Alex Finn posted a thread on July 13 that’s already past 84.8K views. His prediction: within roughly two years, Apple will ship Mac Studios with 1.5 terabytes of RAM, citing a rumor from the AppleTrack account about an unannounced chip nicknamed “M7 Ultra.” With about 300GB of that 1.5TB, Finn says, you’ll be able to run “Opus 4.8 level intelligence” on your desk, no cloud required. His conclusion is bold: within two years, everyone on Earth will have a local model running on their own machine, and people should start preparing now.

I agree with the destination Finn describes. I don’t agree with the variable he thinks is standing in the way.

Three days before I wrote this, IQ Source had already covered an experiment that ran a 744-billion-parameter model on a laptop with no graphics card and just 25GB of RAM. Not 300GB. Twenty-five. It worked, producing correct, coherent output. Memory was never the obstacle. The real obstacle is something else entirely, and bolting more terabytes of RAM onto a Mac Studio doesn’t touch it.

What Finn got right, and where the reasoning breaks

I want to start with what’s right in his thread, because it isn’t nothing. Finn is right that AI running on your own hardware, without depending on a cloud API for every call, is heading toward normal. He’s right that companies should start thinking about that possibility now, before it’s urgent. That instinct holds up.

Where it breaks is the variable he uses to measure the distance between today and that future. His chain of reasoning goes: more RAM capacity means more room to run large models, which means in two years (once the higher-RAM hardware ships) everyone will have frontier-level intelligence on the desktop. That chain assumes RAM is the ceiling that has to be raised. And that specific assumption is already disproven, not in two years, but today, on hardware that already exists.

What the rumor he cited actually says

Let’s look at where the number came from. AppleTrack, an account that regularly surfaces Apple supply-chain leaks, posted that a future chip, nicknamed “M7 Ultra,” would support up to 1.5TB of unified memory. There’s no confirmation from Apple. There’s no announced release date. It’s a spec rumor, the kind that circulates months or years before Apple confirms anything, if it ever does.

Finn took that rumor and stacked his own extrapolation on top: if the chip supports 1.5TB, and if 300GB is enough to run something at the level of Opus 4.8 locally, then in two years anyone will have that on their desk. That’s a chain of “ifs” built on a hardware rumor that doesn’t exist as a shipping product yet. And even if the chip landed exactly as AppleTrack describes it, the underlying argument still wouldn’t hold, because RAM capacity was never the variable that determines whether a model runs at a usable speed.

What we already proved with colibrì

The project is called colibrì, it’s open source, and it was built by a developer who posts on Hacker News as vforno. It runs GLM-5.2, Zhipu AI’s mixture-of-experts model with 744 billion total parameters, on a consumer laptop, with no GPU, using 25GB of RAM.

The technique keeps only the model’s dense, shared core resident in memory, about 17 billion quantized parameters, occupying roughly 10GB. Everything else, hundreds of billions of parameters spread across thousands of routed experts, stays on disk and streams in on demand based on what each token needs. The whole engine fits in a single C file of about 2,400 lines.

It works. It produces correct, coherent text. On 25GB, not 300GB. That alone knocks out the central premise of Finn’s prediction: you don’t need terabytes of RAM to run a model with hundreds of billions of parameters. You already can, on a fraction of what he assumes is required.

The real bottleneck: throughput, not capacity

Now the uncomfortable part, because colibrì isn’t free magic. The result produces a new token every 10 to 20 seconds. That’s roughly how long it takes to read this sentence out loud, word by word, before the next one shows up. For a user waiting on a chat response, that isn’t usable.

Why is it that slow if you’re “just” reading data off disk? Because in a mixture-of-experts model, the router decides, per token, which experts to activate out of thousands available. Almost every new token needs a different subset of experts, and most of those subsets aren’t sitting in the RAM cache yet. The system has to hit disk on nearly every step, and reading from disk, even a fast SSD, takes orders of magnitude longer than reading from RAM.

Here’s the piece Finn’s prediction doesn’t account for. Hand colibrì 1.5TB of RAM instead of 25GB, and yes, you could keep more experts resident in memory at once, which would help. But it doesn’t fix the underlying problem, which is how much data you can move and compute per second, memory bandwidth and compute, not how much total memory you own. It’s the difference between a bigger parking garage and a wider avenue. Both help something. Only one of them clears the traffic jam.

What this actually means for your company

I’m not writing off the destination Finn describes. AI running on local hardware, without a per-call dependency on an external API, is a real direction, and there are cost, privacy, and latency reasons for companies to take it seriously. I’ve written before about why recognizing this cycle, the lab demo that works months or years before someone turns it into something boring and reliable in production, is the real advantage that experience gives you, not nostalgia for having watched the last technology cycle play out.

What I’m pushing back on is using a hardware spec sheet, one that’s still an unconfirmed rumor, to put a date on that destination. The real timeline depends on someone solving the memory bandwidth and per-token compute problem in mixture-of-experts architectures, not on Apple fitting more gigabytes into a chip. That’s the question we ask first in AI Maestro’s discovery engagement when a company brings us the idea of running AI on its own hardware: not how much memory the machine has, but what latency the operation actually needs and whether the model architecture they want to use can sustain it at that volume. You can’t answer that question by reading a rumor about a chip that doesn’t exist yet.