GPU Rent HubMulti-GPU clusters

8x H100 cluster rental on one physical machine.

GPU Rent Hub rents a dedicated 8× H100 SXM node, one single-tenant HGX server with all eight cards on an NVSwitch fabric, for $12,170 a month or $3,420 a week, paid in Bitcoin, Ethereum, USDT, USDC, Monero, Solana or Litecoin with no identity verification.

Nine node configurations, from an 8× RTX 4090 box at $2,810 a month to an 8× B200 HGX baseboard at $28,010.

Updated 2026-09-01

640 GBHBM3 on an 8× H100 SXM node

900 GB/sNVSwitch, every card to every card

$2.11Per GPU-hour, 8× H100 monthly

1Tenant per physical node

01 Definition

What is a multi-GPU node, and how is it different from a cluster?

The word cluster gets used for two different products. Only one of them puts the cards on the same fabric.

The definition we work to

A multi-GPU node is one physical server rented to one tenant, with every GPU on a single baseboard and a direct fabric between the cards. A GPU Rent Hub 8× H100 SXM node is an NVIDIA HGX H100 baseboard: eight cards, 640 GB of HBM3, NVSwitch giving each card 900 GB/s to every other card, 160 vCPU, 2 TB of system RAM, 16 TB of local NVMe and two 100G ports, with root on the bare metal. It is not eight instances joined over Ethernet, and it is not a slice of a machine shared with someone else.

Why the distinction matters

A virtual cluster gives you eight cards that can each run a job. A real node gives you eight cards that can run one job together, and the difference shows up the moment you shard a model, because sharding means the cards talk to each other at every layer.

Everything on the node is yours: CPUs, RAM, NVMe, both uplinks and the fabric. Nothing of ours runs inside the operating system, and a node in stock deploys like any other instance.

02 Which node

How much does an 8x H100 cluster rental cost per month?

$12,170 a month for the eight-card HGX H100 node, $2.11 per GPU-hour. The rest of the range, mapped to the work it is bought for.

Workload	Node	Total VRAM	Weekly	Monthly	Per GPU-hour
Tensor-parallel serving of 70B models, and training that no single card holdsStart here	8× H100 SXMHGX H100 · NVSwitch	640 GB	$3,420	$12,170/mo	$2.11
Full fine-tune with Adam, or a 405B model in FP8The fastest fabric of the nodes that clear both	8× B200HGX B200 · NVSwitch 1.8 TB/s	1,440 GB	$7,850	$28,010/mo	$4.86
Long-context inference where the KV cache dominates141 GB per card, no sharding for 120B	8× H200 SXMHGX H200 · NVSwitch	1,128 GB	$4,610	$16,490/mo	$2.86
Most HBM per dollar, on a ROCm-native stack$7.45 per GB-month	8× MI300XInfinity Fabric · 896 GB/s	1,536 GB	$3,240	$11,450/mo	$1.99
Multi-week training runs on a budgetNVLink generation behind, half the price	8× A100 SXM 80GHGX A100 · NVSwitch 600 GB/s	640 GB	$2,020	$7,130/mo	$1.24
Four-way tensor parallelism, half the nodeSame fabric, smaller commitment	4× H100 SXMNVLink · 900 GB/s	320 GB	$1,750	$6,220/mo	$2.16
Eight independent inference replicasOne model per card, FP8 support	8× L40SPCIe Gen4 peer-to-peer	384 GB	$1,070	$3,740/mo	$0.65
Diffusion and video queues in parallel1.79 TB/s per card	8× RTX 5090PCIe Gen5 peer-to-peer	256 GB	$1,120	$3,890/mo	$0.68
Render farm, one batch job per cardCheapest way to hold eight cards	8× RTX 4090PCIe Gen4 peer-to-peer	192 GB	$790	$2,810/mo	$0.49

Per GPU-hour is the monthly price divided by the GPU count and by the 720 hours in a term, at 100% utilisation. Prices are identical in every region that stocks the node. Single-card prices and specs on the GPU catalogue.

03 The machine

What comes with the node besides the GPUs?

Every figure here is per node, not per card, and all of it is dedicated to you.

Node	Total VRAM	vCPU	RAM	NVMe	Network	Regions
8× B200HGX B200	1,440 GB	224	2 TB	24 TB	2 × 100G	DFW-1
8× H200 SXMHGX H200	1,128 GB	192	2 TB	16 TB	2 × 100G	DFW-1, IAD-1
8× H100 SXMHGX H100	640 GB	160	2 TB	16 TB	2 × 100G	DFW-1, IAD-1, PDX-1
4× H100 SXMHalf baseboard	320 GB	80	1 TB	8 TB	100G	DFW-1, IAD-1, PDX-1
8× MI300XAMD Instinct platform	1,536 GB	192	2 TB	16 TB	2 × 100G	DFW-1
8× A100 SXM 80GHGX A100	640 GB	128	1 TB	8 TB	100G	DFW-1, IAD-1, PDX-1
8× L40SPCIe Gen4	384 GB	128	752 GB	4 TB	100G	DFW-1, IAD-1, PDX-1
8× RTX 5090PCIe Gen5	256 GB	96	496 GB	4 TB	100G	DFW-1, IAD-1, PDX-1
8× RTX 4090PCIe Gen4	192 GB	64	376 GB	2 TB	100G	DFW-1, IAD-1, PDX-1

Each node includes 20 TB of outbound a month, then $4/TB. Local NVMe extends to 30 TB at $9/TB/mo, and a private VLAN is included on every 8-way node. Site detail on the datacenter page.

