The $/GB framework: price any AI machine in 30 seconds

By Daniel S. · July 4, 2026

You can size up almost any AI hardware deal with two numbers: what you paid per gigabyte of usable memory, and what you paid per gigabyte-per-second of memory bandwidth (the speed proxy). Everything else — brand, generation, marketing — is noise until those two numbers are in front of you.

TL;DR

• $/GB tells you how expensive the model capacity is — divide the price by usable memory (VRAM, or unified memory on a Mac).
• $/(GB/s) tells you how expensive the speed is — divide the price by memory bandwidth. Lower is better on both.
• No single "best" machine exists — a used RTX 3090, a used Mac Studio, and an AI mini-PC appliance each win one metric and lose the other, worked out below with current prices. ⏱️ verified 2026-07-05
• Run the math before reading the marketing. A machine that looks cheap on price-per-gigabyte can still be a bad buy if its bandwidth-per-dollar is poor — that's exactly the trap the DGX Spark-class appliances set.
• The framework only prices memory and speed — it deliberately ignores software support, power draw, and form factor, covered in "When this won't help."

What is the $/GB framework, exactly?

It's two ratios, not one. For any AI machine or GPU:

$/GB = price ÷ usable memory (in GB) — cost of capacity: how much model you can afford to load.

$/(GB/s) = price ÷ memory bandwidth (in GB/s) — cost of speed: since generation speed for one user tracks bandwidth ÷ model size (see VRAM: the one number that decides which AI models you can run), bandwidth-per-dollar is a fair proxy for how much speed you bought.

Lower is better on both. The trap is that these two numbers rarely agree on which machine wins — a device can be a great capacity deal and a poor speed deal, or vice versa, which is the entire reason to run both instead of trusting a single spec.

Why not just compare prices directly?

Because raw price tells you nothing about what you got. $1,200 for 24 GB and $1,200 for 128 GB are not comparable numbers until you divide — and neither is comparable to bandwidth until you divide again. Spec-sheet shopping compares GPUs by "how powerful" they sound; the $/GB framework compares what actually determines whether your model fits and how fast it runs once it does.

It also protects you from stale pricing — the same mistake the used-GPU market is currently full of (see that $600 RTX 3090 doesn't exist). A ratio built on a six-month-old price is still wrong; always re-price the top line before trusting the ratio, which is the whole reason every number below carries today's date.

Worked example: a used RTX 3090 (GPU)

A used RTX 3090 has 24 GB of VRAM at 936 GB/s of memory bandwidth — unchanged specs, since it's not a new product. Pricing, checked today: completed eBay sales are running $1,000–1,300, with clean or blower-style cards fetching $1,300–1,600 locally (bestvaluegpu.com tracker, ⏱️ verified 2026-07-05) — consistent with the shortage-driven pricing already documented in that $600 RTX 3090 doesn't exist.

At the midpoint of the completed-sale range (~$1,150):

Read: expensive per gigabyte of capacity, but cheap per unit of speed — the 3090 buys bandwidth efficiently, capacity less so.

Worked example: a used Mac Studio (M1 Max, 32 GB)

The M1 Max Mac Studio has 32 GB of unified memory at 400 GB/s. Individual-seller marketplace listings for this exact configuration (mint condition, 512GB storage) are currently asking $1,391–1,597 (Swappa, ⏱️ verified 2026-07-05) — asking price, not completed-sale price, so treat it as the ceiling; broader tracking puts the realistic used range closer to $1,100–1,400.

At the midpoint (~$1,200):

Read: better capacity value than the 3090, but noticeably worse speed value — unified memory gives you gigabytes cheaply, but at a fraction of a discrete GPU's bandwidth.

