Bare-metal Solana hosting for validators, RPC, and MEV pipelines
When the workload is a mainnet validator, a Firedancer build, or anything that wants the IRQ table to itself: you stop renting slices and take the machine. Intel Xeon and AMD EPYC, 32 to 512 GB ECC, physical NVMe per role, 10 Gbps, optional NVIDIA H200, racked next to the NLN validators. From $253/mo.
$ lscpu | grep "Model name"Model name: AMD EPYC 9575F 64-Core Processor$ lsblk -d -o NAME,SIZE,TRAN | grep nvmenvme0n1 3.8T nvme # ledgernvme1n1 3.8T nvme # accounts$ ethtool eno1 | grep Speed && ping -c3 grpc.nln.clr3.orgSpeed: 10000Mb/s · rtt avg 0.18 ms# the spec sheet, verified from userspace
Configure a machine
Entry metal: high clocks, your own NIC, no hypervisor anywhere.
- ·Full control of IRQ affinity, CPU governor, and kernel
- ·Physical NVMe drives; ledger and accounts DB never share IOPS
- ·10–100 Gbps unshared host networking
- ·IPMI / KVM-over-IP on a dedicated management VLAN
- ·Provisioning: ~24 hours · OS installed, drives validated, burn-in tested
- ·SLO: 99.99% network & power · hardware MTTR < 4 h (business window)
Not for: Core-hungry parallel decode: six cores is six cores.
A steady multi-core box for always-on services.
- ·Full control of IRQ affinity, CPU governor, and kernel
- ·Physical NVMe drives; ledger and accounts DB never share IOPS
- ·10–100 Gbps unshared host networking
- ·IPMI / KVM-over-IP on a dedicated management VLAN
- ·Provisioning: ~24 hours · OS installed, drives validated, burn-in tested
- ·SLO: 99.99% network & power · hardware MTTR < 4 h (business window)
Not for: Single-thread race paths: clocks are modest here.
The highest clocks in the fleet. Built for single-thread race paths.
- ·Full control of IRQ affinity, CPU governor, and kernel
- ·Physical NVMe drives; ledger and accounts DB never share IOPS
- ·10–100 Gbps unshared host networking
- ·IPMI / KVM-over-IP on a dedicated management VLAN
- ·Provisioning: ~24 hours · OS installed, drives validated, burn-in tested
- ·SLO: 99.99% network & power · hardware MTTR < 4 h (business window)
Not for: Wide 64-core parallel decode: pick an EPYC tier for that.
64 cores across two sockets with 256 GB for wide parallel work.
- ·Full control of IRQ affinity, CPU governor, and kernel
- ·Physical NVMe drives; ledger and accounts DB never share IOPS
- ·10–100 Gbps unshared host networking
- ·IPMI / KVM-over-IP on a dedicated management VLAN
- ·Provisioning: ~24 hours · OS installed, drives validated, burn-in tested
- ·SLO: 99.99% network & power · hardware MTTR < 4 h (business window)
Not for: Latency-bound single threads: 2.35 GHz is the trade-off for core count.
64 modern EPYC cores: the RPC / validator baseline.
- ·Full control of IRQ affinity, CPU governor, and kernel
- ·Physical NVMe drives; ledger and accounts DB never share IOPS
- ·10–100 Gbps unshared host networking
- ·IPMI / KVM-over-IP on a dedicated management VLAN
- ·Provisioning: ~24 hours · OS installed, drives validated, burn-in tested
- ·SLO: 99.99% network & power · hardware MTTR < 4 h (business window)
Not for: Mainnet voting; 128 GB is below the 256 GB field standard.
64 cores with more headroom for heavy RPC and indexer stacks.
- ·Full control of IRQ affinity, CPU governor, and kernel
- ·Physical NVMe drives; ledger and accounts DB never share IOPS
- ·10–100 Gbps unshared host networking
- ·IPMI / KVM-over-IP on a dedicated management VLAN
- ·Provisioning: ~24 hours · OS installed, drives validated, burn-in tested
- ·SLO: 99.99% network & power · hardware MTTR < 4 h (business window)
High-frequency 64-core, 256 GB: the mainnet validator turnkey.
