LLM ‘inference tax’ erodes retail Solana bot gains

Agent37 reports a large share of retail traders stop running Solana trading bots within two weeks because the cost of continuous large-language-model queries often exceeds trading profits. The company calls the extra cost an “inference tax.”

The trend followed a wave of online tutorials showing how to prompt models to write trading logic in minutes. Agent37 operates an OpenClaw managed hosting service and monitors hundreds of autonomous AI agent deployments; the company found that creating a strategy is inexpensive while operating it around the clock is costly.

Agent37’s analysis points to unit economics as the core issue. A bot that wakes every five minutes to read charts, parse sentiment and decide on trades generates frequent API calls. Retail traders often use high-end models such as GPT-5.4 or Claude Opus for accuracy, but those models have high per-query fees when used continuously. Agent37 cited examples where model calls cost about $10 a day while the trading profit was roughly $2 a day.

The company describes a common misconception it calls the “frontier model fallacy”: the belief that a top-tier LLM is necessary for simple tactical choices, such as buying Solana after a 5% drop. Agent37 recommends smaller, faster open-weight models combined with strict system prompts for narrow tasks. The firm points to models like Qwen 3.5 Flash as options that can be configured as specialized workers rather than general-purpose oracles, which reduces inference expenses.

Agent37 also identifies a logistics bottleneck for retail traders who try to avoid expensive APIs. Hosting open-source models requires renting optimized cloud infrastructure, configuring and serving the model, maintaining Python environments and continuous execution loops, and monitoring uptime and crashes. Many individual traders lack that DevOps experience and revert to cloud APIs for convenience, then deactivate bots after seeing their accounts decline.

The company proposes platforms that hide infrastructure complexity: tools that let traders visually deploy strategies, route inference to cost-effective models and run logic inside isolated containers. Agent37 says such platforms would remove the need for traders to manage servers or choose between costly APIs and complex self-hosting.

Agent37’s monitoring shows many retail AI trading experiments end within days to weeks because of running costs and deployment challenges. The company’s findings highlight the gap between the ease of designing rules with LLMs and the ongoing expenses of operating those rules in live markets.