Are Tokens the New Headcount?
You might have run into someone recently talking about how their company has adopted AI. A CEO proud that their team is running agents or scaling their operations or asking the AI a huge number of questions every day. The leaders who have been doing this for a bit though are starting to talk a bit differently. They’re thinking about the actual cost and reward.
Recently on the All-In-Podcast, Jason Calacanis disclosed that his agents were running at $300 per day, annualized at over $100k, and were only at a fraction of their capacity. Chamath Palihapitiya added that he’d been forced to institute “token budgets” for his top developers to avoid running out of money. In response Mark Cuban, on X, called it the “smartest counter” to the current job replacement narrative. He argued that the combined cost of the agents and developers to replace a single human can often be not only more expensive, but less valuable as the AI agents lack the real-world judgement of the human employee.
For companies jumping into expanded AI usage, the token costs can often be invisible… right up until they aren’t.
A former CMO who helped take Atlassian public wrote this week that token consumption is starting to become a proxy for how seriously a team is using AI, but nobody really has enough of history to confidently forecast spending on tokens, and where it will go. The reality is that within many companies, employees are starting to burn through tokens, but they’re not yet being held accountable for them.
B2B marketing budgets specifically have been stable for years: roughly 45% headcount, 45% programs, and 10% technology. Technical allocations are starting to rise 2 to 3 points year over year, while headcount remains flat. CFO’s are starting to put budget pressure to have AI not just improve the quality of the output, but to drive down the overall cost. The problem is that most current marketing budget models don’t have a specific box yet for token costs, and it’s a potential bomb lying inside (or outside) the budget.
Some companies are approaching the token costs by bundling all their AI costs into a single enterprise deal, and then spread those costs across the entire organization. At best this is a temporary fix, treating AI usage like office space in a budget line essentially. A cost of doing business. There are several problems with this though, and the model won’t hold once departments start hitting their token limits, while others start to resent the fact that part of their budget is subsidizing the work of another team.
In March, the Wall Street Journal reported that some companies are scrambling to build internal dashboards to better track token consumption after they had several month-end billing surprises. Whether employees or companies want it or not, per-person token tracking is coming.
Another layered problem here is that not all tokens are equal. There is a 375x price differential between the cheapest and most expensive output tokens across the major models. For an enterprise running millions of queries every month, model selection is now a strategic financial decision, not just a technical preference. Because each model does different things better than the rest, advanced users generally have to use a number of different models, all with these different economics.
Going deeper there are discussions about how to approach AI, when considering agencies and their use of AI for their clients. The recent AI Coca-Cola campaign required over 70k prompts and millions of tokens to generate. Some agencies are replacing headcount with AI, reducing their margin costs, and then pass on the token costs to their clients as a separate line item. That seems to be a structural ethics problem just waiting to explode. Some agencies are treating tokens like a production cost line item living next to catering, others absorb them at a company level, others are passing them through.
At some point the smarter question that clients might ask of their agency isn’t “Do you utilize AI?” but “What is your token budget for this project, and how much of what I’m paying you is human vs machine?”
Another thing to consider is the 10% Problem. I once saw a business owner reject a proposal to adopt project management software which would cost him $30k for his company, because the time it would save per person wouldn’t free up this team enough to take on any more clients.
Sometimes automation doesn’t free up any headcount, and it just becomes a new cost.
Even if AI is working as promised, and you’re partially automating someone’s role, it doesn’t automatically generate savings. For instance if you were to use AI to automate 10% of someone’s role, but that person has 5 clients, each taking up 20% of their time, you can’t cut any headcount. You have 8 people doing what 8 people were already doing, but now you have a growing token bill.
Understanding the ROI of your Tokens requires you not just ask whether the AI saved any time, but did that time savings change anything that shows up in our P&L.
At current token prices, replacing a human with AI agents can potentially cost more than the person they are replacing, not less, once you factor in not only the agent and token costs, but the maintenance costs of those agents as well. The economic math only works at scale, and for specific task types, and even then only if token costs stay where they are or get lower. According to IDC, projects of Global 1000 companies could potentially overspend on their AI infrastructure by at least 30% in 2027, if they don’t manage this problem effectively.
And speaking of token costs, don’t count on cheap tokens bailing you out. Many companies, at the encouragement of the big AI vendors, have an assumption that most token costs will go down fast enough to make the economics work, but I think that deserves a little bit of scrutiny. In 2025 Open AI burned $8 billion against $13 billion in revenue, and projects $14 billion in losses for 2026. These companies are not profitable at their current pricing. Uber used to be cheap too, until it started caring about making money. Current token prices are reflecting a race to gain market share, not sustainable unit economics. When you read about the massive infrastructure and energy investments required to actually drive down costs and all the hurdles those will need to jump, you can’t assume that lower token costs in the future are guaranteed.
If there is even a small AI bubble burst, the market corrects, and token costs jump to levels to drive profit, Jason Calacanis $100k intern could turn into a $500k intern overnight.
There’s another issue here. Your highest spenders might actually be your best people. Not all token spend is waste, not all token spend is inefficient. The instinct from folks like Chamath to cap token usage might actually act, as Shashi Bellamkonda put it, like a tax on your best performers dressed up like financial discipline. Guillermo Rauch of Vercel found that his highest token spenders WERE his top performers. A team of his AI agents analysed a research paper, and built a critical infrastructure service in a single day at the cost of about $10k. He estimated it saved him millions of dollars.
The goal therefore shouldn’t necessarily be less token spend, but more visibility into which spend is generating value and which isn’t.
The companies that seem to be getting this right, are the ones building the measurement layer first, tracking how their tokens are being spent, and then letting any governance like spend limits or token caps, follow what their data shows them. Deloitte in January 2026 called for treating token economics with the same rigor as energy or capital allocation including real-time monitoring, budget alerts, and chargebacks to business units. This emerging framework is a centralized AI gateway that tags every API call by every team and project, with soft limits that alert, and hard limits that throttle only when needed.
At the end of the day, the token budget problem is a measurement problem, and measurement is something that marketers and data people already know how to do. The tools might be new, but discipline isn’t. The same instinct that drives most measurement fits here. What tokens drove an outcome? Which were burned with nothing to show? Those aren’t new questions. The companies who treat token spend as a measurement challenge vs a cost cutting exercise will be the ones who move to the front.
All sources for endnotes
The New Enterprise Currency (All-In, Cuban, Calacanis, 375x differential, IDC)
Your 2026 Marketing Budget Has a Hole in It (Carilu Dietrich, Jensen Huang)
Digiday: Agencies and Token Economics (Coca-Cola, pass-through models, Anomaly)
Your Highest Token Spenders Might Be Your Best People (Rauch/Vercel, WSJ, Zapier)
Deloitte: AI Tokens and Spend Dynamics (FinOps, governance framework)
Morph: Real Cost of AI Coding in 2026 (Cursor overages, Claude Code usage data)
WAV Group: Token Costs Are Invisible Until They Aren't (agent cost math)
Want to discuss this topic?
Schedule a Call