01
Full root access
Root SSH from the moment provisioning completes. IPMI/KVM-over-IP attached when the chassis supports it. Reinstall the OS without us touching it.
NOCOPERATIONAL
nodes1,792/2,800
GPUs22,400
SKU types66
regions11
fabric99.99%
flops27.4 ExaFP16
incidents0
live
2026-06-24 12:54:56
Rent Bare Metal GPU Servers: H200, B200, H100 from $1.43/hr
Dedicated GPU clusters. No virtualization. No Docker. Direct hardware access.
Rent dedicated GPU and CPU hardware from 8 to 296 cards per order, fixed-duration from 14 days to 5 years. 99.99% uptime, 10 regions, paid in BTC or USD stablecoin. Built for large-scale AI training, distributed inference and HPC.
01GPU
02Quantity
03Rental term
04Location
05Specs
07Price : 1 year
$ 179,020.80
USD (USDT or USDC)
$ 179,020.80
USD equivalent
H200 : 1 year : $2.59/hr per GPU
08Pay with
>live node previewauto-updating
node-2-usa.bm4U, chassisonline
GPU bay8x H200 : 141 GB HBM3e
HopperH200141 GB
01
HopperH200141 GB
02
HopperH200141 GB
03
HopperH200141 GB
04
HopperH200141 GB
05
HopperH200141 GB
06
HopperH200141 GB
07
HopperH200141 GB
08
CPU + memory2 TB ECC RAM
socketXeon
socketXeon
2x Intel Xeon 8568Y+
Storage12 x NVMe
NVMe
NVMe
NVMe
NVMe
NVMe
NVMe
NVMe
NVMe
NVMe
NVMe
NVMe
NVMe
NetworkInfiniBand 3.2Tbps
RDMA fabricInfiniBand 3.2Tbps
RegionUSA
Power6.0 kW
VRAM total1,128 GB
FP16 total13.36 PFLOPS
SLA99.99% uptime
DC tierTier 3 / Tier 4
Provisioningvia SSH key on delivery
Inventory22,400 GPU fleet · since 2022
10 regions across 4 continents
Fixed-duration rentals
Pay in BTC or USD stablecoin (USDT/USDC)
browse full pricing matrix ↓What you actually get.
Clore Bare Metal is a physical, non-virtualized GPU server. CPU, RAM, disks, network and GPUs are dedicated entirely to your account. You get root SSH from minute one, IPMI or KVM-over-IP when the chassis supports it, and zero isolation between your code and the silicon. Performance matches the vendor spec sheet, not a virtualized fraction of it.
Available in cluster sizes from 8 to 296 GPUs, with terms from 14 days to 5 years, across 10 regions worldwide. Pay upfront in Bitcoin or USD stablecoin. The clock starts when SSH hands over.
When the marketplace is not enough.
The Clore marketplace is great for short on-demand jobs, prototyping, single-host inference. Bare Metal is for committed, multi-GPU workloads: pre-training, distributed fine-tuning, HPC, multi-tenant rendering pipelines. Different products, different tradeoffs.
attribute
bare metal
marketplace
Hardware access
Direct, no virtualization
Inside Docker container on shared host
Cluster size
8 to 296 GPUs per order
Single host (per-server rental)
Billing model
Fixed-duration rental
Per-minute, pay as you go
Minimum term
14 days (UK) or 30 days (most)
No minimum (per-minute)
Currencies
BTC, USDT, USDC
BTC, USDT, USDC, CLORE
SSH access
Root SSH on bare hardware
Container shell, host kernel managed
Power limit
None, full spec-sheet TDP
Host may PL-cap cards
Multi-node IB
Up to 3.2Tbps InfiniBand or RoCE
Not available
Six things only bare metal gives you.
A container rental is fine for prototyping and short jobs. The moment your workload pegs the PCIe bus, needs RDMA across nodes, or runs unmodified for weeks, the kernel-shaped seam between you and the hardware starts to matter. These are the six pillars.
02
No virtualization
No hypervisor, no isolating layers. Your kernel talks directly to the silicon. Single tenant per machine. No noisy neighbours.
