The cloud wasn't built for HFT. Why companies like Dropbox, X, and high-frequency traders are repatriating workloads to bare metal.
AWS and GCP offer incredible elasticity, but they charge a premium for it. For sustained, high-throughput workloads (like a validator or a sequencer), the unit economics of the cloud break down. Specifically, Egress Fees and vCPU markup act as a tax on scale.
Adjust the sliders below to see where the crossover point lies for your infrastructure. Pay attention to the Break-Even Month—this is typically faster than most CFOs expect.
Adjust your infrastructure scale to see the cost comparison
Equivalent: m5.2xlarge (8 vCPU, 32GB RAM)
EBS gp3 vs Enterprise NVMe
Internet data transfer out
Colo saves $121,715 over 3 years
One hidden cost of colocation is ops. In AWS, you click a button to reboot. In a colo, you verify a ticket with a human technician ("Remote Hands") to physically reboot a server. However, modern Bare Metal Cloud providers (like Equinix Metal) bridge this gap by offering API-driven provisioning on dedicated hardware.
Beyond price, the cloud imposes a "Virtualization Tax". Hypervisors (Nitro, KVM) negotiate resources between tenants ("Noisy Neighbors"). This introduces jitter.
In High Frequency Trading, average latency is irrelevant. Tail latency (p99) is what kills alpha.
✓ Standard: Acceptable for most web apps, APIs, and streaming.
When you rent an EC2 instance, you are renting a slice of a CPU. When other tenants on that physical host burst their usage, your L3 cache get evicted. This causes micro-stalls in your signing execution.
| Feature | Public Cloud (AWS/GCP) | Bare Metal / Colo | Impact on HFT | | :------------- | :--------------------- | :-------------------- | :---------------------------- | | Tenancy | Multi-tenant (Noisy) | Single-tenant (Quiet) | Critical (Cache Eviction) | | Network | Virtualized (Overlay) | Direct NIC Access | High (Packet Jitter) | | Clock | NTP (Variable Drift) | PTP / GPS | High (Ordering Accuracy) | | Cost Model | OPEX (Pay-per-hour) | CAPEX + OPEX (Lease) | Medium (Unit Economics) |
Our benchmarks show a 14-20µs overhead introduced by the AWS Nitro hypervisor compared to bare metal execution of the same code. In a race where margins are measured in nanoseconds, this is unacceptable.
When should you stay on AWS, and when should you move to a cage in NY4 or TY3?
Until your AWS bill hits >$10k/mo or latency loses you >10% of trades, the flexibility of the cloud outweighs the TCO savings.
If you are running a sequencer, solver, or heavy validator, bare metal is mandatory for unit economics and competitive latency.
If you decide to move to colocation, you will need a strategy for Hybrid Deployment. You don't move everything at once.
Check out our Deployment Guide for specific hardware recommendations for ZeroCopy Sentinel on bare metal.