- ·Full control of IRQ affinity, CPU governor, and kernel
- ·Physical NVMe drives; ledger and accounts DB never share IOPS
- ·10–100 Gbps unshared host networking
- ·IPMI / KVM-over-IP on a dedicated management VLAN
- ·Provisioning: ~24 hours · OS installed, drives validated, burn-in tested
- ·SLO: 99.99% network & power · hardware MTTR < 4 h (business window)
NVIDIA H200 on dual 32-core EPYC, 512 GB: GPU compute.
- ·Full control of IRQ affinity, CPU governor, and kernel
- ·Physical NVMe drives; ledger and accounts DB never share IOPS
- ·10–100 Gbps unshared host networking
- ·IPMI / KVM-over-IP on a dedicated management VLAN
- ·Provisioning: ~24 hours · OS installed, drives validated, burn-in tested
- ·SLO: 99.99% network & power · hardware MTTR < 4 h (business window)
The whole machine, or nothing
Bare-metal is the right answer when virtualization stops giving you what you need: full control over IRQ affinity, no hypervisor in the path, dedicated NVMe drives so the ledger and the accounts DB don't fight for IOPS, and RAM that nobody else can touch. Our Frankfurt fleet spans Intel Xeon and AMD EPYC platforms — from a 6-core entry box to a high-frequency 64-core EPYC 9575F and a dual-socket NVIDIA H200 GPU server — every machine with IPMI/KVM out-of-band management and a private VLAN to our RPC and gRPC fleet. All servers are private-networked directly to NLN validator nodes. Pricing starts at $253/month for nln.metal.edge (Intel Xeon E-2286G, 6 cores, 32 GB, 2x 960 GB SSD) and runs to $12,758/month for nln.metal.gpu (NVIDIA H200, dual 32-core EPYC, 512 GB DDR5). The 64-core EPYC tiers — base, pro, and max — are built for mainnet validators, RPC mirrors, and Geyser-plus-Postgres pipelines. We provision in ~24 hours from order to root SSH.
Bare metal is not a luxury tier; it is a different failure model. A mainnet validator skips slots when the vote pipeline hits scheduler jitter, and a hypervisor exit is microseconds on paper, skipped slots on the leaderboard. An RPC node thrashes when the ledger and the accounts DB share IOPS. These are hardware problems, and the fix is sold by the machine.
Your kernel on the silicon. IRQ affinity, CPU governor, AF_XDP, huge pages: every knob is yours.
Two to four physical NVMe per box. Ledger, accounts, data, and archive never queue behind each other.
CPU, RAM, drive surface, and NIC tested on every box. ~24 h from order to root SSH, validated.