03
No power limits
Cards run at full spec-sheet TDP. No host-side PL cap, no thermal throttling under sustained load. Hopper at 700W stays at 700W.
04
Any OS, any kernel
Ubuntu, Rocky, Debian, NixOS, ESXi, Proxmox, bring your own image. Docker and Podman on request. Kubernetes allowed when the master is in the same DC.
05
Identical replacement
Hardware failure triggers an identical-SKU replacement inside the SLA window. Same firmware, same performance, no degradation. Downtime beyond 99.99% is compensated automatically.
06
Standardized SKUs
Every node is a published SKU: GPU model, CPU threads, RAM, NVMe layout, NIC and fabric, region, traffic limit. No opaque tiering, no surprises on the invoice.
Full pricing matrix. All 66 SKUs.
Tap any row to see per-region per-term pricing. Every cell here is a real, orderable SKU. $/hr is per GPU and already includes the platform fee. Jump back to the configurator at the top to pick a SKU and check out.
ixgpuarchvramregionsavailabilityper-GPU / hr
01GB200Blackwell192 GB HBM3eUSA1 SKU$4.62 to $5.94 /hr
02B300Blackwell288 GB HBM3eUSAEUJapan3 SKU$5.02 to $5.16 /hr
03B200Blackwell192 GB HBM3eUSAJapanSlovenia3 SKU$3.93 to $5.59 /hr
04H200 SXMHopper141 GB HBM3eIcelandJapan2 SKU$2.81 to $4.18 /hr
05H200Hopper141 GB HBM3eUSAIndiaJapanSlovenia6 SKU$2.09 to $3.86 /hr
06GH200Hopper96 GB HBM3USAEU2 SKU$2.95 to $3.80 /hr
07H100Hopper80 GB HBM3USAThailandJapanSloveniaChina7 SKU$1.81 to $2.62 /hr
08H800Hopper80 GB HBM3China1 SKU$1.68 to $2.45 /hr
09MI325XCDNA 3256 GB HBM3eUSA1 SKU$2.35 to $3.10 /hr
10MI300XCDNA 3192 GB HBM3USA1 SKU$1.96 to $2.60 /hr
11A100Ampere80 GB HBM2eUSAJapanSlovenia3 SKU$1.54 to $2.28 /hr
12L40SAda48 GB GDDR6 ECCUSAJapanSlovenia3 SKU$1.04 to $1.40 /hr
13RTX PRO 6000 BlackwellBlackwell96 GB GDDR7 ECCUSAJapan2 SKU$1.69 to $2.10 /hr
14RTX 6000 AdaAda48 GB GDDR6 ECCUSAFranceJapan3 SKU$1.13 to $1.44 /hr
15RTX A6000Ampere48 GB GDDR6 ECCUSAFrance2 SKU$0.53 to $0.75 /hr
16A40Ampere48 GB GDDR6 ECCUSAFranceJapan3 SKU$0.53 to $0.75 /hr
17V100Volta32 GB HBM2USAFranceJapan3 SKU$0.36 to $0.48 /hr
18L4Ada24 GB GDDR6USAFranceJapan3 SKU$0.41 to $0.52 /hr
19A10Ampere24 GB GDDR6USAFranceJapan3 SKU$0.43 to $0.64 /hr
20T4Turing16 GB GDDR6USAJapan2 SKU$0.26 to $0.40 /hr
21RTX 5090Blackwell32 GB GDDR7USAJapanUKSlovenia4 SKU$0.64 to $1.06 /hr
22RTX 4090Ada24 GB GDDR6XJapanHong KongUK4 SKU$0.40 to $0.94 /hr
23RTX 3090Ampere24 GB GDDR6XJapanHong Kong2 SKU$0.34 to $0.50 /hr
24RTX 3080Ampere10 GB GDDR6XJapanHong Kong2 SKU$0.23 to $0.36 /hr
For workloads that do not fit a container.
01
LLM pre-training and fine-tuning
Multi-node H100, H200 or B200 clusters over 3.2 Tbps InfiniBand or RoCE v2. Run TRL, DeepSpeed, Megatron-LM, vLLM, SGLang on dedicated silicon. NCCL is topology-tuned before handover.
