Somewhere right now, a budget meeting is comparing two lines. One says: junior engineer, fully loaded, roughly a million kroner a year. The other says: enough LLM tokens to generate more code than that junior will write in a decade, for a tenth of the price.
The meeting picks the tokens. The math is clean, the quarter looks great, and everyone moves on.
I think that math is wrong. Not slightly wrong — wrong about what the two lines actually buy. And because I now sit close enough to these conversations to have to argue about them, I want to write down the argument once, properly.
What the clean math gets right #
Let’s be honest about the other side first, because it isn’t stupid.
Tokens are elastic. You scale them up for a migration and down after. They don’t need onboarding, don’t churn after eighteen months, don’t require a laptop, a licence stack, or a performance review. For a well-defined task with a clear spec and a competent person steering, token spend produces working code at a price that makes traditional cost-per-feature look embarrassing.
If your model of an engineer is “a machine that converts requirements into diffs,” the substitution is rational. That’s the model the budget sheet encodes. The problem is that it’s not what engineers are for.
What tokens don’t carry #
Tokens don’t carry pagers. When production breaks at 02:40, something has to take responsibility — not produce a plausible hypothesis, but own the incident, make the judgment call under pressure, and answer for it afterwards. Accountability is not a capability you can rent per million tokens. It lives in people, and it’s most of what you’re paying for.
Tokens don’t accumulate context. A team member who’s been through two incidents, one migration, and a bad quarter knows things about your systems and your org that exist nowhere in writing. That knowledge compounds. Token context resets when the session does. You can point a model at your wiki; you cannot point it at the memory of the last time someone tried exactly this and why it failed.
Tokens don’t push back. The most valuable sentence in engineering is “we shouldn’t build this.” Models are agreeable by construction — they’ll generate whatever you ask for, beautifully. Talent is what tells you the feature is a mistake, the deadline is fiction, or the architecture won’t survive contact with the failure mode nobody put in the prompt.
Tokens multiply. They don’t originate. This is the core of it. An agent amplifies the judgment of whoever is steering it. Strong engineer with tokens: dramatically more output at the same quality of judgment. Weak judgment with tokens: dramatically more output of exactly that. Multiplication by zero is still zero, and now it’s zero at scale, merged before lunch.
The junior problem, again #
The swap has a delayed cost that never shows up in the quarter it’s made: every junior you don’t hire is a senior who doesn’t exist in five years.
Seniors aren’t produced by seniority. They’re produced by juniors doing the unglamorous loop — being wrong, debugging it, being wrong again — with someone experienced nearby. That loop is exactly what token spend replaces first, because junior work is the most automatable work. Cut the bottom rung for four consecutive budget cycles and you haven’t saved money. You’ve scheduled a talent shortage and prepaid for it.
The industry is doing this collectively right now, which means the market will do what markets do: the people who can steer agents and own systems and say “no, don’t build that” are about to get very expensive. Companies buying tokens instead of growing people are bidding up the price of the thing they stopped producing.
The frame I actually use #
It’s not tokens versus talent. That’s the false choice the budget sheet invents. The real question is: what’s your token spend per unit of judgment?
Tokens amplify whatever talent you have. So buy both, deliberately: hire and grow the judgment, then give it as much leverage as it can responsibly steer. On my team that means everyone has serious agent tooling, and nobody’s value is measured by output volume — because output volume is now just a budget number. The scarce thing is people who understand what the output does, take responsibility for it in production, and know when not to generate it at all.
A million kroner buys a junior who might steer a hundred million tokens well in three years. The tokens alone buy diffs. One of these compounds.
I’m four months into management and closer to the money conversations than I’ve ever been, which means I get to watch this argument happen for real — and I don’t yet know if I’ll win it. If you’ve fought this budget fight and have scars or counterexamples, I want them. Find me on LinkedIn.