Tokenmaxing is out — Frugal AI is a new trend

The era of blindly burning through AI tokens is over. Major tech companies are realizing that unlimited use of artificial intelligence leads to sky-high costs and enormous ecological damage. The controversial trend of “tokenmaxing” is rapidly giving way to a mature counter-movement: tokenminimizing. Companies are once again choosing efficiency and purposefulness.

Generating unlimited AI tokens consumes enormous amounts of electricity. The international conference “Watt Matters in AI” in Eindhoven is putting this rapidly growing problem firmly on the agenda. 

The pitfall of tokenmaxing

Nvidia CEO Jensen Huang recently made a remarkable statement at a conference in Silicon Valley. He argued that an engineer earning half a million dollars should spend at least half of that on AI tokens. Huang sees this extreme consumption as a direct measure of human productivity. This philosophy has become known in the tech world as “tokenmaxing.” 

This encourages employees to deliberately set up inefficient processes just to hit their unofficial quotas. Workers had AI assistants routinely generate superfluous code just to climb internal leaderboards. This phenomenon perfectly illustrates Goodhart’s Law: once a metric becomes a goal in itself, it immediately loses its value. The result in the workplace was an explosion of useless data and towering bills. Companies saw their software costs triple with no measurable increase in actual output or innovation.

The financial hangover in the tech sector

Unchecked token consumption quickly led to an unprecedented financial hangover. Major tech companies were shocked by their monthly bills from AI vendors. A single user at Anthropic managed to burn through $150,000 worth of tokens in a single month using the programming tool Claude Code.

Transport company Uber had already exhausted its entire 2026 AI budget by April, forcing it to immediately impose hard limits — employees may now spend a maximum of $1,500 per month per tool. Giants like Meta and Walmart also intervened drastically: they dismantled their internal AI-usage leaderboards immediately and switched to strict cost controls.

This abrupt shift marks the definitive end of the tokenmaxing era. Companies are now moving en masse to the counter-movement called “tokenminimizing” — a business strategy focused entirely on efficiency rather than sheer volume. 

Forward-thinking companies now actively measure token efficiency rather than raw consumption volume. Simple queries are automatically routed to small, fast models; only highly complex analytical problems go to the heavy systems. This targeted routing prevents unnecessary waste of expensive computing power and protects the IT budget.

European efficiency with Mistral

A perfect example of this necessary efficiency drive is the European model Mistral Small 4, which currently ranks extremely high on the price-performance scale. It contains 119 billion parameters in total, but activates only 6 billion per generated word, thanks to a highly intelligent architecture. The model delivers top-tier performance while producing considerably shorter and more concise answers than the competition. In complex reasoning tests, Mistral Small 4 needs only 1,600 characters to give a correct answer, while comparable models like the popular Chinese Qwen require nearly 6,000 characters. Since customers pay per generated token, this conciseness translates directly into major cost savings.

Developers can also manually adjust the required computing power per individual query — low for a simple text summary, high for complex code. The green search engine Ecosia recently made the strategic switch to Mistral, leaving market leader OpenAI specifically to drastically reduce its energy consumption. This real-world example demonstrates conclusively that smaller, efficient models perform excellently in demanding production environments.

A tangible impact on autonomy

The strategic shift to tokenminimizing has direct and very positive consequences for the European economy and autonomy. By consciously choosing efficient, open models, European companies become far less dependent on expensive American cloud services, significantly strengthening much-needed digital sovereignty. Companies that successfully switch to smart model selection report impressive cost savings of 60 to 90%.

The European AI platform Neurometric is already actively guiding organizations through this complex transition, helping companies effectively consolidate their fragmented software infrastructure. Using lighter models directly means fewer servers and far lower operational complexity for IT departments.

The future of artificial intelligence does not lie in building ever-larger, power-hungry systems. The winners of the near future will be the companies that achieve maximum business results with minimal technological resources. Tokenminimizing is forcing the entire tech sector to finally grow up — shifting the focus from brute, wasteful computing power to smart, purposeful innovation. Using AI frugally is not a passing trend; it is the only financially and ecologically sustainable path forward.