PyTorchDeepSpeedMegatron-LMvLLMNCCL 2.22TRL
02
Distributed inference at scale
Sustained throughput without per-tenant contention. Serve LLMs, diffusion or recommendation models 24/7 with predictable latency. Reserved per-GPU pricing for the whole term.
TensorRT-LLMSGLangTritonRay ServeOllama
03
HPC and scientific compute
CFD, molecular dynamics, weather, particle physics, quant finance. OpenMPI over IB, raw RDMA, ECC memory, deterministic placement. What your sysadmin actually asked for.
OpenMPISlurmCUDA-QGROMACSCUDA 12.6
04
3D rendering and VFX
Stack 8 cards per chassis, attach a 60 TB NVMe scratch, render an 8K shot overnight. Reserved terms from 14 days keep render farms hot during a release.
Blender 4.3Octane XRedshiftHoudiniMaya
Where the iron lives.
Tier 3 and Tier 4 data centers across four continents. Choose the region for latency, compliance or data sovereignty. The exact region keys shown here are the same strings the API uses.
region Iceland
Iceland
1 GPU model
H200 SXM
1SKUTier 3+
region USA
USA
19 GPU models
B200H200H100A100L40SRTX 5090
21SKUsTier 3+
region France
France
6 GPU models
L4A40RTX 6000 AdaV100A10RTX A6000
6SKUsTier 3+
region India
India
1 GPU model
H200
1SKUTier 3+
region Thailand
Thailand
1 GPU model
H100
1SKUTier 3+
region EU
EU
2 GPU models
B300GH200
2SKUsTier 3+
region Japan
Japan
18 GPU models
H100L4A40RTX 6000 AdaV100A10
18SKUsTier 3+
region Hong Kong
Hong Kong
3 GPU models
RTX 4090RTX 3080RTX 3090
3SKUsTier 3+
region UK
UK
2 GPU models
RTX 5090RTX 4090
3SKUsTier 3+
region Slovenia
Slovenia
6 GPU models
B200H200H100A100L40SRTX 5090
8SKUsTier 3+
region China
China
2 GPU models
H100H800
2SKUsTier 3+
Wire up to 296 GPUs over real RDMA.
Distributed training and HPC need a real fabric, not a TCP overlay. Most Hopper, Blackwell and Ampere SKUs ship with 3.2Tbps InfiniBand or RoCE v2. NCCL is tuned for the topology before handover.
RDMA, GPUDirect, same row.
Zero-copy GPU to GPU across hosts. All-reduce bandwidth on H100 SXM and H200 clusters lands within a few percent of the Nvidia reference design.
- 3.2 Tbps fabricInfiniBand or RoCE v2 on most Hopper / Blackwell SKUs
- NCCL tunedtopology-aware collectives configured before handover
- RDMA + GPUDirectverbs API, zero-copy host bypass
- Same VLAN, same rowco-located physically, no cross-AZ hops
node-018x H100
node-028x H100
node-038x H100
node-048x H100
IB fabric3.2 Tbps
Every site signs the same paper.
Before a provider's first rack goes live, they ship Clore a documentation package: Tier certificate (or recognized local equivalent), SLA, fire safety report, redundancy diagrams. Then we run acceptance tests: public IPv4, IPMI, iPerf3, disk perf. Only then the SKU appears in the catalog.
Power
- Tier 3: N+1 UPS + diesel
- Tier 4: 2N UPS + diesel
- Documented redundancy scheme on file
Cooling
- N+1 CRAC
- Hot/cold aisle containment
- Per-rack temperature telemetry
Network
- Min 1 Gbps symmetric, 10 Gbps recommended
- Public IPv4 with rDNS, IPv6 available
- BGP, dual-homed, anti-DDoS at edge
Security
- Dedicated management VLAN for IPMI
- Basic ACLs by default
- TLS-only management plane
Operations
- 24/7 NOC, response within 15 min
- Identical hardware on failure
- Monthly uptime + incident reports
Compliance
- Tier 3+ certificate on file (or recognized local equivalent)
- Fire safety compliance
- Documented RPO / RTO and emergency procedures
End of term
- Logged disk sterilization
- Admin access closure
- Audit trail for compliance reviews
From order to SSH, typically within 48 hours.
