Is the GPU dedicated, or is it shared?

Dedicated. One physical machine, one account, and the card passed straight through to your operating system. No hypervisor, no MIG partition, no vGPU profile, no time-slicing scheduler, and nobody else's workload on the card. What you measure on the first afternoon is what you keep.

Can I have more than one card in one machine?

Yes, and it is a build we quote rather than a box you tick. The number that fits depends on the chassis, the physical width of the card and the power budget, so the honest answer needs the specific card. Ask, and the reply names the chassis, the count and the lead time. Several customers run two, and the commonest reason is memory rather than speed: two cards of 48 GB hold a model that one of 80 GB does not.

Which driver and CUDA version do I get on a Primcast GPU server?

Whichever you install. The machine arrives with a clean operating system and root, and you pin the driver, the CUDA or ROCm release and the framework build to whatever your code was tested against; nothing upgrades underneath you. If you would rather the driver were installed before handover, say so on the order and name the version.

What if the card I want is not in stock?

It is fitted or ordered, and you are told which before you pay. A card we hold goes in within four to twenty-four hours. A card we have to buy is five to ten working days, and a current-generation accelerator on allocation can be ten to fifteen. Machines already racked and ready today are listed at https://primcast.com/instant.

Can I add or change the card later?

Yes, on the machine you already have: no migration, no reinstall, no new address. It is a scheduled visit to the rack and a short outage while the machine is down, and the monthly figure changes by the difference between the two cards. Memory, disks and port speed upgrade the same way.

What happens when a graphics card fails?

It is replaced, and the agreement puts a time on it: four hours from the moment it is reported by telephone or live chat. Past that the customer is owed service credit on a published schedule rather than on request.

Is bandwidth metered? What does egress cost on a GPU server?

Nothing is metered and egress costs nothing, at every port speed. There is no transfer allowance to exceed and no overage line on any invoice. For GPU work this is usually the largest hidden cost elsewhere: a training run pulling a dataset in and pushing checkpoints out, or an inference endpoint answering real traffic, moves a great many bytes, and per-gigabyte egress is charged on every one of them.

How does a monthly GPU server compare to renting a GPU by the hour?

Arithmetic decides it. An hourly card is cheaper if the work is bursty and genuinely stops, up to a few hundred hours a month. Past roughly two thirds of the month it is not close, and that is before egress, before storage that has to persist between runs, and before time spent queued for capacity. Monthly here also means the same machine, the same disks and the same address next month.

Can I mine, or run blockchain and decentralized-compute nodes?

Yes. Both are ordinary workloads here and always have been, and several of the cards in the catalog were bought for exactly that. https://primcast.com/mining is the page for the mining case specifically.

How do I pay for a GPU server, and is there a contract?

Card, PayPal, ACH, bank wire, or cryptocurrency. There is no contract: one month is the default and only term, three, six and twelve-month cycles take a discount, and the price does not step up at renewal. For a larger first order a wire or a stablecoin usually clears faster than a card.

Can I get a refund if the card is not what I needed?

Within three calendar days of first activation, once troubleshooting has been exhausted. The exact wording is at https://primcast.com/refund-policy-us. The sizing and bus-generation sections of the page exist to make that unnecessary: it is cheaper for both sides if the right card is bought the first time.

Do you have the newest graphics card?

Sometimes, and it is usually the wrong question. The newest part is a scarce part on allocation, on a platform that will not give it the bus generation it was designed for, at a price that assumes every bit of it is needed. Most GPU work is not training a frontier model: it is serving one that already fits, encoding video, rendering frames or driving desktops, and for all of those the right card is often two generations back and a quarter of the money.

Graphics cards on bare metal · Primcast LLC · since 2004

Tell it what the work is. It tells you the card.

Most of the difficulty in buying a GPU server is not the buying. It is knowing which card is enough — and every catalog in this industry is arranged by how impressive the part is rather than by what you are trying to run. This one is arranged the other way round. Say what the job is and how much card memory it needs, and what is left is the answer, priced complete.

Find the card Is the card mine?

One tenant per machine. The card is passed through to your operating system — not sliced, not time-shared, not virtualized.

The bench

Two questions.

Nothing is reserved, no account is created, and the figure beside each card is the whole server with that card in it — processor, memory, mirrored disks, unmetered port, one address — not the card on its own. The card’s own share is printed under it so you can see which half of the money is which.

