14 September 2026
Heard In AI

Apple's 512GB Mac Studio and the case for owning your AI

On Moonshots with Peter Diamandis, Salim Ismail argued that a Mac Studio with 512GB of unified memory changes AI spending from a perpetual per-token bill into a capital asset, with law firms and mid-sized healthcare organizations as the likely buyers. Two other panelists agreed the machine was worth having and still called Apple's AI record a long-running software failure.

A briefing reports one development when it happens. We correct or clarify it later; a new development gets a new briefing. How our formats work

"The 512 gig unified memory puts a massive amount of intelligence under your desk," Salim Ismail said on Moonshots with Peter Diamandis. For him the interesting part was not the specification but the invoice it replaces. Buying that much memory, he argued, shifts the economics "from paying per token perpetually to buying a capital asset and then using it continuously."

Most people meet AI as a metered service: a model runs in somebody else's data center, and each request is billed by the token, the small chunk of text a model reads and writes. Ismail's claim is that a desktop machine can move that spending from a subscription line into the column where companies put equipment they own. He added that Apple has gone further by starting to lease its hardware to customers.

He thought the obvious step had simply been taken late. If Apple had not shipped this, he said, "they would have been the stupidest company in history and they're not that dumb."

What Apple actually announced

The two new chips are not interchangeable. According to Apple's announcement on 25 August 2026, the M6 in the Mac mini is built on a 2-nanometer manufacturing process and supports up to 32GB of unified memory. The 512GB figure belongs to the M5 Ultra in the Mac Studio, which Apple pairs with 1.2 terabytes per second of memory bandwidth, up to 36 CPU cores and up to 80 GPU cores. Apple describes its UltraFusion interconnect as letting four dies behave as a single processor.

Two numbers do the work in that list. A model's weights — the billions of numbers learned during training — have to sit in memory before the machine can use them, so capacity decides which models fit at all. Bandwidth, the rate at which those numbers can be moved, decides how quickly the machine can work through them to produce an answer. Unified memory means the CPU and GPU draw on the same pool rather than copying data between separate ones. Apple names Core AI, Core ML, Metal and Xcode as the developer tools for the hardware, and lists running and fine-tuning models locally among its intended uses. The episode introduced the Mac Studio as a desktop with enough memory to run some of the largest openly available models locally, with Apple "now pushing local AI as an alternative to AI in the cloud."

Who would buy one

Ismail's buyers are organizations with a reason to keep data off other companies' servers: "law firms, healthcare companies that have HIPAA issues and can't put their stuff on the cloud." HIPAA is the US law governing protected health information. He was most interested in the middle of the market — medium-sized healthcare firms that cannot afford to build a private cloud of their own. "So you get four of these Mac studios cluster together," he said. "Now you've got like a small private data center."

The conditions matter as much as the picture. Keeping processing in the building removes an outside provider from the arrangement, but HIPAA governs how an organization handles patient information wherever it is stored, so on-premises hardware is a way of meeting those obligations rather than an automatic answer to them. And the capital asset is not free to run once bought: four machines still consume electricity, need someone to maintain them and run whatever model version their owner has loaded, while a cloud provider's models are updated for everyone. The saving Ismail describes comes from heavy, continuous use of equipment that has already been paid for.

"At the end of my rope with Apple"

Agreement stopped there. "Salim, I'm at the end of my rope with Apple," Dave said. He began with what he owed the company: nobody he had never met had changed his life more than Steve Jobs, the Apple II he had as a child had changed it, and "when they came out with MacOS based on Linux, changed my life again." Then: "It's absolutely the greatest company I've ever seen, but they have totally missed the boat. And it's just embarrassing that a company with that much cash flow has no AI strategy. They don't even deserve to be a Mag 7 company anymore" — a reference to the group of large US technology stocks the market treats as a bloc.

The new Mac Studio did not move him. "So if the best they can do in the age of AI is add a bunch more RAM to a machine that they already had, I mean, what the hell, man?" He put it at "like 0.01% of what they should have done by now in AI." Ismail did not dispute the record. His reply was that the company had at least done the obvious thing: "At least they did this, right?"

Hardware too good for the software

A third panelist offered another reading, tracing the technology's origins. The unified memory architecture and the neural engine in Apple's M-series chips, he said, descend from the Apple car project that never shipped; that program was why Apple first built a neural engine with access to a large unified addressable space. From there he drew a pattern: Apple's interaction with AI is "almost a history of hardware that's too good for the software, or conversely software that just isn't worthy of the hardware."

He listed the advantages Apple held early — matrix-multiplication acceleration inside a unified addressing space, the manufacturing relationship, and "the raw transistor throughput to basically be Nvidia and then some. And they fumbled it." Apple had Siri before everyone else and had audio interaction, he said, and "Apple software fumbled the ball." Both Tim Cook and Jobs before him had described the company as hardware plus software plus services; in this panelist's view the hardware is incredible, the services are "not that bad," and the software still underuses both. His appeal to Apple's chief executive was direct: "take your amazing hardware and unleash it with much better software that integrates AI natively."

That is where the panel's two arguments meet. Ismail's case for the machine does not depend on Apple shipping better AI software, because the software running on it would be somebody else's open model, chosen by the law firm or clinic that bought the box. The criticism from the others is about everything else Apple sells. A panelist put the stake plainly during the exchange: Apple, he said, is the one brand you can trust with your information — and here we are in the age of AI. Apple's own answer, for now, is a list of developer tools and an invitation to run and fine-tune models on the desk.

Share this article

Go to the original

Sources & further reading

  1. 01

Connected ideas and articles

From the conversation

Podcast episodes

Article history

Updates to this article

Tags

Graylin says China’s AI advantage is deployment, not an AGI finish line

Alvin Graylin argues that China is competing to spread useful AI through industry and overseas developer communities, rather than betting everything on reaching general intelligence first. Provincial competition and open-weight models help explain his account, though the policy contrast is not absolute: America’s AI Action Plan also explicitly promotes adoption.

7 min read

Why a Moonshots panel thinks China's AI tokens go to video and America's to code

Alibaba's Wan 3.0 and a relayed claim that 70% of Chinese AI token use goes to video sent the Moonshots panel into an argument about money: one guest said American labs chase revenue per token while Chinese labs give their weights away, another said video is the only market that will trust a Chinese model. They ended up disagreeing about whether world models or text models reach self-improving AI first.

6 min read

OpenAI's Jalapeño chip has the Moonshots panel asking: could it sell compute to rivals?

On Moonshots EP #284, the panel read out the performance figures OpenAI published for Jalapeño, the inference chip it built with Broadcom, and argued that inference is moving off NVIDIA. One panelist went further, imagining an "OpenAI Compute" cloud that rents capacity to competitors — possibly even hosting an Anthropic model. Dave Blundin explained why CUDA no longer locks buyers in at inference time, and why NVIDIA's next defense is the networking between chips.

6 min read

Parallel sold patient search agents before it could afford a web index

On Training Data, Parag Agrawal explains how his company Parallel entered web search without first building a giant index: it launched a search agent that crawled after a request arrived, replaced outsourced human data collection for insurance, sales and finance customers, and treated the index as a latency optimization to be grown later. He describes the agent-specific architecture behind it, the 200-millisecond Turbo mode Parallel announced in July, and a Google Cloud deal that puts Parallel Search beside Google Search as a grounding option.

7 min read