The AI Bill Is Too Damn High: Tokenomics 101 — Eugene Cheah, Featherless AI

Eugene Cheah, founder and CEO of Featherless AI, a company that sells inference on open models, gave a workshop at AI Engineer NYC 2026 for organisations whose token spend keeps rising faster than their productivity. Cheah opened with two caveats: their examples come from their own company's customers, and the advice is aimed at organisations spending more than roughly 10K a month, below which Cheah thinks finding strong use cases matters more than trimming the bill. Cheah's argument is that most savings come from knowing which work is paying for which tokens, and from not paying for speed nobody is waiting for.

The first step is measurement. Cheah said easily 30 to 40% of the customers who approach Featherless start with a single API key shared across the whole organisation, which makes it impossible to see which department, team or use case is spending. Separate keys also separate two kinds of workload Cheah says have opposite economics: long-running agents, such as coding agents, which reuse large contexts over many turns, and process automation, such as call-centre decisions, message filtering or document processing, which is short, repetitive and a volume game. With that visibility, teams can right-size. A strong model can plan and review while a cheaper model does the token-heavy implementation, and non-technical departments can be given labels such as fast or smart instead of model names, so the model behind a label can be swapped without asking anyone to change settings. People rarely switch models to save the company money, Cheah observed, but they will switch to a cheaper model if it is also faster. For high-volume work Cheah described testing cheaper open models against your own examples, using a new class of cheap classifier models for decisions, and fine-tuning a small model on data generated by an expensive one, where the provider's terms allow it.

The speed point is the least intuitive. Cheah's rule of thumb is that each doubling of speed per request roughly triples the cost. In one example they gave, a server answering one request at a time produced about 112 tokens per second; running 16 requests in parallel roughly halved each request's speed but produced about 10 times the total output, which Cheah put at around 9 to 10 times lower cost. Cheah therefore argues for paying for high speed only when a person, a developer or a customer, is waiting, not for background agents or CI pipelines, and for buying dedicated capacity sized to the steady base load while letting spikes spill over to public cloud.

Cheah's last step has nothing to do with inference. Tokens, Cheah said, are not work done: generating more code or slides can simply move the burden to whoever has to check them, as when reviewers faced with agent changes too large to review end up discarding them and redoing the work. Cheah calls this kind of hand-off a slop grenade, and argues that people should own whatever they hand over and that organisations should measure outcomes rather than usage, while conceding that outcomes are hard to measure. The figures throughout are a vendor's accounts of its customers, not independent data. Another AI Engineer talk looks at the same bill from the other side, weighing token spend against total engineering cost.