Apple pitches Mac Studio silicon as a desk-side alternative to enterprise rack servers

Meta released the Muse Spark Open model as open weights, meaning its trained parameters are publicly available, on Oct. 15, 2026, and had Apple’s help running it on two to four linked Mac Studios powered by M5 Ultra chips.

The model carries 816 billion total parameters while activating roughly 42 billion for each token, a size that suits desktop hardware. 9to5Mac describes Muse Spark Open as Meta’s most capable open-weight model, with a context window of a million tokens. Alexandr Wang, Meta’s chief AI officer, said the model is not Meta’s most capable, and the sources disagree on its standing.

Apple’s Machine Learning Research post says four M5 Ultra Mac Studios run Muse Spark Open above 36 tokens per second, with each machine peaking around 300 watts. Apple cites that on-device inference performance in its desk-side server pitch.

Through Apple’s unified memory architecture, which lets the processor and graphics core share one memory pool, one Mac Studio gives the M5 Ultra’s integrated GPU 512 GB of memory, a capacity edge over servers limited by discrete graphics-card memory. Meta says a pair of linked Mac Studios holding the complete model works as one supercomputer, with one terabyte of unified memory shared across them, positioning two desktops as a rack replacement.

In a launch demonstration, Apple networked four Mac Studios to run a model with one trillion parameters; it found and repaired a programming flaw, all powered by a single wall outlet. Apple presented the demonstration as evidence of local AI capability.

Apple began shipping Mac Studio desktops powered by M5 chips, so businesses can run AI on their own hardware instead of renting data center capacity or paying API fees. Units with M5 Ultra chips and high-end memory reach nearly $20,000, or £15,000, and 512 GB unified-memory models were due to ship in late October.

Apple hardware chief Johny Srouji said local inference carries no cost per token. UltraFusion, Apple’s chip-interconnect technology, fuses two M5 Max chips into the M5 Ultra, doubling CPU and GPU resources while increasing memory capacity and bandwidth. The M5 Max counterpart has an 18-core CPU and a GPU of up to 40 cores. The M5 Ultra Mac Studio with 256 GB starts at $10,799, a small fraction of the $416.16 billion Apple reported for fiscal 2025.

Osaurus, a free open-source Mac tool, manages AI models that run locally and offline, and it can call a frontier model when necessary. Before anything goes to Claude or OpenAI, Osaurus scrubs personal details such as names, phone numbers, and account numbers with a small local model, showing data-sovereignty behavior on the client. Two models in the tool run on 8 GB of memory, and it runs on an Apple silicon laptop with 16 GB of memory and no network connection.

Terence Pae, Osaurus’s co-founder and CEO, said he would not be surprised if local models match today’s frontier models within about a year. That is his outlook on when local models reach parity with frontier systems.

For reinforcement learning and computer-use agent work, OpenAI bought tens of thousands of Macs and deployed them headless without displays or keyboards. Anthropic takes a similar approach, renting Mac mini capacity via Amazon Web Services. That makes Anthropic a second major AI lab using Mac infrastructure, and the OpenAI purchases are evidence of enterprise demand for Apple silicon in AI.

In Apple’s fiscal third quarter, which ended June 27, 2026, Mac revenue rose 29 percent to $10.4 billion, with AI developers buying Mac minis and Mac Studios by the tens of thousands, a financial signal of enterprise AI demand. Some enterprise buyers mount multiple Mac Studios side by side in custom enclosures, an improvised workaround instead of a purpose-built rack. With AI labs outpacing supply, Apple pulled its Mac refresh forward from the usual autumn cadence, a supply response to enterprise demand.

A reported project would give Apple its first rack server for enterprise customers in 15 years, powered by forthcoming M8 Ultra chips and aimed at AI developers, governments, and large businesses. The company is also weighing M-series Ultra servers that could be paired with Nvidia networking equipment and sold to businesses rather than confined to Apple’s own data centers, according to The Information. Apple has not confirmed the project publicly.

