Why Your Best Employees Should Be Expensive
Token spend flips one of the core heuristics we have used to evaluate software. In SaaS, more usage is a good thing. If a team has Salesforce seats and no one is logging in, or the universal truth of reps not updating their pipeline holds true, that is a problem. When MAUs start dropping, investors on boards start freaking out. In the old world, usage maps rather cleanly to value. AI breaks that relationship. With token-based pricing, less usage actually means higher margins for the vendor. If a Claude user never logs in, Anthropic actually has higher margins. Now, it is obvious that most companies are explicitly telling employees to use AI more. It is in internal memos, it is in conversations over drinks after work, it is all anyone can talk about. That tension is new and not fully resolved.
For operators building the next generation of companies, this forces a rethink on how we measure productivity. The 10x engineer who has sharpened his agentic toolkit may carry two salaries, his W2 and his bill to Anthropic, and that is probably a good thing.
Learn from the Best?
Jensen Huang talked about this on the All-In Podcast. He mentioned how he routinely pays engineers 500k a year, and if he found one who only used 5k worth of tokens all year, he would be frustrated. If someone used 250k, he would be happy. The point is not the spend, it is of course what the spend unlocks. The point was made that this is similar to how LeBron James thinks about investing in his body. He has famously talked about spending over a million dollars a year on his body.
If you can look past the NVIDIA engineers to LeBron James comparison, the analogy makes sense. The best performers spend more on the inputs that make them better. All of a sudden, token usage starts to look less like a cost and more like a direct lever on features shipped, code debugged and direct output. If LeBron has ice baths, cracked engineers have Claude.
The Precedent
This isn’t new, we have seen versions of this before. Companies already spend money to help people do their best work. Free meals, better equipment, home office stipends. At high performing companies, there is a clear willingness to invest in anything that removes friction and helps people focus. The difference with AI is how directly the cost scales with usage. Snowflake was an early preview, and as any CTO would know, your best users are often your most expensive ones. AI takes that dynamic and accelerates it because the feedback loop between spend and output is tighter and more immediate.
So the question becomes how far this goes. Strict controls risk suppressing the very behaviour companies are trying to encourage. No controls at all can lead to real cost blowouts. The middle ground is shifting from cost minimization to return on token spend. Not how much did we spend, but what did we get for it.
Our take on how much we should enable top-performers to spend on tokens? I’d say it probably aligns closer with Mr. Huang and Mr. James (although I’m not sure how Claude Code helps our jump shot, unfortunately).