Wharton's Ethan Mollick says AI API tokens transitioned from an overlooked expense to an essential development requirement in one year

Amusing: I'm hearing that at several tech companies, a more frequent topic on internal knowledge sharing sessions is about efficient AI usage.

Cannot really remember topics like "techniques for faster CI/CD" or "how to speed up builds" being THIS widely discussed in the past!

Do you want your best engineers to use the most tokens? Your worst? Can you even figure that out? Are there some use cases to prioritize? Will your other engineers be resentful that they can’t burn as many tokens? How do you deal with competing projects and constraints?

In all fairness, speeding up builds and better tooling have been discussed in the past as well - but more of an ongoing basis, and not all at once, suddenly, across lots of companies!

The token pricing surge is really starting to sting it feels...

Most companies only have very crude understanding of token usage right now, so they veer from focusing on adoption (“everyone should use as many tokens as possible”) to cost control (“can we just use local models?”) depending on the moment and manager. This is all very new.

I have heard from quite a few large organizations that blew through their entire token budget in the first couple months of the year. There aren't even good processes for thinking through how token costs will change over time.

Kinda crazy how fast it’s going

openclaw guy spent $1 mil a month in tokens, meta’s top eng spent half a mil

$200/mo does go far, but it’s also a high bar for anyone not already on a high US tech salary

But big tech businesses are throwing money at it. Genuinely near unlimited budgets.

Opportunity is huge, and risk of spending low and losing is much bigger than risk of spending too much to win. Meta & Microsoft both missed mobile phones and lost billions of dollars. So they will throw everything at the current opportunity.

@emollick Cloud compute had the same lawless phase. Nobody had a chargeback model for EC2 until FinOps emerged around 2016. Tokens will get there, but expect the same 3-5 year governance lag before tooling and org policy actually catch up.

@emollick The interesting shift is that tokens are becoming a leverage allocation problem.

The highest-output operators consume disproportionately more compute.

Companies now have to figure out: who gets amplified, by how much, and where the ROI actually comes from.

@emollick Tokens are clearly turning into the new cloud bill: at first everybody wanted adoption until the invoice showed up.

And all of sudden someone had to start deciding which projects deserve the expensive runs and which ones are just burning money with better autocomplete.

@emollick Compute went from rounding error to line item in 18 months for dev teams. Sales tech is next. Anyone modeling pipeline cost without an AI token line is already behind.

tokens went from zero to a CFO problem in 12 months.

Uber burned through their entire 2026 AI budget by April. Microsoft canceled Claude Code licenses over token costs. the spend is real and the ROI frameworks to justify it don't exist yet.

the organizational problem isn't whether AI is useful - everyone can see it is.

the problem is nobody built the governance layer for "who gets how many tokens and why" before the bills started arriving.

that's not a technology problem. that's a management problem that's about to spawn an entire consulting industry.

@emollick the budget line didn't exist because the token economy didn't exist, and the companies that deployed without that line are now discovering the cost the adoption narrative skipped

@emollick Token futures makets will help

In general give compute constraints tokens should flow to the highest value activities

@emollick Tokens now underpin AI workflows, but defining fair distribution models remains crucial for equitable access and innovation sustainability.

@emollick The same pattern is starting in healthcare AI. Imaging credits and ambient-scribe minutes are showing up as line items that did not exist 12 months ago. Few practices have a framework for who should consume them or on which cases.

@emollick yeah and finance teams are still treating tokens like office supplies when devs are burning through monthly budgets in a week

the "request more tokens" approval flow is the new bottleneck nobody saw coming

I lead AI Engineering in a Not a massive US company

We don't have token budgets right now more or less seeing how much can our engineers push it the highest spend engineer is showing outsized outcomes but also the ones with lower spends are showing higher quality outcome

It is becoming a line item how big it becomes we don't know right now but we have an idea

@totlsota @emollick Using "tokens" like this might be AI's final kick in the nuts for crypto.

@emollick Tokens stopped being a usage metric and became a resource allocation problem. The next issue is which workflows justify spending them at all.

@emollick Token usage is starting to resemble cloud spend a decade ago. Everyone wants AI capability fast, but governance and allocation systems are still catching up.