Buy, lease, rent by the hour or subscribe: the local AI arithmetic

How much does an M5 Ultra Mac Studio lease cost?

Path	M5 Max (36GB)	M5 Ultra (96GB)	After it ends
Check what you own	$0	$0	nothing changes
Buy outright	$2,499 at Apple	$5,499 at Apple; B&H listing; Amazon search	you own it
Apple Card, 12 mo. at 0%	$208.25/mo.	$458.25/mo.	you own it
Apple Upgrade lease, 36 mo.	$48.99/mo.	$110.10/mo.	upgrade, buy it for a one-time payment or return it
Claude Max 20x or ChatGPT Pro	$200/mo.	$200/mo.	it continues
Rent a memory-matched GPU	$0.33/hr. (48GB)	$1.69/hr. (96GB)	it stops when you stop

Apple’s U.S. prices and lease rates are from its Aug. 25 announcement; the rental row is RunPod’s community rate for the nearest memory match, read Sept. 16, and it buys no machine at all. The lease runs through Klarna with credit-based approval, a machine not returned in the condition the lease requires can draw damage fees, and Apple has published rates only for base configurations, so upgraded memory — the thing an AI buyer wants — has no published lease price yet. The B&H link is its M5 Ultra product page and the Amazon link opens search results, not a specific product; confirm the chip, memory, seller and price on the page.

The four options, cheapest first

Start at $0. Check what the computer you already own runs before pricing anything else: a 16GB machine covers 8B-class models, and that check costs nothing.

If it fails, three routes remain — rent a GPU by the hour, pay a subscription, or lease or buy a machine — and the variable that separates them is not the chip or the model. It is how many hours a month you will actually use it.

The base M5 Ultra has 96GB of unified memory and 1.2TB/s of memory bandwidth. At RunPod’s $1.69 community rate for a 96GB RTX Pro 6000, the $110.10 lease payment rents 65 hours a month, a little over two hours a day. Below that, the rented card is cheaper. Above it, the lease is.

Over the lease’s 36-month term, the M5 Ultra costs $3,963.60. The subscription costs $7,200. And a buyer who’d rather own than lease can put the $2,499 M5 Max on an Apple Card at 0% for a year — $208.25 a month, about what the subscription costs, except that after month 12 the machine is paid for and the subscription is just beginning its second year.

What renting costs by the hour

RunPod and Lambda publish a rate per GPU; Vast.ai does not.

RunPod quotes two rates for most cards, and they are two products, not two prices. Its own documentation describes Secure Cloud as “T3/T4 data centers” with “High redundancy,” for “Production, sensitive data,” and Community Cloud as “Peer-to-peer providers” with “Variable” reliability, for “Cost-sensitive workloads.” The same page says RunPod “is no longer accepting new hosts for Community Cloud” while “Existing Community Cloud resources remain available.” Read the cheap column as somebody else’s machine, in somebody else’s building.

Rates below were read on RunPod’s pricing page Sept. 16, 2026, sorted by memory:

Rented GPU	VRAM	Community	Secure
RTX A5000	24GB	$0.16	$0.27
RTX 3090	24GB	$0.22	$0.50
RTX 4090	24GB	$0.34	$0.74
RTX 5090	32GB	$0.69	$0.99
RTX A6000	48GB	$0.33	$0.53
RTX 6000 Ada	48GB	$0.74	$0.84
L40S	48GB	$0.79	$1.09
A100 PCIe	80GB	$1.19	$1.59
H100 PCIe	80GB	$1.99	$2.89
H100 SXM	80GB	$2.69	$3.49
RTX Pro 6000	96GB	$1.69	$2.09
H200	141GB	$3.59	$4.59
B200	180GB	$5.98	$6.79
B300	288GB	$6.94	$7.89

Lambda’s on-demand page, read the same day, lists one rate per instance rather than two tiers, and shows some GPUs at more than one rate depending on the instance. Its page states that prices exclude applicable sales tax, VAT or GST.

Lambda on-demand instance	VRAM	Rate an hour
Quadro RTX 6000	24GB	$0.69
A10	24GB	$1.29
A6000	48GB	$1.09
A100 PCIe	40GB	$1.99
A100 SXM	80GB	$2.79
H100 PCIe	80GB	$3.29
H100 SXM	80GB	$3.99-$4.29
GH200	96GB	$2.29
B200 SXM6	180GB	$6.69-$6.99

Vast.ai publishes no fixed figure to quote. Its pricing page says “Prices are set by the market, not by Vast. More supply means lower prices — and you always see the real rate,” and it sells on-demand, interruptible and reserved instances at rates that move with supply. Interruptible instances are advertised as “50%+ cheaper” and preemptible. We quote no Vast.ai rate because there is no published rate to read.

