Enterprises that spent the first half of 2026 maximizing AI token consumption are now staring at the invoice, and most cannot say what they bought. NEA partner Tiffany Luck, speaking on the Equity podcast, said companies are still figuring out the return on their AI spend, even as the bills from frontier model providers climb past forecast. Uber reportedly exhausted its annual AI budget within a few months, some firms have cut Claude licenses for parts of their organizations, and Meta killed an internal usage leaderboard that had encouraged employees to push consumption higher.
The trend Luck is describing has a name in Silicon Valley: tokenmaxxing. Earlier in 2026, CEOs urged staff to use AI tools as aggressively as possible, treating raw token throughput as a proxy for productivity. The proxy did not hold. Token counts went up, line-item spend went up faster, and finance teams started asking what shipped because of it.
Luck, who began her career convincing companies that e-commerce was inevitable, is now backing AI startups across the consumer and enterprise stack. She told Equity host Rebecca Bellan that one of the most active investment areas she sees is software that helps enterprises track return on AI spend — essentially observability and FinOps for model usage. The pitch writes itself: if buyers cannot tie OpenAI or Anthropic invoices to revenue, growth, or cost savings, the next budget cycle gets ugly.
Key facts
- 01Uber reportedly exhausted its annual AI budget within a few months of the fiscal year.
- 02Some enterprises have cut Claude licenses for portions of their organizations to rein in spend.
- 03Meta shut down an internal AI usage leaderboard tied to the 'tokenmaxxing' push earlier in 2026.
- 04NEA partner Tiffany Luck is backing startups that help enterprises measure return on AI spend.
- 05The reckoning follows a Silicon Valley trend where CEOs pushed employees to maximize token consumption.
The cuts are already visible. Pulling Claude seats from divisions that cannot justify them is a quiet way of admitting the pilot did not pencil out. Meta's decision to retire its internal leaderboard is a more pointed signal — when the company that helped popularize aggressive internal AI adoption removes the scoreboard, the game changes. The question shifts from how much teams use AI to what specific work AI replaced or accelerated.
Luck is more bullish on the consumer side, where she said the technology can produce what she called magic moments — interactions that would not have been possible without a capable model behind them. Personal agents are the category she is watching most closely. The thesis: consumers will tolerate, and pay for, AI that does discrete tasks on their behalf in ways that are obvious and measurable, where enterprises are still negotiating with their own procurement teams.
The IPO calendar is the other pressure point. Luck weighed in on this year's AI public offerings, a class of deals that includes Anthropic and a string of infrastructure plays. Public-market investors are applying the same ROI question to model providers themselves: revenue growth is undeniable, but unit economics on inference, customer concentration, and the durability of enterprise contracts are all still being underwritten in real time.
Forward-deployed engineers — the consultant-style technical staff that frontier labs send into customer accounts to make deployments actually work — are part of the answer and part of the problem. They lift adoption and prove value inside specific accounts, but they also signal that off-the-shelf model access alone is not enough to drive enterprise ROI. That gap is where Luck sees startup opportunity: tooling, evaluation, governance, and measurement layers that sit between the lab and the buyer.
Skeptics will note that the ROI question is not unique to AI. Cloud went through the same cycle a decade ago, when AWS bills surprised CFOs and FinOps emerged as a category. The counter is that AI consumption is harder to attribute because the output is probabilistic and the productivity gains are diffuse. Cutting a Claude seat is easy. Proving that the seat was not generating $50,000 in monthly value is harder, and most enterprises do not yet have the instrumentation to do it.
The reckoning Luck describes reshapes the buyer side of the AI market in 2026. Frontier labs that built revenue on enterprise enthusiasm now need to defend renewals with measurable outcomes, not token counts. Startups that can bolt ROI measurement onto model usage are positioned to ride the second wave of enterprise AI budgets — the one that actually gets approved. And the model providers themselves face a more disciplined customer base heading into the next IPO window, which is not the worst thing for a sector still being asked whether the revenue is real.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
