AI spending just became a board-level problem. We break down why agent-driven token consumption is exploding, and the overlooked lever, knowledge tokens, that actually brings the bill down.
By
Bigdata.com Team
·
AI spending just became a board-level problem. Companies that went all-in on agents last year are now staring at bills nobody budgeted for and scrambling to figure out where the money went.
The overspend is real
Uber reportedly burned through its entire 2026 AI coding budget by April.
Microsoft pulled Claude Code licenses from developers months after rolling them out.
One company reportedly racked up a $500M Claude bill after forgetting to set usage limits.
Per-developer token consumption is up roughly 18.6x in nine months (Jellyfish) but the heaviest users are only ~2x as productive, not 18x.
(Source: TechCrunch, “The token bill comes due”)
The culprit isn’t AI in the abstract, it’s agents. They loop: retrieving, reasoning, regenerating, often unattended. Prices per token have fallen but consumption has exploded anyway.
“The whole conversation shifted from tokenmaxxing and ‘go fast’ to: we need guardrails, how do we control this?” J.R. Storment, FinOps Foundation
The industry’s response so far
A standards body (the Tokenomics Foundation) just launched to bring FinOps-style discipline to token spend. A wave of vendors are racing to add cost dashboards. Goldman Sachs projects global token usage will grow 24x by 2030.
But most fixes so far amount to the same move: route to a cheaper model. Useful, but it treats the symptom, not the cause.
The overlooked lever: knowledge tokens
Everyone talks about input and output tokens. The bill hides in a third category: knowledge tokens, the source content stuffed into the context window so the model can answer with facts instead of guesses.
A typical setup pours an entire document into the model to answer one question and here go ~20,000 tokens, almost all of it noise. Multiply that across an agent looping dozens of times, and that’s where the budget goes.
The fix isn’t a cheaper model. It’s giving the model less to read.
What we just launched
On July 20, we launched the Tokenization of Content on Bigdata.com. AI agents retrieve, license, and pay for premium content in the same unit AI runs on: the token.
Precision retrieval returns only the ~200 tokens that actually carry the answer, instead of the 20,000 surrounding it.
Same question. Same model. Up to 100x less paid context per query and sharper, better-cited, less hallucination-prone answers, since the model is reading signal instead of noise.
“The answer isn’t a cheaper token; it’s fewer, better grounding tokens.” Armando Gonzalez, CEO of RavenPack and Founder of Bigdata.com
Every retrieved token carries a license and citation back to its source, across 170+ providers. Natively connected to Claude, ChatGPT, and Microsoft Copilot; reachable by any agent via MCP or API.
How the billing actually works
Quick hits from Dan Benitez, our SVP of Product’s breakdown:
Price ≠ word count. Foundational content ~$8-10/M tokens; scarce material like expert interviews, ~$86-98/M.
Compute and content are separate line items. Standard tokens = retrieval only. Licensed tokens = retrieval + content license.
Cost tracks complexity, not length. Seven sample research runs averaged $1.56 each, ranging $0.68–$2.69.
A usage dashboard + token calculator mean no bill should ever be a surprise.
More on this from our CEO
Armando Gonzalez: why content should be priced the way AI actually consumes it
The takeaways
The winners will be the ones who stopped paying to feed their AI noise, not the ones chasing the cheapest model.
The tokenization of content is a structural shift in how AI retrieves knowledge: instead of licensing entire datasets, applications retrieve only the passages, filings, transcripts, or data points needed to answer a query.
That precision cuts the context an AI model has to process by up to 100×, lowers inference cost, improves grounded answer quality, reduces hallucination risk, and creates a new usage-based monetization model for publishers.
