Bare-metal Solana hosting for validators, RPC, and MEV pipelines

Most teams running Solana workloads do not need bare-metal. A vds.performance will run a sniper, a market-maker, an indexer with Postgres, and a webhook fanout without breaking a sweat. Pay for a full physical machine when, and only when, the workload actually wants it.

The first “yes” is running a Solana mainnet validator. Not testnet. Not a cluster simulator. The real one where your skipped slots show up on the leaderboard. A validator's vote pipeline is sensitive to scheduler jitter in a way that no virtualization tier completely fixes; the hypervisor overhead on a single-vCPU exit is small in microseconds and large in skipped slots over a year. The economic incentive is clear too: validator commission on a 50k SOL stake at current prices pays a $2,265/month bare-metal box back inside two months. Anything cheaper than dedicated metal is a false saving here.

The second is RAM-bound workloads above 128 GB. Our vds.max tops out at 128 GB RAM with 32 cores; above that, the virtualization tax on memory bandwidth starts to bite. If you're running an RPC mirror on the full validator client, a Postgres indexer with 200 GB shared_buffers, or a Geyser plugin chain doing in-memory account aggregation, you want bare-metal where the memory controller talks directly to the workload.

The third is dual-NVMe workloads where the ledger and the accounts DB must not share IOPS. Solana's validator does tens of thousands of random writes per second to the accounts DB during heavy chain usage; sharing that drive with the ledger (which is mostly sequential) creates contention that shows up as catchup-mode lag. The fix is two physical drives, which is a bare-metal feature. nln.metal.base through nln.metal.ultra ship with four NVMe drives each: two smaller drives for OS and swap, two larger drives for the ledger and accounts DB.

If none of those describe you, save the money. Run on VDS and revisit the question when the workload outgrows it.

The Solana docs publish a minimum hardware spec: 12 cores AVX2, 256 GB RAM, two 1 TB NVMe drives, 1 Gbps network. That spec is the floor for staying in sync. It is not the spec you want for a production mainnet validator that doesn't skip slots during heavy chain usage. The field standard, what the top-100 validators actually run, is closer to this.

The two drives matter more than people expect. Solana writes the ledger sequentially (cheap) and the accounts DB randomly (expensive). On a single drive both contend for the same NAND channels and the validator falls behind during heavy chain activity. Two physical drives let each pattern run at full speed. We rack our nln.metal.pro boxes with the drives in separate PCIe slots, on different lanes, so they don't even share controller bandwidth.

The 10 Gbps network is overkill on a normal day and exactly right on a TPU-flooded one. Stake-weighted QoS on the ingress side has helped, but a 1 Gbps NIC during a memecoin spike still drops UDP packets, which means missed shreds and repair traffic. 10 Gbps gives you headroom that 1 Gbps measurably does not.

The bare-metal cost is the input most validator-economics calculators wave their hands at. So here's the math, with numbers you can plug in.

A nln.metal.pro at $2,265/month is $27,180/year. Vote transaction costs run roughly 1.0 to 1.2 SOL per day at current network conditions, call it 400 SOL per year. That's roughly $58,000 in vote costs at $145 SOL. Total operating cost, $85,180 a year, before any profit margin you keep on commission.

Revenue is operator commission on inflation rewards plus a share of MEV from your block builder. Inflation pays roughly 4.6% per year on stake. At 7% commission and 50,000 SOL of delegated stake, that's 161 SOL of commission, $23,300 at current prices. Add another 100-200 SOL of MEV revenue depending on your relationship with Jito and similar block builders. Net at 50k SOL stake: somewhere between -$30,000 and $0 depending on MEV. You need 75k-100k SOL of stake to be comfortably profitable.

That math changes meaningfully with hardware cost. A $1,500/month box at Latitude.sh adds $7,000 in operating expense, pushing the breakeven stake up. A $720/month OVH box in Gravelines saves $2,000 a year but the network distance to Frankfurt validators costs you 12-18ms on inbound shreds and gossip, which translates to skipped slots, missed credits, and delegators leaving.

The summary: bare-metal hosting is 7-15% of total operating cost. The other 85% is vote fees, and those are the same regardless of where you run.

A few sophisticated validator teams own their hardware and rent rack space directly from Equinix or one of its peers. That can work, especially if you're already running multiple validators across chains and have the operations muscle. Below three or four boxes, the math usually doesn't.

The hidden costs of self-colo: rack space at FR4 runs roughly $1,200/month for a quarter rack (10U usable), plus $300/month/kW power, plus $500/month for a 10 Gbps cross-connect pair, plus the cost of getting hands into the building when something fails. Equinix doesn't do remote hands cheaply. A DIMM swap at 3am is a $400 ticket.

The hidden upsides: you own the box, so the depreciation schedule is yours. You can run whatever firmware you want (relevant for Firedancer optimizations). You can colo with your own switches, giving you full network policy control.

The pragmatic split: rent bare-metal from us until you have three or more validators running profitably and a person whose job includes datacenter operations. At that point, owning the hardware and renting space starts to pencil. Below that, the math says rent.

The hosts most validator operators evaluate alongside us:

Hetzner is dramatically cheaper but the spec doesn't match. Ryzen 9 isn't a validator-grade CPU (no 12-channel DDR5, no ECC on consumer parts), 128 GB RAM is below field standard, and Falkenstein is 14ms from Frankfurt validator density. OVH is in-budget with a fair spec but Gravelines is 11-15ms from FR4. Latitude.sh is the closest direct comparison, similar rooms, no private RPC VLAN. We include the VLAN, IPMI, and ~24 hour provisioning.

Talk to sales for committed annual pricing (typically 10-15% off list) or for custom configurations. Standard SKUs are visible on the pricing page.

Every tier in the bare-metal lineup. All prices are monthly. All servers run in Frankfurt only, private-networked to NLN validator nodes.

nln.metal.edge is the 6-core entry point. It runs a single-purpose workload: an RPC mirror, a Geyser publisher, or a monitoring dashboard that needs physical NVMe without the shared-tenancy of VDS. The 3.8 GHz clock gives single-thread performance for latency-sensitive pipelines.

nln.metal.fast bumps to 12 cores at 4.4 GHz. The high clock makes it suited for single-threaded Solana client code (the vote pipeline is single-threaded in Agave) where clock speed matters more than core count.

nln.metal.base and nln.metal.boost are the same core count (16) and RAM (128 GB) at different clock speeds. Base at 3.0 GHz for steady-state workloads, boost at 4.5 GHz for compute-heavy pipelines. Both ship with four NVMe drives: two 480 GB for OS/swap and two 1.9 TB for data. The dual storage tier eliminates IOPS contention between the ledger and accounts DB.

nln.metal.pro is the validator-ready tier. 24 cores at 2.9 GHz with 384 GB RAM and four NVMe drives (2x 480 GB + 2x 3.8 TB). This is the box that matches the field standard for Solana mainnet validators.

nln.metal.ultra is the Firedancer-class box. 32 cores at 3.25 GHz, 768 GB RAM, 100 Gbps networking, and 2x 480 GB + 2x 8 TB NVMe. Built for Firedancer's multi-tile architecture where memory bandwidth and network capacity are the bottlenecks.