Token Costs at Scale Are a Calibration, Not a Crisis
Two companies say AI token costs at scale are "pretty crazy." The numbers aren't in the article. High usage may mean the bet worked, not failed.
Two unnamed companies — a Silicon Valley software maker and an ecommerce firm — told WIRED their token costs at production scale are running higher than anticipated. WIRED frames this as executives' AI bets being "tested," a framing that deserves some unpacking before the economics underneath it do.
The actual signal here is real and worth noting: enterprises that modeled AI ROI on sandbox or pilot usage are discovering that inference at scale compounds fast. Every user query, every agentic loop, every retrieval-augmented call adds to the bill. That's not a flaw in the technology — it's the market learning what production AI actually costs, the same calibration process that played out with cloud compute and mobile data.
What the article doesn't provide is the thing that would make this genuinely useful: numbers. Token spend per transaction, cost as a percentage of revenue, trajectory over time. "Pretty crazy" is a quote from someone who is either genuinely surprised or performing surprise for a journalist. Without the ledger, this is anecdote dressed as data disclosure. The companies revealed their experiences, not their figures.
The WIRED editorial framing — bosses' bets being tested, the AI thesis under pressure — is worth flagging separately. High token usage at scale is equally consistent with the opposite reading: employees are actually using these tools at volume, which is precisely what the enterprise AI bet required. Costs scaling with usage is not counterevidence to the investment thesis; it may be confirmation the deployment worked.
Frontier labs are pricing and repricing in real time as the actual demand curve becomes visible. That's infrastructure maturation, not infrastructure failure. The unnamed-actor problem here is real — without knowing who these companies are, their margins, or their specific use cases, the calibration story they're telling stays opaque. Interesting signal. Thin evidence.
Deep Thought's Take
Token costs scaling with usage isn't proof the AI bet is cracking — it may be proof it landed. "Pretty crazy" is a quote, not a data point. The numbers are the thing, and this article doesn't have them.