The break-even hours

One division decides this, and you can redo it with your own numbers. Take the purchase price, divide by 36 months to put it on the same footing as a lease or a subscription, then divide that monthly figure by the hourly rate of a rented instance with comparable memory. The answer is the hours a month at which the two cost the same. A lower figure means the machine pays for itself sooner.

Route	Cost a month	Memory-matched rental	Hours at the community rate	Hours at the secure rate
Keep what you own	$0	none needed	—	—
M5 Max 36GB, leased	$48.99	RTX A6000, 48GB	148	92
M5 Max 36GB, bought at $2,499	$69.42	RTX A6000, 48GB	210	131
128GB Ryzen AI Max+ 395 box, $3,649.99	$101.39	RTX Pro 6000, 96GB	60	49
M5 Ultra 96GB, leased	$110.10	RTX Pro 6000, 96GB	65	53
M5 Ultra 96GB, bought at $5,499	$152.75	RTX Pro 6000, 96GB	90	73
Claude Max 20x or ChatGPT Pro	$200	RTX Pro 6000, 96GB	118	96

The arithmetic in full, so anyone can check it. $5,499 divided by 36 months is $152.75. Divided by $1.69 an hour, that is 90 hours a month of a rented 96GB card — three hours a day, every day, before owning the machine is the cheaper way to get 96GB. The $110.10 lease clears sooner, at 65 hours a month, a little over two hours a day. The base M5 Max is the hardest purchase to justify on these terms: $2,499 over 36 months is $69.42 a month, but that buys 210 hours of a 48GB RTX A6000 at $0.33, about seven hours a day, every day, before the machine is ahead. The 128GB Ryzen AI Max+ 395 class is the easiest: at the $3,649.99 GMKtec EVO-X2 price we recorded Sept. 15, it is $101.39 a month, which a rented 96GB card costs in about 60 hours, two hours a day.

The subscription is not exempt from the same test. $200 a month buys 118 hours of a 96GB RTX Pro 6000 at the community rate, or 100 hours of an 80GB H100 PCIe at $1.99. What the subscription buys that no rented card does is the model itself, which is the subject of the subscription section below.

Two things this table is not. It is not a speed comparison: the pairings match memory size, and we publish no measured comparison between a rented RTX Pro 6000 and an M5 Ultra, so we assert none. And it is not a flattering pairing for the purchase. The one mismatch we can state plainly runs the other way: a rented 48GB RTX A6000 has more memory a model can address than a 36GB M5 Max, which also has to hold macOS — so the $0.33-an-hour row is the more capable side of that comparison, not the weaker one. We would rather say that than pick a rental that makes buying look better.

What the table also cannot supply is your own number. Most people overestimate it. Count the hours a model is actually generating, not the hours the machine is switched on.

What renting cannot do

It cannot keep your data on your own hardware. Every prompt, file and line of code goes to someone else’s machine. On RunPod’s community tier, that machine belongs to a peer provider, which the company’s own documentation points at “Cost-sensitive workloads” rather than “Production, sensitive data.” If the reason you want a local model is that the work cannot leave the building, this whole section of the table is closed to you and the rest of the page is your decision.

It bills wall-clock time, not thinking time. RunPod’s pricing documentation says “Pods are billed by the second for compute and storage, with no fees for data ingress or egress.” A pod left running while you read its last answer is a pod you are paying for. The hours in the table above are hours the instance exists, not hours you are typing.

The weights are not free to keep. Container disk data is cleared when a pod stops, so a stopped pod means downloading the model again next session. Keeping it costs storage: RunPod publishes $0.10 per gigabyte a month for container and volume disk while a pod runs, $0.20 per gigabyte a month for a volume disk while it is stopped and $0.07 per gigabyte a month for a network volume. For gpt-oss-120b’s 65.4GB minimum, that is $4.58 a month on a network volume or $13.08 on a stopped volume disk. That division is ours, from RunPod’s published rates read Sept. 16.

It leaves you nothing at the end. This is the strongest argument for the buy branch and it is our own evidence: five of six base Mac Studio models had median used asking prices at or above their original launch prices in our Aug. 18-25 snapshot. Asks are not completed sales, and no one should plan on resale value. But 36 months of rental receipts leave a residual of exactly zero, and 36 months of Mac Studio leave a machine that this market has not been marking down.

Availability is not guaranteed. RunPod says it “is no longer accepting new hosts for Community Cloud” while existing resources remain, and Vast.ai’s cheapest tier is explicitly preemptible. The cheap rate and the reliable rate are different rows for a reason.

Setup is repeated. Each new pod starts empty: drivers, runtime, model, configuration. That one is a general characteristic of renting rather than anything a provider documents, and how much it costs you depends on whether you have scripted it.

Which local AI models fit each Mac Studio?

