In the world of corporate finance, there is an unspoken rule: if resource consumption grows, the budget follows suit. But Coinbase CEO Brian Armstrong is clearly demonstrating that this rule can and should be broken. The company has managed to nearly halve its spending on artificial intelligence, despite exponential growth in token consumption. How is this possible? The answer lies not in cutting limits, but in radically optimizing engineering processes.

Not Restrictions, but Smart Routing

The key takeaway from Coinbase's strategy is this: total control and rigid quotas are a dead-end path. Instead of setting boundaries for employees, the company focused on "correct default settings." Armstrong emphasizes that 91% of employees never hit their limits, so lowering thresholds would have been pointless. The solution was found in intelligent request routing. The system itself determines which AI model to use for a specific task. For strategic planning, the most powerful (and expensive) models are used, while for routine execution, lighter and cheaper ones are employed, including open-source solutions like GLM 5.2 and Kimi 2.7.

Cache and Context: The Two Pillars of Savings

Armstrong highlighted two main tools that allowed the company to break the correlation between consumption and costs. The first is caching. Coinbase restructured its architecture to reuse already processed data. In the LibreChat service, the cache hit rate increased from a catastrophic 5% to an impressive 60%. A cache miss, according to the CEO, is the fastest way to burn through a budget wastefully. The second tool is context savings. The company implemented the practice of starting new sessions when switching tasks, strictly limiting file context, and disabling unused tools. The goal is not to spend fewer tokens, but to spend them wisely, avoiding overspending on unnecessary information.

Deutscher's "Barbell" Strategy

Analyst Miles Deutscher describes a similar approach, calling it "token engineering" and a "barbell" strategy. His method proposes delegating 10% of the most complex tasks (planning and final review) to the smartest models like Opus or GPT. The main 80% of routine work is handled by cheap open-source models. This approach, he says, allows for a 50% or more reduction in AI costs, and he has been using it for several months. This is the best proof that efficiency is not about stinginess, but about smart architecture.

Expert Commentary: The Coinbase story is not just an optimization case study, but a manifesto for a new era in corporate AI. The market is shifting from a race for the "smartest" model to a race for the "most efficient architecture." Companies that are the first to implement intelligent routing and aggressive caching will gain a tremendous competitive advantage, leaving competitors to burn their budgets on meaningless computing power.