I Tried to Buy a Mac Studio This Week. Here's What Happened.

Historical research — October 11, 2026: This records my April 14 shopping experience. Store inventory, rental rates, reported benchmarks and future-chip rumors below describe that period, not current buying guidance. Apple’s March 2025 specification gives the M4 Max Studio a 36GB starting memory configuration. The rental arithmetic is illustrative and does not establish equal throughput, currencies, taxes or total ownership cost.

I've been running our AI stack on a rented L4 GPU on Google Cloud.

$700 a month. Serves a Qwen 14B model. Fine for production.

Not fine for what I want to do next.

I want to self-host a 70B open-source model on my own hardware. Run it locally. Write about the whole process in public. Eventually help other founders do the same.

The plan was simple. Mac Studio. M4 Max. 128GB unified memory. Quiet enough to sit on a desk. Fast enough to run Llama 3 70B or Qwen 72B. Cheap enough to expense, not capitalize.

I opened the Apple Store.

"Available."

"Available" Doesn't Mean What You Think

That was the trap.

"Available" meant the base configuration I was shown. My original 24GB/12-core description was inaccurate for the March 2025 M4 Max Mac Studio, which starts at 36GB. Ships in two to three weeks. Too small for the particular large-model workload I wanted; smaller models can still be useful.

Every config above that — the 48GB, the 64GB, the 128GB that anyone doing real work would actually buy — gone. Not backordered. "Currently Unavailable." Apple won't even take your money.

The 256GB M3 Ultra config? Pulled from the store entirely in March.

I checked Canada. Same. Checked Germany. Three to four months on the 128GB. Checked China. Four to five months. Australia. Extended delays across the board.

This isn't a pre-launch inventory gap. It's rather a global memory supply crisis hitting the exact configurations that matter for AI.

The Datacenter Buildout Is Eating Your Hardware

The reason is structural.

The same three companies that make Apple's unified memory — Samsung, SK Hynix, Micron — are the same three companies building the memory stacks inside every NVIDIA H100 and H200 shipping to datacenters right now.

I did not have a like-for-like production-yield comparison between datacenter memory and Mac memory.

My concern was that major datacenter investment could put pressure on memory supply; the original $650 billion combined figure was not supported by a comparable set of company capital-expenditure disclosures here. All of it memory-hungry. All of it pulling from the same suppliers Apple depends on.

That was my interpretation of the supply pressure, not evidence of suppliers’ allocation decisions.

Apple can't outbid the datacenter buildout. So the Mac Studio sits on a "Currently Unavailable" page while the stripped-down base config pretends everything is fine.

What X Told Me

I posted about it on X. The replies were consistent.

"Don't buy a Mini for this. You want a Mac Studio with 128GB minimum."

"Wait for M5 Ultra at WWDC in June."

"Everyone is buying these. Global RAM shortage. Good luck."

"I'm running 70B on a rented H100 for $2.99/hr. The Mac doesn't pay back."

That last one stuck with me. So I spent the next few days going deeper than I planned.

The M4 Ultra Doesn't Exist

The first thing I learned is that there is no M4 Ultra.

My April reading was that there was no M4 Ultra in the announced Mac Studio lineup. The linked Apple release identifies M4 Max and M3 Ultra; it does not substantiate my original interconnect explanation.

I should not infer a technical impossibility or future product decision from that lineup alone.

Gurman reported in his November 2025 Power On newsletter that Apple killed it entirely — too costly, too few buyers.

At publication I was watching an unconfirmed M5 Ultra/June 2026 rumor; it was not an announced release or a reliable purchase deadline.

So right now, if you want more than 128GB of unified memory on Apple silicon, your only option is the M3 Ultra — a last-gen chip that Apple has stopped accepting orders for in most configurations.

There's a 4x memory ceiling gap between what you can buy and what you actually need. And nothing in between.

The Older Chip Is Faster

Then I looked at the benchmarks and found something that genuinely surprised me.

The M2 Ultra with 192GB runs Llama 3 70B at roughly 12 tokens per second.

The newer M4 Max with 128GB? About 8.3 tokens per second on the same model.

The older chip is faster.

At 70B scale, inference speed is bottlenecked by memory bandwidth, not compute.

The M2 Ultra pushes 800 GB/s. The M4 Max can't move data fast enough to keep up at that model size.

If you're shopping for a Mac to run large models and you're comparing chip generations, the spec sheet is lying to you.

The number that matters isn't on the marketing page.

The Math on Renting vs. Owning

I also ran the math on renting versus owning.

A Mac Studio M3 Ultra with 192GB costs roughly $7,000.

At the quoted $2.99 an hour, eight hours every day for 365 days is $8,730.80 a year. This is a rental-hours calculation, not a matched-throughput comparison.

Dividing a $7,000 purchase by those rental charges gives about 2,341 rented hours, or 9.6 months at that schedule, before running costs, currency, tax, resale and workload differences. That does not establish measured payback.

At 40 rented hours a month, $2.99 per hour is $119.60 a month or $1,435.20 a year. Comparing that with a $7,000 purchase gives about 58.5 months of rental charges before electricity, taxes, currency conversion, maintenance or resale; it does not support a 20x claim.

I'm not running inference eight hours a day right now. I'm experimenting. Building content around the process. I don't know what my sustained workload looks like yet.

Buying hardware I can't get, for a workload I haven't proven, doesn't make sense. Renting does.

What I'm Actually Doing

I'm renting an H100 on DigitalOcean's Toronto data center. $2.99 an hour.

I chose a Canadian data-center location and brought my own model. That location alone does not establish legal jurisdiction, regulatory compliance or that a rented provider cannot access the workload.

I'll write about the experience from the rental side. Honestly. Not pretending supply chains don't exist.

In April, I was treating the unconfirmed M5 Ultra/June rumor as a possible future buying window, contingent on an announced configuration I could actually order.

My plan at the time was to keep renting if that configuration did not ship or lead times were too long.

The Sovereignty Spectrum Has Three Tiers

This experience reshaped how I think about what "own your AI" actually means in 2026.

For the last two years, the story has been simple. Buy a Mac. Download Ollama. Run your models locally. You're sovereign.

That story is broken. Not because the hardware doesn't work — it does. Because you can't buy it.

There's a tier one. People who bought their Mac Studio two years ago.

They own the hardware. They're running models, shipping content, making it look easy on YouTube. Good for them. They timed the window.

There's a tier two. People like me, trying to buy in this week. The Apple Store leads with the base config to make it look available.

The spec you actually need is gone. NVIDIA's DGX Spark is $2,999 and competitive — but it doesn't officially sell in Canada. This tier is broken.

And there's a tier three. Bare-metal rental. H100s on DigitalOcean Toronto, Lambda, Hetzner. Three dollars an hour.

Canadian data center if you want it. Bring your own model. That was the rental path I chose while inventory was uncertain. Rented hardware still has an operator, contract and capacity constraints; it is a different form of control from owning the machine.

My April choice was rental while I tested a workload and watched inventory; this diary cannot establish where most self-hosting would happen in 2026. My checked configurations were unavailable in that April search; I did not establish that suitable hardware was unavailable everywhere.

If You're Thinking About Buying a Mac for AI

Two things.

One. Check the exact configuration and delivery estimate against your workload. My April search showed why a storefront availability label was not enough for my particular purchase.

Two. My M5/WWDC plan was a dated, unconfirmed possibility. A current buying or rental decision needs current announced specifications, actual availability, prices and a tested workload comparison.

The path to owning your AI just got narrower than the marketing made it look.