M5 Ultra Mac Studio Is Official: Specs, Pricing, and Pre-Orders
Quick Verdict: Apple's official specs are out, pre-orders are open, and the base $5,499 / 96GB configuration ships September 22. If you want the full 512GB memory ceiling, you're waiting until late October regardless of when you order — Apple built the staggered rollout our earlier RAM-ceiling analysis predicted almost exactly. We've placed an order ourselves; full hands-on testing is coming when it arrives.
The Official Specs
| M5 Max (base Studio) | M5 Ultra | |
|---|---|---|
| Starting price | $2,499 | $5,499 |
| Base memory | 36GB | 96GB |
| Base storage | 512GB | 1TB |
| Max memory | — | 512GB (256GB available at launch; 512GB ships late October) |
| Memory bandwidth | 614GB/s | 1.2TB/s |
| CPU cores | — | up to 36-core |
| GPU cores | — | up to 80-core |
| Neural Engine | — | 32-core |
| Storage range | 512GB – 16TB | 512GB – 16TB |
| Ports | 4x Thunderbolt 5, Ethernet, USB-A | Same |
| Display support | Up to 8 displays | Up to 8 displays |
Apple's own performance figures put the M5 Ultra at up to 2.4x the CPU performance and 4.7x the GPU performance of the original M1 Ultra, with AI workloads specifically cited at up to 9.8x faster. The M5 Max's memory bandwidth also jumped from 546GB/s on the prior generation to 614GB/s.
On the AI side specifically — the part we've been tracking most closely across this whole series — Apple's messaging leans directly into on-device LLM deployment and agentic workflows running locally without cloud dependency. That's not just marketing language layered onto existing hardware; it lines up with what we flagged as a likely "Agentic Accelerator"-class capability back when we were only working from rumors.
How This Compares to What We (and Everyone Else) Predicted
Worth being straight about this rather than just declaring victory: back in June, MacRumors reported Apple had tested M5 Ultra configurations up to 768GB of unified memory. That number never materialized. The actual shipping ceiling is 512GB — a real reduction from what was floated, and MacRumors' own coverage of the official announcement calls it out as "a significant reduction from earlier expectations."
Our own read at the time was more conservative than the 768GB rumor, and it landed close to what actually shipped:
- We predicted: "a shipping ceiling somewhere in the 256GB–384GB range feels far more plausible than the full 768GB Apple reportedly tested." Reality: 256GB is what's actually available at launch. The 512GB tier exists on paper but doesn't ship until late October — effectively confirming the staggered, supply-constrained rollout we expected, even if the eventual ceiling (512GB) lands a bit above our range once that October tier arrives.
- We predicted: the base 96GB configuration was "the safer prediction — it's stayed constant through two generations." Reality: correct. Still 96GB base, unchanged.
- We predicted: memory bandwidth rumored at 1,000–1,100GB/s. Reality: 1.2TB/s (1,200GB/s) — a bit higher than the rumor, but the right order of magnitude.
Not a perfect call, but closer than the 768GB headline number that was floating around all summer. The pattern we flagged back then — Apple engineering for a higher ceiling than it can reliably supply, same as the M3 Ultra's repeated pullbacks — held up.
Pricing Reality Check
$5,499 for the base M5 Ultra is a $200 increase over the outgoing M3 Ultra's $5,299 starting price for the equivalent 96GB configuration — a modest bump given everything we've covered about DRAM pricing over the past few weeks. If you want to go past the 256GB tier available now, you're looking at a wait either way: the 512GB configuration isn't orderable for delivery before late October no matter when you place the order today.
We Ordered One
We placed an order for the M5 Ultra with 256GB of unified memory. Current estimated delivery: late November — noticeably later than Apple's September 22 general availability date for the tier, which tells you something about real-world fulfillment lead times that the official press materials won't. If you're planning around a specific delivery window, budget more runway than the headline ship date suggests.
When it arrives, we'll run it through the same local-AI-workstation workloads we've been covering in this series — agentic coding tools, local model inference, the works — and put real numbers against the DIY dual-GPU Linux comparison we published back when this was still speculation. Full hands-on review coming once the unit is actually on the desk, not before.
Sources: Apple — Mac Studio, MacRumors — Mac Studio