The fleet sheet
| SKU | CPU | RAM | NVMe bays | Network | Built for | Price/mo |
|---|---|---|---|---|---|---|
| nln.metal.edge | 6 cores @ 4 GHz | 32 GB ECC | ▮ 960 GB▮ 960 GB | 10 Gbps | Latency-critical single services · Dedicated relays · Small Geyser pipelines | $253 |
| nln.metal.core | 12 cores @ 2.2 GHz | 64 GB ECC | ▮ 960 GB▮ 960 GB | 10 Gbps | Indexers · Webhook fanout · Background workers | $346 |
| nln.metal.turbo | 16 cores @ 4.3 GHz | 192 GB ECC | ▮ 1.92 TB▮ 1.92 TB | 10 Gbps | MEV hot paths · Sniper engines · Anything where one thread decides the trade | $647 |
| nln.metal.dual | 64 cores @ 2.35 GHz | 256 GB ECC | ▮ 3.84 TB▮ 3.84 TB | 10 Gbps | Full-chain Geyser pipelines · Indexers + Postgres · Non-voting RPC mirrors | $926 |
| nln.metal.base | 64 cores @ 3.1 GHz | 128 GB ECC | ▮ 1.92 TB▮ 1.92 TB | 10 Gbps | RPC mirrors · Geyser + Postgres pipelines · Testnet validators | $1,227 |
| nln.metal.pro | 64 cores @ 3.2 GHz | 192 GB ECC | ▮ 1.92 TB▮ 1.92 TB | 10 Gbps | Full RPC nodes · Heavy indexer pipelines · Multi-service stacks | $1,622 |
| nln.metal.max | 64 cores @ 3.3 GHz | 256 GB ECC | ▮ 3.84 TB▮ 3.84 TB | 10 Gbps | Mainnet validators (Agave) · Firedancer · Full RPC nodes | $1,902 |
| nln.metal.gpu | 64 cores @ 3.55 GHz | 512 GB ECC | ▮ 1.92 TB▮ 1.92 TB | 10 Gbps | DePIN / GPU workloads · On-chain ML & inference · Accelerated analytics | $12,758 |
Built for mainnet validators
| Official docs floor | Field standard (2026) | nln.metal.max · $1,902/mo | |
|---|---|---|---|
| CPU | 12 cores, AVX2 | 64 cores | 64 cores @ 3.3 GHz (EPYC 9575F) |
| RAM | 256 GB | 256 GB | 256 GB DDR5-6000 ECC |
| Storage | 2 × 1 TB NVMe | 2 × large NVMe, separated | 2 × 3.84 TB NVMe, ledger and accounts on separate drives |
| Network | 1 Gbps | 10 Gbps unshared | 10 Gbps unshared + private VLAN |
| Survives | staying in sync, mostly | heavy chain traffic | heavy chain traffic, with margin |
Will the box pay for itself? Slide and see.
Assumes 4.64% network inflation and 1.1 SOL/day vote cost, based on current network conditions, not a forecast. Jito tips and MEV revenue excluded (upside).
- Commission revenue
- $2,944
- Vote transaction cost
- −$4,851
- Hosting (nln.metal.max)
- −$1,902
- Net / month
- −$3,810
at these assumptions, on this hardware
Firedancer images, GPUs, and shapes we haven’t printed
nln.metal.max (64 cores @ 3.3 GHz, 256 GB DDR5) is sized for the tile architecture and its memory appetite. We have hosted Firedancer operators since the frankendancer phase, and ship a tuned image (governor, IRQ affinity, AF_XDP) on request. Need more RAM? We quote custom builds.
EPYC 9684X with 1 TB RAM. NVIDIA H100 or L40S for GPU workloads. 100 Gbps fabric, multi-rack. If Supermicro, Dell, or HPE sells it, we rack it in 7–10 days, on the same private VLAN as the rest of your fleet.
IPMI and serial-over-LAN are always on, so most 3am problems never need a human in the aisle. When one does: 24/7 remote hands at $150/incident, hardware MTTR under 4 hours in the business window.
The end of the isolation spectrum
Tell us the workload, we’ll tell you the box
Bursts CPU on signals, idles between slots. Network-bound 95% of the time.
The bot spends its life waiting on the wire. 8 shared vCPU covers the duty cycle with room to spare, and the money stays in the strategy.
Step up ifmpstat shows steal > 1% in peak hours, or your loop decodes every transaction.
Hot loop on the gRPC feed, decode everything, race a single thread to the trigger.
Pinned cores keep the decode loop consistent: a tail event during your neighbor's burst is a missed trade.
Step up ifthe race comes down to one thread: 4.3 GHz Ryzen metal beats 3.0 GHz virtualized.
Constant ingestion from Yellowstone into Postgres on the same box. Never idles.
16 pinned cores and NUMA-local RAM let ingest and Postgres share a box without sharing tails. Unix-socket DB writes, zero network hop.
Step up ifyou index more than a couple of programs or keep > 40 GB hot in Postgres.
Receive events, transform, fan out over HTTP. Spiky concurrency, light CPU.
16 vCPU absorbs concurrency spikes; shared vCPU is irrelevant when the work is mostly I/O.
