Our Take
The crash from tokenmaxxing to budget cuts reveals that most enterprises still have no idea whether their AI spend actually works.
Why it matters
Silicon Valley's AI hype cycle hit a hard constraint this year: actual costs. As companies move past proof-of-concept, the question shifts from 'can we use this?' to 'should we pay for this?'—and most don't have a clear answer yet.
Do this week
Finance lead: audit your AI spend by department and model (GPT vs. Claude vs. in-house) before end of Q1, so you can kill licenses that aren't tied to a measurable outcome.
Tokenmaxxing hit the fiscal wall
Silicon Valley spent the first half of 2024 in what could charitably be called an AI free-for-all. CEOs pushed teams to maximize token consumption, treating API access as a boundless resource. Uber reportedly exhausted its annual AI budget in a few months. Other enterprises cut Claude licenses in certain departments. Meta scrapped an internal leaderboard that was meant to track AI adoption.
This pattern—explosive early adoption followed by abrupt pullback—tells a simpler story than most coverage suggests. It wasn't enthusiasm that changed. It was the bill.
ROI measurement is still missing
Tiffany Luck, a partner at NEA who has spent years embedded in enterprise AI adoption, framed the real problem in a TechCrunch Equity podcast appearance: companies are "still figuring out their AI ROI." That's not a technical problem. It's an accountability problem.
The gap between early hype and actual return-on-investment is where the venture capital industry now lives. Startups are stepping in to help enterprises measure AI spend against tangible outcomes. That market signal—the need for ROI tracking tooling—is more important than any individual product release.
What's missing from this year's AI story is not capability. Models work. What's missing is a credible way to tie AI spend to business metrics. Without that link, AI budgets behave like discretionary spend in a downturn: first to get cut.
Map AI spend to a measurable outcome now
If your team deployed Claude or GPT in production this year, you need a number: how much time did it save, how much error did it eliminate, what revenue did it drive. Not a guess. A number you can defend to CFO review.
Companies that can't articulate this difference will face another round of license cuts in Q1 2025. Those that can will be ready to scale.