Yesterday, my free year of Perplexity Pro expired.
It was fine. Decent research tool for generic “thought leadership” drawing from online sources. Bad for academic research in fields where many of the relevant sources aren’t readily available online. It’s a lot better with computer science than philosophy, for example. Output was frequently irrelevant or inaccurate for literary stuff.
But here’s the thing. Even if Perplexity Pro’s performance was acceptable across the board, I wouldn’t pay $20 a month for it – let alone $200 for Max. There are plenty of tried-and-true ways to dig out articles and books, as well as free LLM-based tools that could point me in the right (or wrong) direction. Besides, I’ve got cats to feed.
The reality is that application-layer companies like Perplexity are likely taking losses on many of their users, whether that’s offering people like me premium subscriptions for free or not passing on the full cost of tokens to heavy users of flat-rate plans. Subscriber numbers are padded, operating costs are high, and these companies are losing money.
It’s a similar story for the companies making the models these applications rely on. Until recently, OpenAI and Anthropic’s flat-rate plans mostly obscured the true inference costs associated with tasks. When Anthropic introduced metered billing for enterprise customers earlier this year, we saw companies place tighter controls on AI spend.
Used correctly, these tools can do useful stuff. But often that usefulness doesn’t justify their actual cost (let alone their environmental impact). AI companies can’t keep subsidising their users’ tokens forever. And there’s limited evidence that either enterprises or consumers will be willing to pick up the bill.
There’s no such thing as a free token. If the customer isn’t paying the full cost of inference, the AI company – and ultimately its investors – is covering the shortfall. What we’re left with is questionable unit economics and companies that rely on unprecedented levels of funding to keep juicing their growth.
What could possibly go wrong?