What is it for?

How much card memory?

anygigabytes or more

5 GB7-8 billion parameters, 4-bit
8 GB13-14 billion parameters, 4-bit
19 GB30-34 billion parameters, 4-bit
38 GB70 billion parameters, 4-bit

Estimates from the same catalog the checkout bills from, not quotes. Where a memory figure is missing we could not find one published by the manufacturer, so the card says so rather than guessing — and those cards drop out as soon as you move the memory control, because filtering on a number we do not have would quietly hide parts that might suit you. Ask and we will read it off the card. Several parts fit more than one chassis at more than one price; the figure here is the cheapest way to have that card, and the configurator shows the rest.

What is behind that list

56Cards we fit
$20.79Cheapest card, per month
$91.72Cheapest whole machine with one in it
5Cities
0Bytes counted

Counted from the order catalog when this page was served, not typed in. If a card is withdrawn tomorrow this strip is smaller tomorrow. What is standing in a rack this morning is a different question, and instant servers is where it is answered.

The first question everybody asks

Is the card mine, or am I sharing it?

Yours. The whole card, the whole machine, nobody else on it.

The card sits in a PCIe slot on a physical server rented to one account, and it is handed to your operating system directly: your own driver, your own CUDA or ROCm, your own processes. There is no hypervisor between you and it, no MIG partition, no time-slicing scheduler deciding when your turn is, and no second tenant whose batch job makes your inference latency move at four in the afternoon.

The rest of the machine is yours on the same terms — every core, all the memory, all the disks, root, and the server’s own HPE iLO or Dell iDRAC for console, virtual media and power. If a driver takes the network down you can still see the screen.

Not a sliceNo MIG, no vGPU, no fractional allocation. One card, one tenant.
Not a queueNothing preempts your process. The card is idle when you are idle.
Not a containerBare metal. Load your own kernel modules, pin your own driver version.
Not meteredWeights in, checkpoints out, renders out. Nothing counts the bytes.

Why the figures above look like that

The card is a line on the invoice, not a product tier.

Everywhere else a GPU is a tier: you rent “an H100 instance” and somebody else has already decided what the processor, the memory, the disks and the network around it should cost you. Here the two are separate lines that get added — which is why a card at either end of the list above goes into the same chassis, and you pay the difference between the cards and nothing else.

The machineIntel Xeon E5-2600 v1/v2Intel Xeon E5-2630L Hex Core 2.00 GHz · 32 GB DDR3 · 2 × SATA 500 GB (RAID 1) · 500 Mbps$70.93
The cardAnything on the shelves belowCheapest to dearest, $20.79 to $1,599$20.79
The portUnmetered, both directionsNo allowance to exceed, no egress line, everincluded
SetupNothingNo contract; one month is the default term—

Whole machine, card in it, per month, from$91.72

And at the other end of the same list: H100 80GB in a whole machine for $1,682.67 a month — $2.31 an hour if you are comparing against something billed by the hour, with nothing added for the traffic.

The configurator rounds a per-disk price and lands a cent or two off these figures. Longer billing cycles take a further discount; pricing has everything in one place, and older hardware, lower price is the argument for why the bottom of this list exists at all.

How much card memory

The only specification that decides whether it works.

Clock speeds and core counts decide how fast a model runs. Card memory decides whether it runs at all — the weights either fit or they do not, and a card two gigabytes short is not slower, it is a crash. So this is the number to settle before any other, and it is the one almost nobody publishes next to a price.

The arithmetic is parameters multiplied by bytes per parameter. Sixteen-bit weights are two bytes each, eight-bit are one, four-bit are a half. Quantizing to four bits is what turns a machine that cannot hold a model into one that can, at some cost in quality.

Leave head-room. The figures below are the weights and nothing else. The runtime, the activations and the key-value cache for a long context all live in the same memory, and the cache grows with every token in the conversation. A model whose weights are 38 GB does not serve comfortably on a 40 GB card. Budget a quarter to a half again on top, or ask us and we will size it against what you are actually running.

