Two years ago, processing a million tokens through a frontier AI model cost $33. Today, the same work costs 9 cents. That 99.7% price collapse is reshaping how companies buy, budget, and think about AI - and it's pushing tokens toward becoming something closer to a corporate currency than a billing metric.
Tokens are the small chunks of data that AI models process. Every word you type into ChatGPT, every response it generates, gets broken into tokens first (roughly one token per word, though longer words split into pieces). Every AI interaction in every company is now measured in these units.
From Software Seats to Token Budgets
The shift is already underway. Companies are moving from traditional per-user software licensing to token-based consumption models, and the numbers are staggering. OpenAI reported that average reasoning token consumption per organization jumped approximately 320 times over the past 12 months. That's not a typo - 320x in a single year.
Nvidia CEO Jensen Huang pushed the idea further at a recent event, suggesting that employee token allocations could become standard practice across the industry. His framing was blunt: "I could totally imagine in the future every single engineer in our company will need an annual token budget," with estimates that those budgets could reach half of an engineer's base salary.
That's a corporate expense line that didn't exist three years ago, potentially rivaling compensation costs for technical staff.
The Measurement Problem Nobody's Solved
Here's the tension: tokens measure volume, not value. A badly written prompt that burns through thousands of tokens might produce garbage. A tight, well-structured query might deliver exactly what you need for a fraction of the cost. The metric captures how much AI you consumed, not whether it accomplished anything.
Consider a concrete example from the PYMNTS analysis: an AI agent that costs $4 in inference tokens to save a customer service rep 15 minutes of work. That's negative ROI. The token spend is real and measurable. The value created is questionable.
Deloitte's 2025 survey backs up the gap between spending and results. Nearly half of business leaders expect it will take up to three years to see ROI from basic AI automation. Only 28% of global finance leaders report measurable value from their AI investments so far. Meanwhile, cloud computing bills jumped 19% in 2025 as generative AI became operationally central.
If token consumption gets tied to performance evaluations - which seems inevitable once budgets are assigned - workers will optimize for AI interaction frequency rather than task quality. We've seen this movie before with click-through rates and logged hours.
Where the Money Actually Goes
The cost picture is more complex than per-token pricing suggests. According to Deloitte's analysis, roughly 50% of "AI factory" costs come from networking, power, cooling, facilities, and the software stack - not the GPUs everyone fixates on. Three-year simulations showed on-premise AI infrastructure delivers over 50% savings versus API-based solutions once token production hits certain scale thresholds.
That creates a bifurcated market. Smaller companies and individuals pay per-token through APIs like OpenAI and Anthropic, absorbing variable and sometimes unpredictable costs. Large enterprises increasingly bring token production in-house, treating AI compute like a utility.
For the people who actually use AI tools daily, the practical question is straightforward: your employer is about to start tracking how many tokens you consume, and that number will eventually show up in budget conversations alongside headcount. The companies building AI tools are racing to make tokens cheaper. The companies buying those tools are trying to figure out whether cheaper tokens actually translate to better work. Nobody has a convincing answer yet, and the 320x consumption growth suggests most organizations are still in the "spend first, measure later" phase.