Apple wants companies to stop paying per token and buy a desk full of Macs instead for their AI work. In the same week as the new Macs dropped, OpenAI and Anthropic cut down their API prices. We asked our newsletter readers which approach they’d choose.
The Pitch That Started This
On September 22, Apple began shipping its new Mac mini and studio desktops with a unique launch. Instead of selling them on speed or design, it sold them on a running cost argument. Apple’s idea is simple. A one-time hardware purchase could let users run models on their own machines without having to pay cloud providers for every token.
A quick note on what pay per token means when doing AI work. When businesses use AI models like ChatGPT or Claude for their workflows through the cloud, they usually pay by the “token.” A token is a small chunk of text, roughly three quarters of a word. Both the prompts a company sends and the responses it receives are tallied to its total bill. A short query might come up to only a fraction of a cent. That said, an AI agent that spends hours scanning documents, analysing information and rewriting code can easily consume millions of tokens. This is why usage can add up very quickly, much like a taxi meter that never stops running. Apple is positioning itself as the alternative solution by effectively shifting the cost structure.
The numbers behind this pitch aren’t small either. The Mac mini starts at $899 while the Mac Studio with Apple’s new M5 Ultra chip goes for $5,499 and the fully specced versions are priced at close to $20,000. Apple, in its press release, says the Mac Studio lets users run massive models entirely on device “without counting tokens or worrying about rising cloud costs.”
To prove the point and show its capabilities, Apple connected four Mac Studios together and ran a trillion-parameter model to hunt down a graphics bug. This is the sort of workload that is usually associated with a data centre, but Apple says the entire setup ran from a single wall socket.
Apple’s edge here traces back to a decision it made in 2020 for completely different reasons. When it moved to its own chips, it put the processor and memory on the same package to improve battery life. That design happens to suit AI work very well, because large models need fast access to a lot of memory. Apple stumbled into an advantage years before anyone was asking for one.
There’s already some evidence the market agrees. Mac minis sold out in parts of China as an open-source agent tool called OpenClaw spread, and the high-memory Mac Studio configurations have been hard to find since launch.
Meanwhile, The Meter Got Cheaper
The same week as Apple began shipping the Macs, the two largest model labs went to war on token pricing. This made the entire hardware versus pay-per-token model a genuine debate.
Anthropic slashed Claude Opus 5.5 to $4 per million tokens and $20 per million output, which was a 20% reduction compared to the previous version. Roughly 90 minutes after this news, OpenAI launched GPT-6 Sol and Luna at around half the price of its previous model. Both the companies are now openly competing on what it costs to finish a task rather than solely on benchmark scores.
This raises the question: If renting keeps getting cheaper, then what is the point of investing thousands of dollars into hardware that depreciates?
The price of an individual token tells only one part of the story. Gartner forecasts that inference costs per agentic workflow will rise more than fivefold through 2028. More capable and affordable models can encourage businesses to automate additional tasks, run agents for longer and make more model calls. Consequently, lower unit prices do not guarantee lower total spending.
The Rest Of The Field Is Coming
Apple’s market share in this field is much weaker than the announcement made it seem. According to IDC data cited by Reuters, Apple’s current share of the enterprise desktop market sits at 4.6% versus 91.3% held by Windows. Microsoft is eyeing the same buyers with CEO Satya Nadella describing on-device AI as “unmetered intelligence” at Microsoft Build 2026 in early June and on an X post.
On October 7, at a Windows event in San Francisco, Microsoft launched the Surface Laptop Ultra starting at $2,599 and a Surface RTX spark dev box, a compact desktop with 128GB of memory, priced at $5,999. Meanwhile, Asus, Dell, HP, Lenovo and MSI opened pre-orders for their own RTX spark machines. What’s happening now is that the local AI pitch is becoming a much larger industry debate.
What The Result Says
This is a newsletter that grew out of crypto, a space where “own your stuff” is close to an article of faith. Self-custody, running your own node, not trusting a third party with something you can hold yourself. Readers who follow this space are very familiar with that instinct, whether or not they personally live by it. Apple’s pitch, in a way, is the hardware version of “not your keys, not your coins.” It’s not surprising that it resonated.
What’s more interesting is that it didn’t resonate more.
Own it (38.71%): The plurality, and the phrasing of the option probably helped. “A running meter” is a deliberately uncomfortable image, and anyone who has watched an API bill climb during a busy month knows the feeling. This group is betting that usage only goes one way, which is exactly what Gartner’s projection suggests. If you believe your AI workload will keep growing, a fixed cost starts to look like a hedge. There’s also a privacy dimension here that doesn’t show up in a cost comparison. Data that never leaves your desk can’t leak from someone else’s server.
Pay per token (~31.6%): Close behind, and this group has the stronger argument on the specific facts of that week. Prices had just been cut by up to half. The best models are still the ones you rent, not the ones you can download and run on a Mac, and the gap in capability matters for serious work. Hardware also ages. A machine bought today will be running last year’s models on last year’s chip in two years, while the API quietly upgrades underneath you. Add electricity, maintenance and the effort of setting it up, and “no upfront risk” is a reasonable thing to value.
Depends on the workload (~29.7%): Almost a third of readers refused to pick a side, and honestly, this is the answer most engineers would give. The economics flip depending on what you’re doing. Steady, repetitive, high-volume jobs that run all day favour owning. Occasional, bursty tasks that need the most capable model favour renting. Plenty of teams will end up doing both, running cheaper open models locally for routine work and calling the frontier APIs for the hard problems.
Put the three together and you get a close split, with no option getting near half. This audience was persuaded by the ownership idea without being convinced it wins outright, and the numbers suggest that’s a fair call.
The money for an entry M5 Ultra Mac Studio would buy hundreds of millions of tokens at OpenAI’s new $2 and $10 rates, which is a long runway for one developer and a few weeks for a company running agents around the clock. Falling prices favour the renters for now, but if Gartner’s projection holds and total bills keep climbing, finance teams will start running exactly the sums Apple and Microsoft wants them to run.
The more significant change may be that people are considering local AI in the first place. Just two years ago, the idea of keeping an AI model on a personal computer was far from a mainstream consideration.
The smartest crypto minds already read our newsletter. Want in? Join them.