Setting the Budget

Earlier this year, Jensen Huang said a $500,000 engineer who spends less than $250,000 a year on tokens would make him "deeply alarmed."

That number doesn't hold up. Nvidia sells the infrastructure that processes tokens (the units AI usage is billed in), so more spend means more demand for its chips means more revenue for Nvidia. This is a sales pitch dressed up as advice. It also doesn't survive contact with reality: no engineer can spend $250,000 a year on tokens and have it be an effective spend. Past a point, more tokens just means more re-reads and looping agents, not more output.

What's real: companies are pushing spend past old assumptions, then pulling back hard once finance sees the bill. Meta's internal usage hit 73.7 trillion tokens in 30 days, tracked on an internal leaderboard nicknamed "Claudeonomics." Staff called the race to top it "tokenmaxxing." Then the company reversed course, its CTO writing that "token usage alone is not a measure of impact of any kind." Uber burned its entire annual AI budget in four months, then capped spend at $1,500 per engineer per month.

The ceiling is higher than assumed. So is the crash back to earth once someone reads the invoice.

What this article covers: the budget for everyday employee AI use, the tools and subscriptions your people reach for day to day. It leaves out three things that get costed differently: AI built into your product (that's a pricing and margins decision), AI "employees" and always-on agents (a separate topic, later), and the custom automations you build in-house. None of those fit a simple per-person budget.

One more thing these numbers assume: that you buy everything off the shelf. You don't have to. Two levers bring the bill down. Most of the SaaS seats here have a build-it-yourself version, like your own indexed search in place of Glean; that swaps a licence fee for engineering time. And the model usage underneath, which you pay for either way, gets cheaper on open-weight models served by other providers instead of the frontier labs. Neither is free, but together they can land a team well under these numbers.

Three tiers, one shared stack that gets richer as you go up. One thing to flag before the numbers: a flat subscription is one bill, and metered usage (billed by how much you consume, like an electricity meter) is another. It's the metered side that can get away from you.

The base stack, for everyone

Every employee gets the same foundation, whatever their job. Three things.

A top-lab plan. One $100-a-month "5x" plan from a frontier lab: Claude Max 5x, or the equivalent from another top lab. Not the $20 entry tier. The $20 plan runs out mid-afternoon for anyone actually leaning on it, and a person who hits the wall every day stops reaching for the tool. The $100 plan is the one that survives daily use.

A meeting note-taker. Something like Granola, running quietly on every call so notes and action items write themselves. Around $20 a head. It pays for itself the first week.

Agentic file search. A system like Glean that indexes your documents, chat, and tickets, then answers questions across all of it. Around $30 to $50 a seat at enterprise scale. This is the gap between "I have a chatbot" and "I have a chatbot that knows how my company works."

That is the floor. Everyone gets it, no exceptions.

The average employee

For most of your company, the base stack is the whole budget. A hundred dollars for the plan, twenty for the note-taker, thirty for search. This will be roughly $150 a head, per month. These are people who use AI to write, summarise, research, and answer questions, not to run heavy automated workflows. They don't need more, and paying for tools they won't use is just money spent to look ambitious.

The power user

Some teams live in AI all day, and their tooling should reflect it. A sales team on heavy sales tooling (something like Gong for call analysis, Clay for prospecting). A legal team on legal-specific tooling (something like Harvey). These people move up to the $200 "20x" plan (Claude Max 20x) because they hit the 5x limits, and they carry $100 to $200 of role-specific tooling on top of the base stack. This will be roughly $400 a head as a floor.

The engineer

Engineers are power users with a coding stack bolted on. Same $200 "20x" plan (Claude Max 20x, which includes Claude Code), same base stack underneath. On top of that: code review tooling (something like CodeRabbit, around $25 a seat), dictation for talking through a problem instead of typing it out (Wispr Flow, around $15), and headroom for agents that run against the API. This will be roughly $500 a head as a floor. Their extra usage, the metered spend on top of the flat plans, is capped at $1,500 a month, the same ceiling Uber landed on. The cap is there for the month an agent runs away from someone.

Engineering is where token billing bites hardest, because a coding agent can loop for an hour with nobody watching. That is exactly what happened at Uber: Claude Code spread across 5,000 engineers faster than finance had modelled. The average engineer ran $150 to $250 a month, but the heaviest ran $500 to $2,000. The floor is low. The tail is long. That gap is the whole reason the cap exists.

Flat plans, metered caps

Here is the distinction that keeps the bill predictable. A flat subscription is a known number. Claude Max 20x at $200, Claude Code included, cannot surprise you. Metered usage, billed by the token as it's consumed (what providers call API access), is the part that runs away, because an agent has no natural stopping point. It will happily re-read the same files a hundred times.

So put people on flat plans wherever the work fits inside them, and treat the $1,500 as a ceiling on the metered part, not a target to hit. Uber's bill blew up because it was buying tokens, not seats. Flat where you can, capped where you can't.

The formula

Everything above collapses into one line:

Monthly = ($150 × employees) + ($400 × power users) + ($500 × engineers)

Count "employees" as everyone who isn't a power user or an engineer. Treat the total as an estimate, not a quote: a rough guideline to start from, since every company is shaped differently. And remember what the formula can't show you: anything you build on top of these plans, agents and internal tools, runs on metered usage and has no ceiling unless you set one.

Next: which tools

This gives you the number. It doesn't tell you which tools to actually buy, and the tools are where most of the money is won or lost. Part 03 goes slot by slot: which plan, which note-taker, which search layer, which coding stack, and how to tell the tools that earn their seat from the ones that just have good marketing. It starts with dictation: the voice dictation tools worth setting up.