Step up ifconcurrency is modest, same architecture, $160 less.
Your own read-only RPC node, ledger and accounts on separate volumes.
32 pinned cores, 128 GB, and four NVMe namespaces handle a non-voting mirror, virtualized, so you can resize as it grows.
Step up ifthe mirror is permanent: a full 64-core physical box, physical drives, no hypervisor.
Learning the operator workflow: vote, restart, snapshot, monitor.
16 vCPU / 64 GB carries a testnet validator. The point is the runbook practice, not the hardware.
Step up ifyou want the exact I/O behavior you will see on mainnet.
Voting on mainnet. Skipped slots cost stake. Hypervisor jitter is disqualifying.
64 high-frequency EPYC 9575F cores, 256 GB, and dual 3.84 TB NVMe is the field-standard build. The vote pipeline never crosses a hypervisor.
Step up ifyou prefer a dual-socket 64-core layout with the same 256 GB.
Custom Geyser plugin feeding your own downstream: full-chain firehose in, structured data out.
Full-chain Geyser is a sustained-everything workload: real drives for the sink, real cores for the decode, 128 GB to breathe.
Step up ifyou want more RAM headroom for the working set: 192 GB on the EPYC 9555P.
The next validator client: tile architecture, AF_XDP, large memory appetite.
Built for it: 64 high-frequency cores, 256 GB, and a Firedancer-tuned image (governor, IRQ affinity, AF_XDP) on request. Need more RAM? We quote custom builds.
Racked where the slots happen
The field guide: boot, verify, vote
#!/usr/bin/env bash
# /opt/solana/run-validator.sh on a nln.metal.max host
# Drive layout (set by NLN provisioning):
# /mnt/ledger -> NVMe #1 (3.84 TB) - solana ledger + snapshots
# /mnt/accounts -> NVMe #2 (3.84 TB) - accounts DB (separate IOPS)
exec /usr/local/bin/agave-validator \
--identity /opt/solana/keys/validator-identity.json \
--vote-account /opt/solana/keys/vote-account.json \
--known-validator 7Np41oeYqPefeNQEHSv1UDhYrehxin3NStELsSKCT4K2 \
--known-validator GdnSyH3YtwcxFvQrVVJMm1JhTS4QVX7MFsX56uJLUfiQ \
--known-validator DE1bawNcRJB9rVm3buyMVfr8mBEoyyu73NBovf2oXJsJ \
--known-validator CakcnaRDHka2gXyfbEd2d3xsvkJkqsLw2akB3zsN1D2S \
--only-known-rpc \
--ledger /mnt/ledger \
--accounts /mnt/accounts \
--snapshots /mnt/ledger/snapshots \
--limit-ledger-size 200000000 \
--rpc-port 8899 \
--rpc-bind-address 127.0.0.1 \
--private-rpc \
--no-voting \
--dynamic-port-range 8000-8020 \
--gossip-port 8001 \
--log /var/log/solana-validator.log# nln.metal.max sanity check
lscpu | grep -E '^(Model name|Socket|Core|Thread|MHz)'
# Model name: AMD EPYC 9575F
# Socket(s): 1
# Core(s) per socket: 64
# Thread(s) per core: 1
# RAM
sudo dmidecode -t memory | grep -E 'Speed|Configured Memory|Type:'
# Configured Memory Speed: 6000 MT/s
# Type: DDR5
# Total: 256 GB
# NVMe layout
lsblk -d -o NAME,SIZE,MODEL,TRAN
# nvme0n1 3.8T ... nvme (ledger)
# nvme1n1 3.8T ... nvme (accounts)
# Network: 10 Gbps unshared
ethtool eno1 | grep Speed
# Speed: 10000Mb/s# All bare-metal boxes ship with dedicated IPMI on a separate VLAN.
# Credentials are handed over with the server on first delivery.
# Power cycle from the CLI
ipmitool -I lanplus -H ipmi-fra-bm-042.nln.internal \
-U admin -P "$IPMI_PASS" power cycle
# Serial-over-LAN console for kernel-panic / install screens
ipmitool -I lanplus -H ipmi-fra-bm-042.nln.internal \
-U admin -P "$IPMI_PASS" sol activate
# Web KVM and virtual-media ISO mount available from the dashboard.# Validator economics on a nln.metal.max ($1,902/mo) at typical inflation.