The Information and Bloomberg reported that cancellation or a 2029 shipment without Nvidia technology remains possible if Apple’s interconnect work advances. Sources differ on how certain the project is. That leaves launch timing contingent on Apple’s interconnect progress.

UltraFusion links only two M-Max dies within a single chassis at a speed of 2.5 terabytes per second, and was not built for rack-scale use, so Apple needs an outside interconnect. Nvidia’s NVLink 6 connects as many as 72 accelerators in one domain at 3.6 terabytes per second for each XPU, 14 times PCIe Gen 5 bandwidth. NVLink Fusion demands Nvidia hardware in the deployment, so Apple cannot build a pure M8 Ultra-to-M8 Ultra link without Nvidia silicon. Apple sits on the UALink Consortium board, a group developing an open alternative to NVLink.

Apple’s last rack server, the Xserve, was a 1U machine with PowerPC G4 processors that launched in May 2002, and a Virginia Tech cluster called System X briefly ranked near the top of the world’s most powerful supercomputers. Apple ended Xserve sales on January 31, 2011, steering customers toward the Mac Pro or Mac mini instead.

Baltra, Apple’s AI server processor, was set to ship during 2026 for Private Cloud Compute, though Apple has since delayed it. Apple’s AI servers run on Mac-based M2 Ultra chips, which proved insufficient for complex generative AI work. Engineers could not get the performance they needed running Google’s Gemini models locally on Apple’s own servers for an upgraded Siri, so Apple relied on Nvidia GPUs in Google’s cloud for compute-intensive parts of that upgraded Siri effort. The delay explains why Apple needs new server silicon for its own data centers.

To speed server silicon, Apple held talks with semiconductor startups and investment bankers about buying companies, departing from its conservative M&A approach. The purchase of Israeli AI startup Q.ai cost nearly $2 billion earlier in 2026.

About a year ago, John Ternus backed the server project while serving as SVP of hardware engineering, and he became Apple’s chief executive on September 1, 2026. The company lacks a dedicated enterprise engineering organization, enterprise sales infrastructure, and an enterprise support network. The M8 Ultra chip is not real yet; it would follow unreleased chips.

This marks Apple and Meta’s first public partnership as collaborators since about 2017, after Apple’s 2021 privacy changes and other friction between the companies, ending years of public non-cooperation. Mark Zuckerberg posted on X on Oct. 15, 2026: “runs great on new Mac Studios thanks to work with @johnternus and Apple.”

John Ternus, Apple’s CEO, replied on X in praise of on-device AI on the Mac Studio. Meta Chief AI Officer Alexandr Wang said the companies worked for months and that Daniel Gross, an ex-Apple executive now at Meta, played a key role; he also said Apple tunes the model using MLX, Apple’s machine learning framework. The companies disclosed no financial terms, exclusivity period, or duration for the collaboration.

Mount Thor raised $45 million in equity funding so it can rent Macs as cloud infrastructure for computer-use work, with CRV and Conviction co-leading the Series A. Chad Byers and Offline Ventures co-led the seed round, and its strategic angels have ties to Cognition, Core Automation, Daytona, Meta, OpenAI, Together AI, and Vercel.

Renters get exclusive physical Macs or isolated macOS virtual machines, with two VM slots per host. The VM service is in preview, and customers connect by SSH or remote desktop, with traffic through private authenticated tunnels. An inference lab for open-source models optimized for Apple silicon is in the plans, but Mount Thor has published no comparative benchmarks or prices. It took Mount Thor under six months to open its first data center, with colocation planned next in major U.S. metros and then in Europe and Asia.

Meta’s diluted EPS for Q2 came in at $6.18, missing the $7.22 estimate, while free cash flow fell to $784 million and capital spending reached $30.12 billion. For Q3, Meta is guiding to revenue of between $61 billion and $64 billion.

On Oct. 15, 2026, Apple’s stock fell 0.28 percent to $342.00, putting its market value at $4.99 trillion, while Meta shares slipped 0.11 percent to $736.72. Both moves came on the announcement day.