Mac Studio M5 Ultra · Vienna · dedicated monthly hosting

What one box can
actually earn

Four SKUs between €6,599 and €12,429. The hardware is the easy part — the spread between the cheapest and the most valuable thing you can rent it as is bigger than the spread between the SKUs.

Commodity Mac cloud€420–790Per month. Competing head-on with MacStadium and MacWeb. 17–30 month payback.
Private LLM box€900–1,500Sold on data sovereignty, not tokens/sec. 8–13 month payback.
Creative / editor seat€650–1,000Plus storage, which is where the real margin is. 10–22 month payback.
Attached storage€10–15/TBNever sold on its own. An attach to a box already let — which is the whole catch.
The model

Run your own numbers

Everything below recalculates live. Prices are net of VAT — you reclaim the 20%, so a €12,429 box is a €10,358 asset. Defaults assume you host it yourself on a business line rather than paying colocation.

Monthly rent—
Occupancy—

Share of months the box is actually let. 85% ≈ six weeks empty a year.

Attached storage—

Only ever sold with a box. Start at 48 TB, expand out of revenue.

Price per TB—

—

Fixed opex—

Power ≈ €22 · connectivity ≈ €60 · support & monitoring ≈ €50 · AppleCare ≈ €8.

Payback——
Net cash / month—after all running costs
Profit over 3 years——
Break-even rent—below this you lose money

Monthly ledger

Payback by customer type — same box, same occupancy

Bars capped at 48 months. Storage add-on included where set.

The four SKUs

Only one of the two upgrades is worth buying

Token generation on these chips is bound by memory bandwidth, and all four Ultra configs share the same 1.2 TB/s. The extra 16 GPU cores buy you prompt processing, video encode and 3D render — not tokens per second. The RAM buys you models the smaller box physically cannot load.

ConfigCPU / GPUMemoryGrossNet of VATWhat it's for
Ultra base30c / 64c96 GB€6,599€5,499Entry ticket. Runs a 27–32B model with a generous context window, or a 70B at Q4. Cheapest way to test whether anyone actually buys.
GPU upgrade36c / 80c96 GB€8,029€6,691€1,430 for render throughput. Pays off only if you sell it to editors and colourists, where export time is the product.
RAM upgradeAI pick30c / 64c256 GB€10,999€9,166The only config that unlocks a different customer: 120B-class models, or a 235B MoE at Q4, or four 27B instances side by side. Nothing cheaper on the market does this in one box.
Both upgrades36c / 80c256 GB€12,429€10,358You're paying for both upgrades when no single customer needs both. Buy it only if you'll genuinely time-share the box between AI tenants and render work.

The practical read: two boxes at €10,999 and €8,029 beat one at €12,429 twice over. You get two revenue lines, two failure domains, and each machine is aimed at a customer who values exactly the upgrade it has. The €12,429 SKU is the one Apple wants you to buy; it is not the one that rents best.

Three markets

The same hardware, priced three ways

What you can ask depends almost entirely on what the buyer is comparing you against. Against a Mac hosting provider you are a commodity. Against a GDPR problem you are a solution.

Mac cloud & CI

€420–790

Build farms, iOS CI runners, Xcode fleets. Sticky annual contracts, high occupancy — and a race to the bottom on price.

Compare
MacStadium M2 Ultra 128 GB at $449/mo; MacWeb M3 Ultra 256 GB at $799/mo
Upside
Predictable, low-touch, fills the box while you find better customers
Risk
You cannot out-price American providers at scale. This is a floor, not a business.

Private LLM box

€900–1,500

An EU company that wants a model it chose, running on iron in Austria, with nothing leaving the building. You are not selling compute — you are selling a defensible answer to a compliance question.

Compare
A 20-seat ChatGPT Team subscription is ~€600/mo and the data leaves
Throughput
~45–60 tok/s single stream on a 27–32B at Q4; roughly 150–250 tok/s aggregate across 4–8 concurrent users
Risk
One tenant per box — the model pins the memory. No overselling.

Creative seat

€650–1,000

Editors, colourists, render offload. Weaker on its own, stickier once their footage lives on your array — though the array only ever follows the box, never leads it.

