#22 Token Economics: Why It’s Hard for Large Companies to Track Their AI Spending
You can also listen to the episode with video on Spotify.
Enterprises around the world are struggling to predict and track their AI spending. What makes this a hard problem to solve, and what can CFOs do now to get a handle on token spending? This episode is ideal for anyone at an enterprise today who is trying to manage these costs.
Speaker: Ravi Kuppan
LinkedIn: https://www.linkedin.com/in/ravikuppan/
Yarken is a Betatron Venture Group portfolio company.
Episode learnings
AI spend has flipped from predictable per-seat licensing to variable, spiky consumption — which is why it's landed on the CFO's desk. Ravi's rough figure: about 75% of enterprises are over their AI budget, and he cites Uber burning a full-year token budget in three months and Salesforce spending ~$300M on one vendor.
Most enterprises are paying several times for overlapping capability: Copilot bundled into M365, GitHub via Microsoft, plus standalone Claude or ChatGPT seats. His analogy — buying a Ferrari when a Nissan does the job — unless you genuinely have power users.
The harder problem is invisible spend: AI features embedded inside SaaS products you already buy. Getting to true AI total cost of ownership means joining contracts, the GL, and actual usage data in one place. A Fortune 50 company he spoke to still can't state its AI cost.
Self-hosting isn't an escape. One university built on-prem AI infrastructure and found the energy cost of a single rack doubled its entire data centre bill.
Practical advice for 2027 planning: name an owner now; deploy an AI gateway (a proxy that reveals model, user, business unit, and query) so you can attribute and charge back; and block anything from going to pilot or production without a governance metric attached.
Tokens aren't comparable across providers — no like-for-like benchmark exists, which makes forecasting genuinely hard.
US enterprises remain wary of Chinese models; where they're used, it's self-hosted. Data sovereignty, MCP data flows, and security harnesses are becoming procurement questions.
His contrarian call on the data centre buildout: per-token prices fall, but total spend won't — consumption expands to fill it. The durable question becomes revenue generated or cost removed, not cost alone.
Hosted by Arshad Chowdhury, Managing Partner at Betatron: Betatron.co X (Previously Twitter): @arshadgc
LinkedIn: https://www.linkedin.com/in/arshadgc
