Leading technology firms, including Meta, Nvidia, and Databricks, are establishing a unique culture where engineers are celebrated for maximizing their expenditure on AI tokens. These tokens represent the fundamental unit for processing data when an AI model generates text.

The New Metric: Token Consumption as Productivity Indicator

The prevailing sentiment among these tech giants is that high token usage directly correlates with superior output. Meta CTO Andrew Bosworth stated regarding this spending, "This is easy money... No limit." This approach contrasts sharply with widespread corporate belt-tightening and layoffs occurring elsewhere.

Databricks Champions High Usage

Databricks CEO Ali Ghodsi highlighted an engineer who utilized over $7,000 in AI tokens within a two-week span in January. Ghodsi used this example during an address to his engineering department not to criticize, but to encourage greater adoption. The engineer accessed models from Anthropic and OpenAI via Isaac, Databricks' internal coding tool.

Databricks, valued at $134 billion, is part of a Silicon Valley trend pushing developers to fully embrace AI-generated code. Companies are quantifying this adoption by tracking the monetary value of tokens consumed, aiming to maximize productivity.

Understanding Token Costs

The cost of these atomic units varies significantly based on complexity and model quality. Basic tasks on cheaper models might cost cents per million tokens. However, premium models for complex computations can range from $20 to over $100 per million tokens.

For instance, Anthropic charges $25 per million "output" tokens for its flagship model, Claude Opus 4.6. This demonstrates the potential scale of expenditure when engineers utilize the most advanced tools available.

Startups and Giants Set Aggressive Spending Goals

Smaller AI-focused companies are also aggressively pushing token budgets. Alven, a real estate AI startup with fewer than ten employees, reported spending $16,000 on tokens in February. Cofounder Julio-Cezar Scerbina announced a goal to reach $60,000 the following month, believing winners will be those who spend the most tokens.

At the $1.9 billion enterprise AI startup Writer, even non-engineering staff are accumulating billions of tokens used. The push is intense across the industry to ensure maximum AI integration.

Nvidia and Meta Set High Benchmarks

Nvidia CEO Jensen Huang has publicly tied high salaries to high token consumption. Speaking on the All-In podcast, he warned that he would be "deeply alarmed" if a $500,000 engineer spent less than $250,000 worth of tokens annually.

Meta's Andrew Bosworth echoed this sentiment, noting his top engineer was spending the equivalent of their salary in tokens but achieving five to ten times the productivity. He encouraged this behavior without imposing spending caps.

Incentivizing Usage at Sendbird

Sendbird, which provides customer support chat solutions for companies like DoorDash, uses a leaderboard system to rank token spenders from "Beginner" to "AI God." An "AI God" is defined as someone consuming at least 100 million tokens daily, according to CEO John Kim.

The company offers perks like coffee gift cards and company swag for top users, with plans for larger rewards such as extra vacation days. The objective is to maximize productivity by deploying autonomous AI agents to work overnight.

The Impact on AI Service Providers

This explosion in token usage has significantly benefited AI coding services. Anthropic’s Claude Code has seen its annualized revenue surge to $2.5 billion, positioning it as a perceived leader in the space.

OpenAI’s Codex maintains 1.6 million weekly active users. Furthermore, Cursor, a popular tool, reportedly grew its annualized revenue to over $2 billion in the last three months, according to an insider source.

Balancing Spending with Production

While the push for spending is strong, leaders acknowledge the risk of blind consumption. Sendbird CEO John Kim noted that while eight out of the top ten token spenders are usually highly productive, others may be purely experimental.

Kim compared the current environment to the 1990s, when lines of code written were the primary measure of a software engineer. He clarified that token spending is not the ultimate performance metric; the true indicator is how much of the AI-generated code successfully makes it into production.