Retail investors are quickly giving up on cryptocurrency trading bots built with generative artificial intelligence, with high inference costs for large language models cited as the main reason. Token costs to keep the bots running are eroding profitability more than the strategies themselves, the analysis said.

Cryptopolitan, a blockchain outlet, reported on April 29 that many retail traders shut down custom AI trading bots within about 2 weeks of starting. The biggest reason is the cost burden arising from continuous API calls and the inference process.

The analysis is based on the experience of a contributor who runs Agent37, a managed hosting company built on OpenClaw. The contributor said that after observing hundreds of deployments of autonomous AI agents, retail investors are running into high maintenance costs in practice.

Methods for building a Solana trading bot in minutes using tools such as Claude are spreading quickly on YouTube and online communities. The contributor argued that things change in actual operation.

The key point is that the cost of designing a strategy and the cost of continuously running it are completely different. The contributor said LLMs have made it close to free to create momentum indicators or trading rules, but keeping an AI bot running around the clock to read and analyse market data and make trading decisions creates ongoing costs.

The contributor called it an "inference tax". The analysis said this structure can be fatal for retail investors. For example, if an AI bot is set to analyse charts and market sentiment every 5 minutes and decide whether to execute a Solana swap, token usage keeps accumulating.

Many users choose top-tier models such as GPT-5.4 or Claude Opus, but these models can become sharply more expensive when run continuously for long periods. In some cases, the contributor said users spent $10 a day on API costs while actual trading profits were around $2. The contributor called it "a situation where the cost of intelligence becomes larger than the value of the trade".

Perceptions spreading in the AI crypto market were also cited as a problem. People believe a top-level general-purpose AI model is needed even to run a simple trading strategy, but that is not the case, the contributor said. "What you need for a strategy of buying when Solana falls 5 percent is not a genius-level AI, but a fast and cheap model combined with strict system prompts," the contributor said.

As an alternative, the contributor pointed to low-cost open-weight models such as Qwen 3.5 Flash. By adjusting system prompts to fit a specific trading strategy, users can employ a model like one specialised for a particular task instead of a general-purpose AI, and in that case inference costs can be lowered to near zero, the contributor argued.

Realistic barriers to entry also exist. Building a low-cost model in a local environment or on a private server requires substantial technical knowledge. The contributor called it a "new logistics bottleneck". Because most retail investors are not DevOps engineers, they cannot handle complex server build and deployment processes, eventually return to expensive API-based models, and then give up running the bot due to cost burdens, the contributor said.

That has led to a view that competitiveness in the AI trading market will depend not simply on better prompt design but on how easy infrastructure can be made. Platforms are needed that automatically distribute tasks to low-cost models and run them in isolated environments when users enter only a strategy. "Infrastructure should not get in the user's way," the contributor said.

The market also sees the case as showing that barriers to entry for retail algorithmic trading are shifting from coding skills in the past to hosting and inference costs. For AI automated trading to become mainstream, high inference costs and complex deployment environments must be addressed first, the analysis said.