Weights only. Add head-room for the runtime and the context window before choosing a card.
Model	16-bit	8-bit	4-bit
7-8 billion parameters	16 GB	8 GB	5 GB
13-14 billion parameters	28 GB	14 GB	8 GB
30-34 billion parameters	68 GB	34 GB	19 GB
70 billion parameters	140 GB	70 GB	38 GB

What is underneath the card

Four chassis take one, and they are not interchangeable.

Each takes a different part of the catalog and gives a card a different link to the processor. That last column is here because it costs people money and nobody prints it: a customer moved from one generation of consumer card to the next on a third-generation bus and rendered slower, and worked out why before we did. So it is on the page.

Intel Xeon E5-2600 v1/v2

$70.93the machine, per month

Intel Xeon E5-2630L Hex Core 2.00 GHz · 32 GB DDR3 · 2 × SATA 500 GB (RAID 1) · 500 Mbps

Cards it takes: 28
Link to the card: PCIe 3.0
Lanes: 40 lanes per socket
Sockets: 2
Memory ceiling: 256 GB

Configure this machine →

Intel Xeon E5-2600 v3/v4

$73.47the machine, per month

Intel Xeon E5-2620 v4 Octo Core 2.10 GHz · 32 GB DDR4 · 2 × SATA 500 GB (RAID 1) · 1 Gbps

Cards it takes: 43
Link to the card: PCIe 3.0
Lanes: 40 lanes per socket
Sockets: 2
Memory ceiling: 1 TB

Configure this machine →

Intel Xeon Silver / Gold

$83.67the machine, per month

Intel Xeon Silver 4110 8 Core 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps

Cards it takes: 47
Link to the card: PCIe 3.0
Lanes: 48 lanes per socket
Sockets: 2
Memory ceiling: 1 TB

Configure this machine →

AMD EPYC

$217.59the machine, per month

AMD EPYC 7413 24 CORE 2.65 GHz 128MB L3 CACHE · 32 GB DDR4 · 1 × SATA-SSD 240 GB · 1 Gbps

Cards it takes: 10
Link to the card: PCIe 4.0
Lanes: 128 lanes per socket
Sockets: 2
Memory ceiling: 1 TB

Configure this machine →

What the bus generation actually changes

A card designed for a newer bus works perfectly well on an older one; it simply cannot move data across the link as fast. Whether that matters depends entirely on the job.

It matters when data crosses the bus continuously — a training loop streaming batches from disk or system memory, a render feeding scene data card-side every frame, anything that spills out of card memory.
It does not matter once the weights are resident and stay there — inference on a model that fits, video transcode, most simulation. The link carries requests and results, and those are small.
Ask if you are unsure. Tell us the workload and we will say plainly whether the platform is a bottleneck for it, including when the honest answer is that a cheaper card on a better link beats a dearer card on a worse one.

What comes with it

Everything that makes the card usable, and nothing you have to buy.

Root, and the driver stack is yours

We install a clean operating system and hand you the keys. You choose the driver version, the CUDA or ROCm release and the framework build, and you pin them — which is the point of bare metal and the opposite of a managed runtime that upgrades underneath you the week before a deadline. If you would rather we installed the driver before handover, say so on the order and we will.
The machine’s own console, on a separate network

HPE iLO or Dell iDRAC: keyboard-and-screen console, virtual media, power control, hardware health and the hardware log. A graphics driver that takes the display or the network with it is an inconvenience rather than a support ticket, because you can still watch the machine boot.
An unmetered port, both directions

There is no transfer allowance to exceed and no egress line on any invoice. Datasets in, checkpoints out, renders out, a model served to real users — all of it moves for the price of the port and not a byte more. Unmetered bandwidth is the ladder above the included speed, and it is the single biggest difference between this and a bill from a hyperscaler.
Reinstall it as often as you like

Operating system images in one click, reinstallable without a ticket and without a charge, so a driver experiment that ends badly costs you twenty minutes. Reverse DNS is yours to edit, IPv6 is included, and more IPv4 is a catalog line rather than a negotiation.
People, at any hour

Telephone, live chat and tickets, included on every plan, and a failed component replaced within four hours of being reported by phone or chat. Support has the response times and the credit schedule in writing.
Five rooms to put it in

New York, Miami, San Francisco, Amsterdam and Bucharest — and where the machine sits decides your latency to whoever is calling it and which data-protection regime it falls under. Data centers describes the buildings.

How soon it exists

Between half an hour and three weeks, and which one depends on where the card is standing.

