What AI usage is really telling us about enterprise adoption

Many public AI conversations still revolve around leaderboards: which model ranks highest, which provider is 'winning', and which benchmark score matters most.

But for organizations building AI in production, those questions are becoming less useful than understanding how AI is actually being used.

One reason is the growing debate around "tokenmaxxing" - the practice of maximizing AI usage, often driven by internal adoption targets, incentives or leaderboard-style competitions.

But focusing on token consumption can encourage activity for activity's sake, rather than measuring the business value AI actually delivers.

Recent examples, including Amazon reportedly shutting down an internal AI leaderboard and Uber capping employee AI spending after rapidly exhausting its annual budget, highlight the risks of treating usage as the primary success metric.

Yet the latest data suggests something more nuanced is happening.

According to Vercel's July AI Gateway data, token volume grew by 29% in June while spend increased by 27%, with the average price per token remaining flat.

Rather than simply consuming more AI, organizations are becoming deliberate about where they deploy different models and how they balance cost with performance.

The teams building real AI systems are increasingly routing tasks dynamically across multiple models depending on cost, reliability and reasoning.

For example, a low cost model may be useful for handling summarization, while a premium reasoning model is reserved for high stakes decisions.

In practice, AI is more about orchestrating layers across many models than building systems around a single provider.

The latest AI Gateway data shows organizations distributing workloads across different classes of models rather than relying on a single provider.

Open-weight models now process 29% of gateway tokens while accounting for less than 4% of spend, while frontier models continue to dominate higher-value reasoning workloads.

Rather than taking a one-size-fits-all approach, organizations are selecting different models according to the task at hand.