RAM prices are spiking. Get more AI out of the Mac you already own.

Memory is getting more expensive, and the headlines blame an AI-driven shortage squeezing PC builders. If you've been wanting to run real AI on your own machine, the obvious move just got expensive. So here's the contrarian take: don't buy the RAM. The amount of intelligence you can run locally is set far more by the software than by how many gigabytes you bolt on, and that's where there's room to win without spending a cent on hardware.

The price spike is real, and it's pointed at memory

Memory got pulled into the AI buildout. The same demand filling data centers tightens supply for the sticks that go in a desktop, and consumer prices follow. I'm not forecasting prices, but the direction is the point: the thing everyone tells you to buy first when you want to run bigger models is the thing that just spiked. Upgrading into a shortage means paying a premium for the privilege of brute-forcing a problem that has a software answer.

Why "just add RAM" is backwards right now

The standard advice treats RAM as the wall: a 70-billion-parameter model needs X gigabytes, a bigger one needs more, so size your machine to the model. That math assumes every weight has to sit in memory at the same time. It doesn't. A mixture-of-experts model only activates a fraction of its parameters for any given token, so most of those weights are idle most of the time. Paying full RAM price to keep them all resident is paying for capacity you aren't using on that token.

Buying more memory also solves exactly one problem and creates a few. It's a one-time spend at a bad price, it's capped by what your Mac's chip actually supports, and it does nothing about the recurring cost most people are actually trying to escape: the cloud subscription. Several outlets spent late May 2026 on that exact theme: Axios ran "AI sticker shock," another piece argued "the solution might be cancelling my AI subscription," and a widely shared post described "AI subscriptions are a ticking time bomb for enterprise." The answer to expensive AI isn't more expensive hardware on top of an expensive subscription. It's getting more out of what you already have.

Capability-per-gigabyte, with the numbers

Outlier's patent-pending paged inference engine streams expert weights from your SSD instead of holding the whole model in RAM. The headline number: a 397-billion-parameter model carries about 209 GB of weights, and it runs at roughly 14 GB peak resident memory on a 64 GB Mac Studio. The full model lineup spans about 2.4 GB to 209 GB on disk, but disk is cheap and RAM is the expensive, spiking part, so the engine spends the cheap resource to conserve the dear one.

This is not magic, and it shouldn't be sold as such. It's four ordinary ideas stacked: mixture-of-experts so only the needed weights are touched, 4-bit quantization so each weight is smaller, SSD streaming so idle experts live on disk, and Apple's unified memory so the chip and the engine share one fast pool. Each piece is real engineering with real trade-offs, and the deeper mechanics are in how a 397B model runs on consumer hardware and what paged MoE inference is. The trade-offs are speed and ceiling: Core 27B runs about 20.7 tok/s on an M1 Ultra, slower than cloud flagships, and cloud models are stronger on the hardest work. You're trading some speed and top-end strength for never renting a model and never paying the shortage tax. The weights themselves are open and inspectable on HuggingFace: Nano, Lite, Quick and Core in Outlier's own repos, Plus and Vision 3.8 as public community conversions of Qwen's models.

What your current Mac already runs

Start with what's free. Nano and Lite run locally and privately on a 16 GB Apple Silicon Mac, no account, no purchase. For a lot of daily work, that's the whole job. Nano scored 52.0% on a 200-question MMLU sample (build 1.11.804), so keep it to light drafting and quick questions. From there the mid tiers (Core 27B, Vision 3.8 27B) run on common configurations, and we published a 54-prompt side-by-side of Core 27B and Claude Opus so you can compare the answers yourself. The 397B Plus tier is the one that genuinely wants a 64 GB machine. The honest map of which tier fits which Mac is in how much RAM you need for local AI. Notice what's missing from that list: a trip to buy memory you don't have to buy. When RAM gets expensive, the cheapest upgrade is software that already knows how to do more with the gigabytes you've got.

Frequently asked questions