Worked example: an AI mini-PC appliance (DGX Spark-class, 128 GB)

The ASUS Ascent GX10 — an NVIDIA GB10-based "DGX Spark" appliance — ships with 128 GB of LPDDR5x memory at 273 GB/s. The identical 1 TB storage SKU currently lists at $3,099.99 direct from Amazon and $3,997.00 direct from Newegg — the same hardware, a ~29% spread depending purely on which retailer you buy from (Amazon listing; Newegg listing, both ⏱️ verified 2026-07-05) — its own small lesson in checking the channel, not just the spec sheet.

At the Amazon price ($3,100):

Read: this is the whole point of running two numbers. On $/GB alone, the GX10 looks like the best deal on this page — cheapest capacity per gigabyte of the three. On $/(GB/s), it's the worst by a wide margin: nearly 10× more expensive per unit of speed than the 3090. A single-metric shopper sees a bargain; the two-metric framework shows a machine that holds a lot of model and delivers it slowly.

So which one should I buy?

There's no universal winner — that's the finding, not a dodge. Match the metric that matters to your actual workload:

• Chasing raw speed on models that already fit in 24 GB? The 3090's $/(GB/s) wins clearly — it's the cheapest way to buy bandwidth.
• Need to hold a large model (30B+) quietly and efficiently, speed being secondary? The Mac Studio's balance of decent $/GB and low power draw is the more livable daily-driver trade.
• Need to hold a genuinely huge model (100GB+) and speed is a distant second concern? The appliance's $/GB is real — it's the cheapest way to a large coherent memory pool, as long as you accept its bandwidth ceiling on anything that needs to feel fast.

When this won't help

• It doesn't price the software ecosystem. A cheap $/GB machine that only runs a narrow, awkward software stack (driver support, framework compatibility) can cost you more in time than a pricier, better-supported one saves you in dollars.
• It ignores power and noise. A GPU with great $/(GB/s) that pulls 350W and needs a loud fan is a different cost story for a 24/7 home server than a Mac Studio idling at a fraction of that — that difference doesn't show up in either ratio.
• It ignores form factor and expandability. A GPU needs a case, a PSU, and a motherboard with room for it; an appliance is self-contained; a Mac is fixed once you buy it. If you need to add a second GPU later, a discrete card's ratio doesn't capture that flexibility.
• It says nothing about the model you actually want to run. A great $/GB·s ratio on hardware that can't fit the model class you need is still the wrong buy — check capacity fit first (per the VRAM post), then use this framework to compare hardware that already clears that bar.
• These prices are already aging. Every number here is dated today; the shortage driving them has been rising, not falling, through 2026. Re-run the math with a fresh price before acting on it.

FAQ

What counts as "usable memory" for the $/GB calculation? VRAM on a discrete GPU, or total unified memory on Apple Silicon. Don't subtract for OS/driver overhead for this quick framework — it's a comparison tool, not a precise capacity planner.

Why divide by bandwidth instead of just comparing GB/s directly? Because bandwidth alone doesn't account for price — a card with double the bandwidth at triple the price is not automatically the better buy. Dividing price by bandwidth turns "how fast" into "how much that speed cost you," which is the number that's actually comparable across very different machines.

Does this framework work for brand-new hardware, not just used? Yes — the two ratios apply to any priced machine. New hardware just tends to have less price variance to watch for than the used market, where completed-sale price and asking price can differ sharply.

Why did the DGX Spark-class appliance look bad here if it has 128 GB? Because 128 GB of memory at only 273 GB/s of bandwidth is a wide memory pool fed by a comparatively narrow pipe. It's genuinely the cheapest way to hold a very large model, but that same model won't generate quickly — the two-metric framework is what catches that trade-off a single "128GB!" headline number hides.

Last updated 2026-07-05. Time-sensitive: every price in this article is dated 2026-07-05 and drawn from a shortage-affected, still-rising market — re-check current completed-sale prices before using these ratios to make a purchase.

Sources (verified 2026-07-05): used RTX 3090 pricing — bestvaluegpu.com price tracker; used Mac Studio M1 Max pricing — Swappa listings; ASUS Ascent GX10 pricing — Amazon, Newegg; shortage context — that $600 RTX 3090 doesn't exist.