Most orders provision in under an hour. Larger clusters or busy regions can queue: the contractual SLA is up to 48 hours from confirmed payment to root SSH. The same five steps apply for an 8-GPU node or a 296-GPU multi-region cluster, no phone call, no sales engineer, no PO.
01
Configure
configurator above
Pick a GPU, quantity, spec package, term and region in the configurator at the top of this page. The price preview updates in real time. Drop in your public SSH key.
02
Pay upfront
BTC or USD stablecoin
The full term is paid in one shot in Bitcoin (BTC) or USD stablecoin (USDT or USDC on Ethereum or BNB Chain). No subscription, no auto-renew.
03
Provisioning
minutes to 48 hours
After payment confirms, the provider runs acceptance tests: public IPv4, IPMI access, iPerf3 throughput, disk performance. Most orders complete in under an hour. Larger clusters or busy regions can queue up to 48 hours. The order page tracks every check.
04
SSH handover
root, your key
Your public key is dropped into root on the node. The order page surfaces the IP, IPMI credentials when supported, and a connection string. You are in.
05
Operate
24/7 NOC behind you
Monthly uptime reports, identical-SKU replacement on failure, response within 15 min on incidents. End of term: logged disk sterilization, admin closure, full audit.
Frequently asked questions.
[01]How is Bare Metal different from the Clore marketplace?
ansBare Metal is your physical machine: CPU, RAM, disks, network and GPU. No virtualization, no Docker, root SSH. The marketplace rents single hosts inside isolated containers. Bare Metal is sold in clusters of 8 to 296 GPUs, fixed-duration rental.
[02]Which GPUs are available?
ansB300, B200, H200 (SXM and PCIe), H100, A100, L40S, RTX 5090, RTX 4090 and RTX 3080. See the full pricing matrix above for SKU-level pricing per region per term.
[03]What durations can I rent for?
ans14 days (UK only), 1 month, 3 months, 6 months, 1 year, 2 years, 3 years, 4 years, 5 years. Not every SKU offers every tier: B300 starts at 2 years, India H200 and Thailand H100 start at 1 year, China H100 starts at 3 months.
[04]What payment currencies are accepted?
ansTwo on /bare-metal: Bitcoin (BTC) on Bitcoin L1, and USD stablecoin (USDT or USDC) on Ethereum and BNB Chain. No other networks for stablecoin on bare metal orders. The CLORE token is accepted on the marketplace but not on bare metal orders.
[05]What is the SLA?
ans99.99% uptime per node, Tier 3 or Tier 4 data centers, 24/7 NOC, on-site smart hands for hardware replacement.
[06]Can I cluster nodes over InfiniBand?
ansYes. Most Hopper and Blackwell SKUs include 3.2 Tbps InfiniBand or RoCE v2 per node. NCCL is topology-tuned before handover. Multi-node clusters are co-located in the same row where possible.
[07]When do I get SSH access?
ansProvide a public SSH key (ssh-ed25519 or ssh-rsa) when placing the order. After payment confirms and the cluster is provisioned, the order status page surfaces IP and SSH connection info.
[08]What happens if a server goes down?
ansThe provider is contractually obligated to deliver an identical replacement with no degradation (same SKU, same firmware) within the SLA window. Downtime beyond the 99.99% threshold is compensated automatically.
clore bm order --gpu H100 --qty 8 --region USA --term 1y
The machine is yours.
The clock starts on provisioning.
Order a non-virtualized cluster, pay in BTC or USD stablecoin, get root SSH on hardware that nobody else is sharing. 8 to 296 GPUs, 14 days to 5 years, 10 regions.
For data center operators
Own racks of Hopper, Blackwell or Ada hardware?
List your fleet on Clore Bare Metal. Get matched with reserved-term renters paying upfront. Income paid in USD stablecoin or BTC, your choice. Clore Partners is the high-volume placement track for legal entities operating Tier 3+ facilities.
Income
USD or BTC, your choice
Min term
2 weeks, standardized SKU
Requirements
Tier 3+ DC, 24/7 NOC, IPMI
Settlement
100% upfront for the full term