04 Interconnect

What is the difference between an NVSwitch node and a PCIe node?

Two groups of machines in the table above, and the gap between them is about thirty to one.

NVSwitch and NVLink

On the HGX platforms, every card reaches every other card at 900 GB/s on the H100 and H200 nodes, 1.8 TB/s on the B200 node and 600 GB/s on the A100 node. The MI300X platform does the same job with Infinity Fabric at 896 GB/s. No CPU sits in the path.

That bandwidth is what makes tensor parallelism realistic. Splitting a single model across eight cards means an all-reduce at every layer, hundreds of times per forward pass, and the cards spend their time waiting on that exchange rather than on arithmetic if the link is slow.

PCIe peer-to-peer

The L40S, RTX 5090 and RTX 4090 nodes have no NVLink. Cards talk over PCIe: about 32 GB/s in one direction on a Gen4 x16 link, about 63 GB/s on Gen5, shared with everything else crossing the root complex. Against the 900 GB/s of an H100 SXM card that is roughly a twenty-eighth of the bandwidth.

In practice that means one model per card. Data-parallel training, eight inference replicas behind a load balancer, eight render jobs: all fine. An 8-way tensor-parallel shard on a PCIe node is possible and slow, and the wall-clock time costs more than the node saved.

Aggregate memory bandwidth is a different number, often confused with the fabric. Eight H100 SXM cards sum to 26.8 TB/s of HBM3, eight B200 to 64 TB/s, eight A100 to 16.3 TB/s. That is how fast each card reads its own memory, not someone else's.

05 Sizing

How much VRAM does a multi-GPU node need for your model?

Count bytes per parameter first. Almost every node decision falls out of that one number.

The sizing rule

Weights take 1 byte per parameter in FP8 and 2 bytes in BF16. To serve a model, add the KV cache and about 2 GB of runtime overhead. To fully fine-tune it with Adam in BF16, budget about 16 bytes per parameter: 2 for weights, 2 for gradients and 12 for optimiser state and the FP32 master copy. On the VRAM guide's figures a 70B needs 75 GB to serve in FP8, 145 GB in BF16, 156 GB for LoRA and 1,288 GB to fully fine-tune with activations counted. Only the last of those is a node-sized number.

What that buys you

A 70B model in BF16 needs 145 GB, so it will not fit on one 80 GB card but sits on an 8× H100 SXM node with 495 GB spare for batch and KV cache. A full fine-tune of the same model at 1,288 GB clears only on the 8× B200 or the 8× MI300X. Everything else here needs LoRA or ZeRO offload, as set out on the GPU for fine-tuning page.

LoRA on a 70B is not one of these jobs. At 156 GB the VRAM guide puts it on a single MI300X at $1,590 a month, so a node costs you seven times the money and an evening of device maps for nothing.

Ranked by gigabytes per dollar the order is blunt: the MI300X node is $7.45 per GB-month, the L40S node $9.74, the A100 node $11.14, the H100 node $19.02. The cheap ones are cheap for reasons, ROCm on one and no NVLink on the other.

06 Discounts

How does the multi-GPU discount work, and does it stack?

5% at two cards, 8% at four, 10% at eight, applied to the single-card price. It stacks with the prepaid term discount.

H100 SXM, one host	List price	Discount	You pay	Per GPU-hour
1 card	$1,690	none	$1,690/mo	$2.35
2 cards	$3,380	5%	$3,211/mo	$2.23
4 cardsNVLink, half baseboard	$6,760	8%	$6,220/mo	$2.16
8 cardsFull HGX node, NVSwitch	$13,520	10%	$12,170/mo	$2.11
8 cards, 3 months prepaid10% then 5%	$40,560	14.5%	$34,685total	$2.01
8 cards, 6 months prepaid10% then 10%	$81,120	19%	$65,718total	$1.90

The discount applies to cards in the same host, not to cards spread across the fleet. Prepaid terms are 3 months at 5% and 6 months at 10%, taken off the already discounted node price. Full table on the pricing page.

The floor is $1.90 per GPU-hour for eight NVSwitched H100 SXM cards, against $2.35 for a single card on a rolling month. Payment does not change with the size of the order: fund a balance in BTC, ETH, USDT, USDC, XMR, SOL or LTC and the term is drawn from it, no card and no identity document. See how crypto payment works, what no-KYC GPU hosting means, and why a monthly term beats an hourly meter.

BTCETHUSDTUSDCXMRSOLLTC

07 Beyond one node

Can you rent 16 to 64 GPUs across several nodes?

Yes, with InfiniBand between the nodes, quoted per case rather than sold from the catalogue.

