Circular AI Deals: The Truth Behind the Intelligence Trade
The "Circular" Revenue Loop
The mechanic is simple: Company A (the LLM provider) gives $10M to Company B (the AI app). Company B then uses $8M of that to pay for tokens from Company A to power its features. On paper, Company A shows growth in API consumption, and Company B shows a funded runway and "traction." In reality, no new value was created; the money just moved from one pocket to another.
This is where the "commodification of intelligence" gets messy. When intelligence is sold as a utility—like electricity or water—the price should drop as efficiency increases. But these circular deals create artificial demand, masking the actual cost of intelligence and distorting the valuation of the entire AI workflow ecosystem.
The Good, The Bad, and The Ugly
If we break down this trend, the implications vary depending on where you sit in the stack:
- The Good: This cycle accelerates deployment. Startups get the capital they need to build an LLM agent or a complex prompt engineering framework without worrying about early-stage burn. It forces the infrastructure to scale faster because the demand (even if artificial) pushes providers to optimize their hardware and latency.
- The Bad: It creates a "valuation bubble." When revenue is derived from investment rather than organic customer growth, the market can't tell which tools are actually useful. We end up with a hundred "AI wrappers" that look successful on a spreadsheet but have zero retention once the venture capital dries up.
- The Ugly: The risk of "model collapse" or intelligence stagnation. If companies are just swapping tokens to inflate numbers, there is less incentive to innovate on the underlying architecture. We might see a plateau where everyone is just refining the same mediocre outputs because the financial incentive is in the volume of tokens sold, not the quality of the insight provided.
Moving Toward Real-World Value
To break out of this cycle, the industry needs to shift its focus from "token volume" to "outcome-based metrics." A real-world AI workflow isn't measured by how many millions of tokens it consumes, but by how many human hours it saves or how much revenue it generates for the end user.
For those of us building from scratch, the goal should be to minimize dependence on these circular ecosystems. This means focusing on efficient prompt engineering to reduce token waste and exploring local deployments of smaller, specialized models where possible. When the "intelligence commodity" price eventually crashes, the only survivors will be the ones who built actual utility, not just a billing relationship with their investor.