// Case studies
How real Solana workloads run on Subglow
Three anonymized operator case studies — a Pump.fun sniper, a KOL copy-trade shop, and a multi-DEX arb desk. Real architectures, real numbers, honest trade-offs. Names and specific identifying details are obscured at each operator's request.
Solo operator running a Pump.fun sniper on Sniper plan
A solo trader building a bot that buys newly-minted Pump.fun tokens immediately after a bonding-curve launch event. Competing directly with dozens of other snipers on the same tokens.
// Architecture
One gRPC stream against grpc.subglow.io in AMS with a Pump.fun-only filter (vote=false, failed=false). Bot runs in the same AMS datacenter on a rented VPS. Second stream monitors Jito bundle status. Signing happens on-box (custodial, non-hot-wallet) and bundles submit to Jito AMS.
// What changed
- ✓Pre-parsed JSON eliminated a Rust Borsh-decode path the operator had been maintaining — removed ~300 lines of code and ~10ms of per-event latency.
- ✓Moved from a credit-metered competitor that was charging $400–900/mo depending on market activity. Flat $99 ended the billing anxiety during Pump.fun meme cycles.
- ✓Competing directly with other snipers, p99 latency matters more than p50. The AMS colocation and pre-parsing together are the edge.
If your workload is a single filter on Pump.fun and your bot runs in Europe, a Subglow Sniper plan + AMS VPS is a hard combination to beat for the money.
Four-person team running a copy-trade bot tracking 120 KOL wallets
Small team operating a hosted copy-trade product for retail users. Needs to detect trades from a rotating watchlist of 120 Solana KOL wallets across Pump.fun, Raydium, and Jupiter, and fan out execution to ~800 concurrent user sessions.
// Architecture
Three gRPC streams split by program (Pump.fun / Raydium / Jupiter) for parallelism, plus two streams for specific high-value KOLs with account_required filters, plus one for Jito status. Rust matcher in AMS fans out to a WebSocket layer that the browser-signers subscribe to. Non-custodial — each user's key sits in their browser.
// What changed
- ✓Pro's full-speed delivery + 10 concurrent streams lets the team keep program-level streams separate instead of multiplexing everything through one filter. Simpler code, lower p99.
- ✓Custom Program IDs on Pro let them add Phoenix and Meteora coverage without renegotiating a contract — they just updated the filter config.
- ✓At their scale (~$2,800/mo in aggregate 0.2% copy-trade fees), the $249 infra line is a rounding error. Flat pricing simplifies the P&L.
Multi-KOL copy-trade shops benefit most from Pro's stream headroom and custom Program IDs. Split by program, don't try to cram everything through one filter.
Two-person arb desk running statistical arbitrage across 4 Solana DEXes
Quant-style arb desk that watches pool-state accounts across Raydium, Orca, Meteora, and Phoenix, and fires atomic transactions through Jito when pricing dislocates by more than a configured basis-point threshold.
// Architecture
Dedicated private gRPC endpoint colocated in FRA alongside the arb bot. Each monitored pool has its own account filter subscription for minimum match evaluation cost. Bundle submission goes through Subglow's FRA Jito relay — same rack as the validator and the bot, minimum possible RTT.
// What changed
- ✓Moving from a credit-metered competitor to Dedicated cut the median event-to-decision time roughly in half — most of that gain came from skipping the Borsh decode step because Subglow pre-parses Raydium and Orca swap instructions.
- ✓Dedicated infrastructure meant no shared-tenant backpressure during market spikes. Previously the desk was losing trades specifically during the windows where they had the biggest edge.
- ✓Custom pipeline work on Subglow's side (a bespoke pool-state snapshot delta stream) replaced a weekly DevOps effort the desk had been absorbing internally.
At the arb / HFT end of the spectrum, Dedicated stops being an infrastructure line and becomes a competitive moat. If your edge is measured in milliseconds, a shared tier will leak it.
Your workload probably looks like one of these
If you recognize your bot in any of the case studies above, the fastest way to check the math is a free trial against your real workload — no credit card, standard filters included.