SOL_PRICE_USD = 145.0
ANNUAL_INFLATION = 0.0464 # current Solana inflation
COMMISSION = 0.07 # what you charge delegators
VOTE_COST_SOL_PER_DAY = 1.10 # vote tx fees, current network conditions
HOSTING_USD_PER_MONTH = 1902.0
STAKE_SOL = 75_000 # delegated stake target
# Annual inflation rewards at this stake
gross_sol = STAKE_SOL * ANNUAL_INFLATION
operator_sol = gross_sol * COMMISSION
operator_usd = operator_sol * SOL_PRICE_USD
# Costs
vote_cost_usd = VOTE_COST_SOL_PER_DAY * 365 * SOL_PRICE_USD
hosting_usd = HOSTING_USD_PER_MONTH * 12
net_usd = operator_usd - vote_cost_usd - hosting_usd
print(f"Gross: ${operator_usd:,.0f} Vote: ${vote_cost_usd:,.0f} "
f"Hosting: ${hosting_usd:,.0f} Net: ${net_usd:,.0f}")
# Note: vote fees dominate. On commission alone, breakeven sits
# near 170k SOL staked. Jito tips / MEV share (excluded here)
# are what close the gap for most operators below that.How the fleet is run
- ·VPS pools run fair-share vCPU scheduling with headroom held in reserve
- ·VDS and metal capacity is never shared; pinned capacity is held hot
- ·Spare capacity stays racked and powered, so orders never wait for a host to be carved
- ·AMD EPYC platforms, DDR4 ECC across the fleet
- ·Datacenter-grade storage with redundancy (mdraid1) on virtualized pools
- ·Every metal box burn-in tested before handover: CPU, RAM, drive surface, NIC
- ·VPS: nightly snapshots, restore to a healthy host < 15 min
- ·VDS: live migration off a failing host, target < 8 min
- ·Metal: IPMI / KVM-over-IP always on, hardware MTTR < 4 h in the business window
- ·Committed QoS bandwidth on VDS; 10–100 Gbps dedicated ports on metal
- ·Dedicated private VLAN to the NLN RPC / gRPC / WS fleet included on every VDS and bare-metal server
- ·Fair-share scheduler on shared uplinks; one tenant cannot saturate a host
- ·Full root. No locked kernels, no restricted outbound ports, no inspection
- ·Key-only SSH from first boot; password auth ships disabled
- ·Bring your own ISO: Ubuntu, Debian, Rocky, Arch, NixOS all routine
- ·99.99% monthly rolling, fleet-wide
- ·Credits applied automatically on breach, no ticket required
- ·Only enforcement on your box: outbound SMTP abuse limits
Frequently asked questions
Related products
Step down from bare-metal when you don't need a full physical machine but still want pinned cores.
Lowest tier. Right for testnet bots, build runners, and bursty workloads.
Use our RPC fleet directly, or run your own RPC on bare-metal for full custody.
Custom Geyser plugins with managed publishing. Pairs with bare-metal hardware.
The streaming layer your bare-metal will publish to or consume from.
Take delivery of a machine this week
nln.metal.edge $253/mo through nln.metal.gpu $12,758/mo. Frankfurt, validator-adjacent, IPMI included, ~24 hours from order to root SSH. Burn-in tested before you ever see it.
Custom configurations and committed-annual pricing are quoted within 24 hours by an engineer who has racked the hardware, not a configurator PDF.
- 01Create an accountEmail and a password. No card required to look around.
- 02Pick tier + paste your SSH keyThe exact specs and prices on this page. No checkout surprises.
- 03Root SSH lands in your inboxKey-only auth, your chosen OS image, ready for the NLN data fleet.