The 512GB Option That Lost $17,000 Overnight — And What It Teaches You About Your Next Mac Purchase

The 512 GB Option That Lost $17,000 Overnight — and What It Teaches You About Your Next Mac

On August 25, 2026, Apple unveiled the M5 Ultra. The 512 GB config? That doesn't ship until late October. Four days later, the used price of a machine that had been treated as an unaffordable collector's item collapsed from $20,000 to $3,000. Not because the hardware got worse — but because the scarcity that had been propping the price up evaporated.

This story isn't an aside about price collapse. It's the clearest demonstration of a rule that should dictate every hardware purchase decision in 2026: You don't buy compute. You buy an option on scarcity — and exactly one capability decides whether it's worth it: holding the model you actually want to run.

The 4×-Multiplier Trick Designed to Fool You

Apple markets the M5 Ultra with "up to 4.3× faster AI performance." That's not a lie — but it's a number that measures the wrong moment. Apple's own ML team measured this in "Exploring LLMs with MLX and the Neural Accelerators in the M5 GPU":

The 4× figure describes the moment before the first token. It says nothing about how fast the model writes afterward — and that's precisely what you feel in day-to-day use.

Prompt processing scales with compute. Generation scales with bandwidth.

That's why the real question isn't "how many TFLOPS?" but: "What's my bottleneck — and which model do I want to hold?"

The One Number That Actually Matters: Bandwidth

Bandwidth is what determines everything you feel when generating. And here is where the M5 generation actually changes something:

From M4 Max (546 GB/s) to M5 Max (614 GB/s) is +12% — same league, new packaging. From M5 Max to M5 Ultra is +95% — a genuine class jump. If you're choosing between the M5 Pro and the M5 Max, you're buying the same league twice. Go to the M5 Ultra and you change leagues.

Unified Memory is both a feature and a tax. macOS reserves 8–12 GB before your model even loads. Every browser tab, every userspace tool, draws from the same pool. When you read "64 GB," you should read "~54 GB for the model."

The DRAM Crisis as a Price Machine

Why are prices exploding at all? Because memory is the scarcest resource in the tech industry in 2026. Tim Cook called it a "100-year flood" crisis. The numbers:

• iPhone 17 Pro memory module: from ~$25–29 to ~$70 (+230%)
• Gartner 2026 forecast: DRAM +125%, NAND +234% over the year
• Micron: HBM sold out for 2026, meeting only 50–66% of demand
• Market size: from ~$220B (2025) to ~$890B (end of 2026) — AI inference is eating the gap

This isn't a capacity problem; it's an allocation problem: the fabs are migrating to the higher-margin HBM chips. The exact same force that's making the GB10 boxes more expensive is what hit the 512 GB option.

The $50 Collision

And then there's the calculation nobody has on their radar: An M5 Ultra with 256 GB/1 TB costs $9,499. Two DGX Spark bundles together cost $9,449 — a difference of 50 dollars.

But 2×128 GB isn't 1×256 GB. Two Spark nodes shard the work: a single 70B token takes 269 ms over the network bridge, or 133 ms with two nodes — in exchange for two operating systems, two network interfaces, and interconnect latency. A 256 GB Mac holds the model in a single memory pool.

2×128 ≠ 1×256.

The Lesson

If you paid $20,000 for a 512 GB machine a year ago, you weren't buying compute. You were buying an option on scarcity. Apple closed that option in March, reopened it in August — and the value of the old option evaporated. The hardware performed exactly the same. It was the premium that died.

Before you buy now, answer three questions:

1. Which model do I really want to hold? — the capacity ladder, not wishful thinking.
2. Do I read more or write more? — prefill is compute-dominant, decode is bandwidth-dominant.
3. Can I shard my workload? — if yes, 2× Spark can be the smarter math; if no, the memory pool wins.

Buy for your bottleneck, not for the marketing number. It isn't the chip name that decides — it's the single capability that counts: holding the model you want to run.

FAQ

Why is Apple's 4× figure misleading even though it's accurate? That figure measures prompt processing (time-to-first-token), which scales with compute. Generation — the part you actually feel while writing — scales with memory bandwidth and improves only 19–27%. Apple isn't lying; the number just measures the moment before the first token, not the writing experience after it. If you anchor on the 4× figure, you're measuring the wrong bottleneck.

Why is 512 GB of Unified Memory so special? Because it's the only way to hold a 700B-class model on a single machine. 256 GB covers the 300B class, but Kimi K3 (2.8 trillion parameters) doesn't fit even in 512 GB — the ceiling is real. Memory is therefore the decisive bottleneck, not compute. That's exactly why RAM size determines which model class you can realistically run.

Is the M5 Ultra worth the premium over the M5 Max? Only if you genuinely want to push the 700B class or very long contexts. The bandwidth jump from 614 to 1,200 GB/s (+95%) is the only true class jump in this generation — everything below it is the same league. For a pure decode workload on 70B models, the M5 Max 128 GB is the more sensible buy. The premium only pays off if you actually need the capacity or the bandwidth.

When does 2× DGX Spark make sense over 1× M5 Ultra? When your workload is naturally shardable: CUDA clusters, multiple parallel jobs where the interconnect latency is tolerable. At a $50 price difference, it's an architecture question, not a budget one. A single 70B token costs 269 ms across the network bridge — or 133 ms with sharding, but with two operating systems. If, on the other hand, you want to hold a single large model in one memory pool, the M5 Ultra wins.

What does the DRAM crisis have to do with my purchase? It's the price machine behind all of it. With DRAM +125% and NAND +234% over the year, the cost of every bit of memory rises — and the scarcity is an allocation issue, not a capacity one. The fabs are migrating to higher-margin HBM chips, so commodity DRAM stays tight. If you're buying now, treat memory as the most expensive and scarcest part of the machine and prioritize RAM size.

Should I buy now or wait for the 512 GB version? It depends on your bottleneck: if you want the 700B class, waiting for the 512 GB config (late October) is worth it. For most use cases (up to the 300B class), the 256 GB version is the right choice. If you have shardable workloads, you can start today with 2× Spark. The DRAM crisis makes memory expensive — but it also makes it precisely what actually matters.

Sources

• Apple ML Research: "Exploring LLMs with MLX and the Neural Accelerators in the M5 GPU" (machinelearning.apple.com)
• Apple Newsroom, 25.08.2026: M5 Max/M5 Ultra announcement, "up to 4.3x faster AI performance"
• Apple product specifications: bandwidth, Unified Memory configurations
• Gartner DRAM/NAND 2026 forecast · Micron HBM sellout · Tim Cook memory comments (July 2026)
• Ars Technica / Tom's Hardware (March 2026: 512 GB option removed)
• wccftech / OnMSFT / Macfax / MacRumors (M3 Ultra 512 GB resale market)
• ComputerBase (independent DGX Spark measurements) · NVIDIA forum / heise / overclock3d (GB10 price trajectory)
• Yume_X (29.08.2026, price observation) · own verification 31.08.2026