The great AI experiment is hitting a wall. Tech companies that encouraged employees to go all-in on AI experimentation are now scrambling to contain costs as token bills spiral out of control. Uber revealed it burned through its entire 2026 AI budget in the first four months of the year, with COO Stuart Almada acknowledging it was becoming "harder to justify" internal AI expenses. OpenAI CEO Sam Altman added fuel to the fire earlier this month, calling AI costs a "huge issue" for customers.

The Rise and Fall of Tokenmaxxing

The spending frenzy had its own culture. Employees at Meta and Amazon competed on internal leaderboards tracking who could burn through the most tokens—a practice dubbed "tokenmaxxing." Heavy token use became shorthand for productivity, with Nvidia CEO Jensen Huang going so far as to say he'd be "deeply alarmed" if a $500K software engineer wasn't spending $250K annually on tokens. That was March. The math is catching up now. The core problem isn't simple queries—asking ChatGPT what's for dinner costs almost nothing. It's the agentic workflows that are bleeding companies dry. Cognitive scientist and AI researcher Gary Marcus explained: "Those agents internally pose many, many queries in the process of getting you to your answer... sometimes it takes 500 times as much, a thousand times as many [tokens]." For complex coding tasks and chain-of-thought reasoning, those token counts compound fast.

The Corporate Clampdown

Uber's response? A $1,500 monthly cap per employee per coding tool. That's not innovation-friendly policy—that's triage. Canadian startup leaders at a recent conference echoed the pain (via Betakit), describing growing internal AI expenses that are eating into runway. Nestor Maslej, CEO of an AI consulting firm and former editor-in-chief of Stanford's AI Index Report, put it bluntly: companies have "moved away from very kind of naive experimentation" toward confronting integration realities.

Pricing Models in Flux

The AI providers themselves are scrambling to adapt. Anthropic's Enterprise plan layers a flat fee with token-based charges. GitHub Co-Pilot shifted to token-linked pricing at the start of June. OpenAI is reportedly considering lowering token costs to poach users from Anthropic. Chinese startup DeepSeek just dropped a 75% discount on its primary model—aggressive positioning in a crowded market. Maslej offered context: "To me, this is all evidence of a technology that's still in very early days in terms of not only its capability, but also how it's priced." That's generous. The reality is simpler: customers are balking at invoices they didn't budget for.

Key Takeaways

  • Uber blew its 2026 AI budget in four months; COO says costs getting harder to justify
  • Agentic AI workflows can consume 500–1000x more tokens than simple queries, per Gary Marcus
  • Jensen Huang's March advice: $500K engineers should spend $250K/year on tokens. That advice aged poorly.
  • Uber responded with $1,500 monthly caps per employee per coding tool—not a sign of confidence
  • AI providers are reshuffling pricing models as customers push back on token bills

The Bottom Line

The tokenmaxxing era is over. Companies aren't stopping AI adoption—they're demanding proof it pencils out. For AI vendors, the next phase isn't about building fancier demos; it's about making inference costs transparent and demonstrating actual ROI. The startups that help enterprises track, optimize, and justify their token spend are going to print money.