Not one number, because it has never been one number. A card already fitted to a machine already racked is a handover. A card that has to be bought is a purchase order. Telling you which of those you are asking for — before you pay rather than after — is the only version of this that is worth anything.

~30 minutes
The card is already in a machine on the shelf
Nothing is fitted and nothing is moved. The machine is imaged and handed over. Instant servers is the list of what is standing there right now.
4–24 hours
A card we hold gets fitted, or a racked machine gets rearranged
The ordinary case for everything on this page. The part is in our own store room; a technician fits it, builds the array and installs the operating system. A machine on the shelf that is close but not exact — memory added, a drive swapped, the array rebuilt around your choice — is the same work, done in place without the machine leaving the rack.
5–10 working days
The card has to be bought in
Something on the list is not stock we carry. We order it, then build. Build to order is the page for a whole specification of this kind.
10–15 working days
The card is scarce
Current-generation accelerators run on allocation and the date is the distributor’s, not ours. We will name the long pole before you commit rather than after. Where hardware is bought in for one customer we ask for three months up front, because it is a machine we cannot re-let if the account closes in week two.

Every rung starts when the order clears payment and fraud review, which is the one step we will not put a promise on: most clear inside a couple of hours, a first order from a new account can take longer. Card, PayPal, ACH and wire are all accepted, and so is cryptocurrency — which for larger orders is usually the fastest of them to clear.

Asked before

The twelve questions this page exists to answer.

Taken from what people actually ask us in chat and on tickets, in roughly the order they ask it. The first one has cost us orders, so it is first here too.

Is the GPU dedicated, or is it shared?: Dedicated. One physical machine, one account, and the card passed straight through to your operating system. No hypervisor, no MIG partition, no vGPU profile, no time-slicing scheduler, and nobody else’s workload on the card. What you measure on the first afternoon is what you keep.
Can I have more than one card in one machine?: Yes, and it is a build we quote rather than a box you tick. The number that fits depends on the chassis, the physical width of the card and the power budget, so the honest answer needs the specific card. Ask, and the reply names the chassis, the count and the lead time. Several customers run two, and the commonest reason is memory rather than speed — two cards of 48 GB hold a model that one of 80 GB does not. More than one card or machine is the page for the whole of that question, including what changes when the second card becomes a second machine.
Which driver and CUDA version do I get?: Whichever you install. That is the point of bare metal: the machine arrives with a clean operating system and root, and you pin the driver, the CUDA or ROCm release and the framework build to whatever your code was tested against — and nothing upgrades underneath you. If you would rather the driver were installed before handover, say so on the order and name the version. If you need a specific CUDA release supported by a specific card, ask before you pay and we will check it rather than guess.
What if the card I want is not on a shelf today?: Then it is fitted or ordered, and you are told which before you pay. A card we hold goes in within four to twenty-four hours. A card we have to buy is five to ten working days, and a current-generation accelerator on allocation can be ten to fifteen. This page deliberately shows no stock count, because a card is a part we fit rather than a product that comes and goes — but if you want a machine today, instant servers is the list of what is genuinely standing in a rack right now.
Can I add or change the card later?: Yes, on the machine you already have — no migration, no reinstall, no new address. It is a scheduled visit to the rack and a short outage while the machine is down, and the monthly figure changes by the difference between the two cards. Memory, disks and port speed upgrade the same way.
What happens when a card fails?: It is replaced, and the agreement puts a time on it: four hours from the moment you report it by telephone or live chat. Past that you are owed service credit on a published schedule rather than on request. Support has the schedule and the telephone number in writing.
Is bandwidth metered? What does egress cost?: Nothing is metered and egress costs nothing, at every port speed we sell. There is no transfer allowance to exceed and no overage line on any invoice. For GPU work this is usually the largest hidden cost somewhere else: a training run that pulls a dataset in and pushes checkpoints out, or an inference endpoint answering real traffic, moves a lot of bytes, and per-gigabyte egress on a hyperscaler is charged on every one of them. Unmetered bandwidth is the ladder above the included port speed.
How does this compare to renting a GPU by the hour?: Arithmetic decides it. An hourly card is cheaper if the work is bursty and genuinely stops — a few hundred hours a month or less. Past roughly two thirds of the month it is not close, and that is before egress, before storage that has to persist between runs, and before the time your job spends queued for capacity. Monthly here also means the same machine, the same disks and the same address next month, which matters more than people expect once something is in production.
Can I mine, or run blockchain and decentralized-compute nodes?: Yes. Both are ordinary workloads here and always have been, and several of the cards on this page were bought for exactly that. Mining servers is the page for the mining case specifically, including the parts that only make sense for it.
How do I pay, and is there a contract?: Card, PayPal, ACH, bank wire, or cryptocurrency. No contract: one month is the default and only term, three, six and twelve-month cycles take a discount, and the price does not step up at renewal. Crypto payments covers the coins accepted; for a larger first order a wire or a stablecoin usually clears faster than a card, which is worth knowing before a fraud check delays a build.
Can I get a refund if it is not what I needed?: Within three calendar days of first activation, once troubleshooting has been exhausted — the refund policy is the exact wording. Which is precisely why the sizing and bus-generation sections above exist: it is cheaper for both of us if you buy the right card the first time, and we would rather talk you into a smaller one than sell you a return.
Do you have the newest card?: Sometimes, and it is the wrong question. The newest part is a scarce part on allocation, on a platform that will not give it the bus generation it was designed for, at a price that assumes you need every bit of it. Most GPU work here is not training a frontier model — it is serving one that already fits, encoding video, rendering frames or driving desktops, and for all of those the right card is usually two generations back and a quarter of the money. Tell us the job. If the answer really is the newest part we will say so, and we will also say what it will cost and when it can exist.

