Uber is officially signaling the end of the „tokenmaxxing” era, a period defined by rampant, indiscriminate AI usage in the enterprise. In a recent post on X, Uber CTO Praveen Neppalli Naga highlighted „very interesting trends on AI costs,” noting that this marks a pivotal shift away from the trend that emerged in early 2026. Tokenmaxxing previously encouraged employees to integrate AI into as many workflows as possible, with some companies even tying AI usage to performance reviews. However, Uber’s latest data suggests a maturation of the market, moving from a focus on quantity to a focus on efficiency and strategic deployment.
The shift at Uber is driven by impressive operational metrics. Since the beginning of the year, the number of employees utilizing frontier AI tools has quadrupled, yet this surge has coincided with a significant decline in per-token costs. Naga attributes this cost reduction to specific engineering improvements, including enhanced prompt caching processes, the adoption of better default models, and granting engineers greater visibility into their AI consumption patterns. By experimenting with open-weight models, Uber has managed to optimize its spending while scaling its usage, proving that higher adoption does not necessarily have to equate to skyrocketing expenses.
Uber’s finance chief, Balaji Krishnamurthy, echoed these sentiments during the company’s second-quarter earnings call, emphasizing the ability to deliver productivity gains cost-efficiently. He noted a doubling in code output for engineers, underscoring that the next phase of AI adoption is about efficiency and tangible results. This marks a stark contrast to earlier in the year when Uber made headlines for blowing through its 2026 budget for Anthropic’s Claude Code. By May, COO Andrew Macdonald had already expressed concerns about the trade-offs of AI investments, stating that proportional productivity gains were not materializing relative to the skyrocketing costs.
This pivot is not unique to Uber; the broader tech industry is grappling with how to secure better returns on inflated AI expenditures. Companies like Coinbase are experimenting with model switching, assigning complex tasks to frontier models while offloading routine jobs to cheaper alternatives. This industry-wide recalibration has also birthed a new wave of consultancies and infrastructure startups focused on helping enterprises scale their AI cost-effectively. As Naga succinctly put it, the future will not be characterized by who spends the most tokens, but by how efficiently they are used—a clear departure from the excesses of the tokenmaxxing era.
Ez a cikk a Neural News AI (V1) verziójával készült.
Forrás: https://www.businessinsider.com/uber-cto-praveen-neppalli-tokenmaxxing-era-end-2026-8#article.