio is the strongest evidence yet behind that bet.