Compare
Renderro: €49/mo holding fee plus $1.99–9.99/hr; Vagon from $0.25/hr
Reality
macOS gives you one console session. You schedule editors sequentially, you do not run them concurrently.
Risk
Most editors want the box on their desk. Sell overflow and render, not primary edit.
Your NAS question, corrected

Storage is an attach, not a business

Hetzner sells 20 TB for €40.60 a month. LucidLink sells the same terabyte for roughly forty times that, and post houses pay it. Both companies are fine. So the market does not price storage on €/TB at all — and the only question that matters is which of those two you're standing next to.

Against Hetzner you lose, completely, and you should. Their box is SFTP and SMB across the public internet to Falkenstein, capped at ten concurrent connections. It is an excellent vault and a hopeless working volume — you cannot scrub a 4K timeline off it. If the customer wants a bucket, send them there and keep the phone call short.

What you actually sell is not storage, it's locality: media on the same 10/25GbE segment as the Mac the editor is already renting from you. Hetzner structurally cannot sell that, because their compute isn't your box. That's why €10–15/TB is a defensible number sitting between a €2 bucket and a €74 filesystem — but only ever as a line on an invoice that already has a machine on it.

Which means the NAS has no independent demand. Nobody in Vienna wakes up wanting to buy 8 TB from a man with a Synology. You never sell it standalone, you never market it, and its entire economics are derivative of Mac occupancy. Empty boxes mean an empty array — this is correlated risk, not diversification. Buy the array after the second box is let, populate four bays for around €2,500, and expand out of revenue rather than out of optimism.

Two tiers, and let Hetzner do half the work. Working storage — RAID6, explicitly not backed up, client keeps their camera originals — at €10–12/TB, where your marginal cost is nearly nothing. Vault storage at €14–18/TB with the cold copy sitting on a Hetzner Storage Box at €2/TB. Don't compete with them on the tier where you're weakest; resell them and keep the margin.

Two operational facts to plan around. Ingest: 40 TB over a business fibre line takes days, so the first delivery arrives on physical drives and you charge €150–300 to take it in; after that only proxies cross the wire. Backup: the day you hold a client's only copy of a shoot and lose it, the business is over — which is exactly what the €2/TB cold copy is for. On wiring, Thunderbolt 5 gives you 120 Gb/s direct-attach for the hot project; the Mac Studio's built-in 10GbE is fine for one box and tight the moment two machines pull from the same array.

Before you commit

Six things that change the answer

Who is the provider of an uncensored model
If you operate an abliterated model as a service — your endpoint, your system prompt, your model choice — you are plausibly the provider under the EU AI Act and DSA, and you own what it emits. If you rent a bare box and the customer loads their own weights, you are infrastructure. The revenue is similar; the liability is not remotely similar. Structure it as the second one, and put it in the contract.
One tenant, one model
A loaded model pins unified memory for as long as it is served. You cannot oversubscribe an AI box the way you oversubscribe a web server, and you cannot run two tenants' models at once without halving the context each of them gets. Every AI box is single-tenant. That is what justifies the price, and also what caps it.
Occupancy is the whole game
At 100% the private-LLM 256 GB box pays back in under eight months. At 65% it takes thirteen. Nothing else in the model moves the answer that much — not the SKU, not the electricity, not the rent. Sell the first box before you buy the second.
This needs six customers, not a funnel
Two AI boxes and a couple of edit seats fills everything you'd buy in year one. That is not a marketing problem with a CAC and a conversion rate — it is forty phone calls into a list you can write down. There are perhaps 50–80 post houses and working colourists in Austria, and the sovereign-AI buyers are the same SME crowd already in reach through KontorBund and the aws/FFG ecosystem. If six of them can't be found by talking to people you nearly know, that isn't a storage problem or a pricing problem — it means the venture has no market, and it is much cheaper to learn that before €11,000 leaves the account.
The tax side is worth a phone call
The Investitionsfreibetrag sits at 20% through December 2026, up from the usual 10% — on a €10,358 net asset that's a €2,072 extra deduction. But it requires a useful life of at least four years, and computer hardware is often written down over three. Worth a specific question to your Steuerberater before the invoice date, not after. Given the sovereign-AI track you're already pitching to aws and FFG, the hardware may also be a fundable line item rather than a pure capex hit.
Apple silicon holds value — use it
The model assumes 45% residual after three years, which is conservative for an Ultra. That residual is roughly a third of the total return. It also gives you a real exit: if occupancy never materialises, you sell the box at a modest loss rather than being stuck with depreciated commodity GPUs nobody wants. The downside here is unusually well-protected.