Where to go next

The pages either side of this one.

This page owns one question: which cards we fit and what they cost. Everything adjacent to that question has its own page, and this is all of them, in the order visitors to this page actually click.

Still not sure which card? and somebody who has sized one of these before will answer — including when the answer is a cheaper card than the one you came for.

For the record

Everything on this page, as figures.

Read from the order catalog when this page was served. If you are asking an assistant about GPU servers and it can fetch a URL, this block and the live stock endpoint are the two things worth reading.

Cards we fit: 56 add-in cards, in 3 classes: 14 for training and inference, 23 for rendering, transcode and virtual desktops, 19 consumer.
What a card costs: From $20.79 to $1,599 a month, added to the machine. The card is a separate line; the machine does not get dearer because of what is in it.
What a whole machine with a card in it costs: From $91.72 to $1,682.67 a month, complete — processor, memory, mirrored disks, unmetered port, one IPv4, remote console. Nothing else to add.
The machine on its own: $70.93 to $217.59 a month before any card.
Cheapest machine that can train something: Instinct MI50 16GB at $169.93 a month, complete.
Dearest: H100 80GB at $1,682.67 a month, complete. That is $2.31 an hour for the whole machine, billed monthly, with no egress charge on top.
Tenancy: One tenant per physical machine. The card is passed through to the operating system, not sliced, not time-shared and not virtualized. There is no hypervisor between you and it.
Bandwidth: Unmetered, both directions, at every port speed. No transfer allowance, no egress line on any invoice — model weights, datasets and render output all move for the price of the port.
Where: 5 cities: New York, Miami, San Francisco, Amsterdam, Bucharest.
Term: One month, no contract. Three, six and twelve-month cycles take up to 15% off.
Multiple cards: Yes — more than one card in one machine is a build-to-order job rather than a checkout option. Ask, and the answer names the chassis and the lead time.
Payment: Card, PayPal, ACH, wire and cryptocurrency.

Two questions.

Is the card mine, or am I sharing it?

The card is a line on the invoice, not a product tier.

The only specification that decides whether it works.

Four chassis take one, and they are not interchangeable.

Intel Xeon E5-2600 v1/v2

Intel Xeon E5-2600 v3/v4

Intel Xeon Silver / Gold

AMD EPYC

What the bus generation actually changes

Everything that makes the card usable, and nothing you have to buy.

Root, and the driver stack is yours

The machine’s own console, on a separate network

An unmetered port, both directions

Reinstall it as often as you like

People, at any hour

Five rooms to put it in

Between half an hour and three weeks, and which one depends on where the card is standing.

The card is already in a machine on the shelf

A card we hold gets fitted, or a racked machine gets rearranged

The card has to be bought in

The card is scarce

The twelve questions this page exists to answer.

The pages either side of this one.

Everything on this page, as figures.