Unified memory is the whole game for local models: the model must fit in it. Using canirun.ai’s requirement figures, at roughly 4-bit quantization:

Memory	Machine	What fits (examples)
36GB	M5 Max base, $2,499	Qwen 3.5 27B and smaller
64GB	M5 Max step-up	Llama 3.3 70B, Qwen 3 Coder Next 80B, GLM-4.5 Air
96GB	M5 Ultra base, $5,499	GPT-OSS 120B, Mistral Small 4, Qwen 3.5 122B
256GB	M5 Ultra step-up	Qwen 3 VL 235B, Hy3 295B, MiniMax M3 428B
512GB	M5 Ultra, late Oct.	GLM-5.3 753B

Two readings of that table. The base M5 Max, for all its GPU cores, is a 27B-class machine — Llama 3.3 70B needs 39.1GB and the machine has 36, with an operating system to run. The interesting tiers start at 64GB, and the $110.10 lease’s 96GB is genuinely capable: GPT-OSS 120B and Qwen 3.5 122B are current, well-regarded open models. The architecture trend works in the Studio’s favor here — recent large models are mixtures of experts with a few billion active parameters per token, which is exactly the shape a machine with enormous memory and 1.2TB/s of bandwidth serves well. Fitting a model is not the same as running it fast; measured and estimated speeds by model size show what to expect from each tier. Before pricing either path, check what the Mac you already own runs: a 16GB machine covers 8B-class models at no cost.

The ceiling is also real. GLM-5.3, the largest open model that fits anything Apple sells, needs the 512GB machine that arrives in late October at an unannounced price. DeepSeek V4 Pro wants 820GB and Kimi K3 wants 1.4TB; no Mac runs them, and nothing Apple announced changes that inside a year.

Can a local Mac replace Claude Max 20x or ChatGPT Pro?

No. The $200 subscriptions sell the frontier: the strongest closed models, improving on someone else’s capital expenditure, with no machine to own, cool or resell. The open models in the table above are a tier below them on most benchmarks and always have been. Anyone canceling a Claude subscription expecting GPT-OSS 120B to be the same product will be disappointed inside an afternoon. Renting does not change that answer either: a rented card runs the same open weights a bought machine does.

What the local machine buys instead is a different bundle: tokens that cost nothing at the margin, prompts and code that never leave the building, no rate limits, any model you can download and a residual asset — this year’s used Studio market prices the machines it can’t restock at or above list. The honest comparison isn’t “which is smarter.” It’s whether unlimited private tokens from a 122B-class model are worth $110 a month to you while the frontier costs $200 — and whether you will use them more than the 118 hours a month that $200 rents.

Two subscriptions make the arithmetic sharper. A developer paying for both Claude Max 20x and ChatGPT Pro is at $400 a month, $14,400 over three years — the price of leasing three M5 Ultras with money left over, or buying one 256GB-class machine outright.

And there is one thing buying cannot do that renting can: reach hardware no consumer purchase holds. RunPod listed a 288GB B300 at $6.94 an hour on its community tier Sept. 16, more memory than every Mac Studio configuration except the 512GB machine that has no price yet. It does not reach far enough for DeepSeek V4 Pro or Kimi K3 on one card either; multi-GPU pods exist and we have not priced one.

When buying is the right answer

Three conditions, and one is enough.

You are over the break-even hours. More than about three hours a day of 96GB-class work, every day, and the $5,499 machine costs less than renting one. More than two hours a day and the $110.10 lease does.

The data cannot leave. No hourly rate fixes this. It is the condition that makes the whole rental column irrelevant regardless of the arithmetic.

You want the asset. A rented hour leaves nothing. Our own Mac Studio resale figures are the most favorable residual evidence we have, and they are asking prices, not sales.

If one of those applies, the decision finishes on these pages. Local AI hardware by budget places every tier from $0 up, the GPU guide matches a model to a 16GB, 24GB or 32GB card with dated used prices, and five 128GB Ryzen AI Max+ 395 machines are priced and compared for the class in the table above. If owning is the answer and the budget runs past $4,000, which machine to buy for local AI above $4,000 carries this arithmetic into the used Mac Studio market, where secondhand asks on Sept. 15 were frequently not a discount.

When will the 512GB M5 Ultra ship?

Late October; its price is unannounced. Not scheduled: anything about which models exist a year from now. This year’s open releases populated every memory tier Apple sells; whether next year’s do the same is not knowable from here, and we won’t pretend otherwise. What we can do is rerun the fit table when the 512GB machine gets a price — and rerun the lease math if those “from” rates move. Hourly rental rates move faster than any of that, so re-read the provider pages before you use the break-even table; we rerun them when the 512GB machine is priced, and sooner if these rates move enough to change a row. For the broader hardware comparison, see local AI memory prices per gigabyte and by bandwidth.