Tell sales the shape

GPU count, model, region, term, and InfiniBand HDR at 200 Gb/s or NDR at 400 Gb/s. Write to sales.

Quote and lead time

16 to 64 GPUs, priced against the node rates above. Lead time 2 to 6 weeks, depending on the fabric.

50% at order

Half the first term up front, in any of the seven coins. That reserves the hardware.

50% at hand-over

The balance once the fabric passes bring-up and you have root. No identity check at any step.

Networking between instances in a region

Instances in the same region can share a private VLAN: layer 2 between your machines with no public hop, $25 a month, included on every 8-way node. That is the tool for a storage box feeding four 8× A100 nodes.

It is not a substitute for InfiniBand. A VLAN rides the same 100G Ethernet as everything else: fine for data loading and checkpoints, wrong for gradient exchange at every step.

Cross-region clusters are not a product. DFW-1 Dallas, IAD-1 Ashburn and PDX-1 Hillsboro are separate failure domains, and the latency between them makes a distributed training job pointless. One region per cluster.

If you would rather own the cards, the same HGX hosts take colocation at $240 per GPU per month, or buy the hardware outright.

08 Honest limits

When is a multi-GPU node the wrong thing to rent?

If your job fits on one card, a node is a waste of money. An 8× RTX 4090 node is $2,810 a month against $390 for a single card, and a 13B model in FP8 is 22 GB all in, weights, cache and overhead. Eight cards pay for themselves only with eight streams of work, or a model that does not fit.

If you need eight cards for two days, use an hourly provider: a week of an 8× H100 node is $3,420 whether you run it for six hours or 168. And if the constraint is memory rather than throughput, a single H200 with 141 GB or an MI300X with 192 GB often removes the need for a node.

Rent a node when

The model does not fit on one card, the run lasts weeks, you need eight replicas behind one address, or the job is sharded and needs 900 GB/s between cards.

Rent single cards when

The jobs are independent, the load is unpredictable, or you want to test on one H100 SXM before committing to eight.

09 Questions

Multi-GPU and cluster rental, asked plainly.

How much does an 8x H100 cluster rental cost per month?

$12,170 a month, or $3,420 for a week. That is one HGX H100 node with eight NVSwitched cards, 640 GB of HBM3, 160 vCPU, 2 TB of RAM, 16 TB of NVMe and two 100G ports. It works out at $2.11 per GPU-hour over a 720-hour month, or $1.90 on a six-month prepaid term.

Is a GPU Rent Hub node a real machine or a virtual cluster?

A real machine. One physical server, one tenant, root on the bare metal, with all eight GPUs on a single HGX baseboard and NVSwitch between them. Nothing is shared with another customer, including the CPUs, the RAM, the NVMe and both uplinks. We do not sell virtual clusters stitched together over a network.

What is the difference between NVSwitch and PCIe for multi-GPU training?

NVSwitch gives each card 900 GB/s to every other card on the H100 and H200 nodes, and 1.8 TB/s on the B200. PCIe peer-to-peer moves about 32 GB/s per direction on Gen4 and 63 GB/s on Gen5. Tensor parallelism across eight cards is realistic on the first group and painful on the second.

How much VRAM does an 8x H100 node have in total?

640 GB of HBM3, being eight cards of 80 GB, addressable across the NVSwitch fabric. For comparison, an 8× H200 node has 1,128 GB, an 8× B200 node 1,440 GB and an 8× MI300X node 1,536 GB. A 70B model in BF16 needs 145 GB of that, leaving room for a large batch and KV cache.

Does the multi-GPU discount stack with prepaid terms?

Yes. The multi-GPU discount is 5% at two cards, 8% at four and 10% at eight in the same host. Prepaying takes a further 5% for three months or 10% for six, off the already discounted price. Eight H100 SXM cards prepaid for six months is $65,718, which is $1.90 per GPU-hour.

Can I rent more than eight GPUs in one cluster?

Yes, 16 to 64 GPUs with InfiniBand HDR or NDR between the nodes, quoted per case through sales rather than sold from the catalogue. Lead time is 2 to 6 weeks, payment is 50% at order and 50% at hand-over, in the same seven coins. Clusters stay inside one region: DFW-1, IAD-1 or PDX-1.

Can I rent an HGX node without KYC and pay in crypto?

Yes. An account is a username and a password, with no email, phone number or identity document at any order size. Fund a balance in Bitcoin, Ethereum, USDT, USDC, Monero, Solana or Litecoin and the term is drawn from it. That applies to a $190 single card and to a $28,010 B200 node alike.

Which node should I rent to fine-tune a 70B model?

For LoRA, none. It needs 156 GB, which one MI300X holds at $1,590 a month, and the VRAM guide gives the same answer. For a full fine-tune with Adam, budget about 16 bytes per parameter, 1,288 GB with activations, which clears only on the 8× B200 or the 8× MI300X node at $11,450.

Deployment detail in the documentation, fleet state on the status page.

Eight cards on one fabric, by Friday.

Nodes in stock deploy like any other instance. For 16 GPUs and up, talk to sales and expect 2 to 6 weeks.

Create an account →