Tokenmaxing is out - Frugal AI is the new trend

# Tokenmaxing is out — Frugal AI is the new trend Source: [https://ioplus.nl/en/posts/tokenmaxing-is-out--frugal-ai-is-the-new-trend](https://ioplus.nl/en/posts/tokenmaxing-is-out--frugal-ai-is-the-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\. ## The pitfall of tokenmaxing Nvidia CEO Jensen Huang recently made a[remarkable statement](https://www.businessinsider.com/jensen-huang-500k-engineers-250k-ai-tokens-nvidia-compute-2026-3)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\." It is an extremely dangerous and short\-sighted way of thinking\. [.png&w=2048&q=75)](https://wattmattersinai.eu/) Measuring productivity by raw token consumption is just as absurd as evaluating a programmer by the number of lines written or bugs fixed\. It creates a perverse incentive — it 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](https://www.nytimes.com/2026/06/18/technology/ai-token-minimizing.html)\. 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](https://www.nytimes.com/2026/06/18/technology/ai-token-minimizing.html)\. Nvidia, of course, benefits directly from this waste — the company sells the expensive hardware needed to generate all those[unnecessary tokens](https://www.businessinsider.com/jensen-huang-500k-engineers-250k-ai-tokens-nvidia-compute-2026-3)\. ## 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](https://thenextweb.com/news/tokenminimizing-companies-cap-employee-ai-spending)\. 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](https://www.nytimes.com/2026/06/18/technology/ai-token-minimizing.html)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\. Blindly fixating on massive,[expensive American models](https://www.reddit.com/r/theprimeagen/comments/1ttpnn7/chinese_ai_is_30x_cheaper_than_claude_and_chatgpt/)proves economically unsustainable, especially as Western experts warn that token prices will rise sharply once cheap venture capital in Silicon Valley dries up\. ## Choose the right tool for the job Tokenminimizing requires a fundamentally different and more mature approach to AI\. At its core, it means choosing the right model for the specific task\. Many companies currently use the heaviest language models on the market for relatively simple tasks — like using an industrial press to push a small nail into a block of wood\. It works perfectly well, but it's completely over the top and extremely expensive\. A smart IT strategy looks critically at what the user actually needs\. Sometimes, a locally installed model is far more logical and secure\. A local server setup costs a one\-time investment of[around $10,000](https://www.reddit.com/r/theprimeagen/comments/1ttpnn7/chinese_ai_is_30x_cheaper_than_claude_and_chatgpt/)— in stark contrast to daily cloud service costs that can easily reach $500 per engineer\. 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](https://mistral.ai/news/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](https://medium.com/@AiDocTakes/mistral-small-4-the-open-source-model-that-does-everything-and-costs-you-nothing-to-license-539818925327)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](https://huggingface.co/mistralai/Mistral-Small-4-119B-2603)— low for a simple text summary, high for complex code\. The green search engine Ecosia recently[made the strategic switch to Mistral](https://support.ecosia.org/article/1006-ai-search), 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\. ## The ecological necessity of frugality The shift to tokenminimizing is not just a financial necessity — it is, above all, crucial for the preservation of our environment\. 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](https://wattmattersinai.eu/)\. A recent UN report paints an extremely alarming picture: data centers consumed an estimated 448 terawatt\-hours of electricity worldwide in 2025, with AI accounting for a full 20% of that total\. 80% of this energy is spent simply answering everyday user queries\. Europe is responding with strict regulation\. The European data center sector must be fully climate\-neutral by 2030, and large tech companies must make their ecological footprint fully transparent under new European sustainability laws\. Blindly burning tokens simply no longer fits this new reality\. Companies that cling to tokenmaxing will soon run hard up against ecological and legal limits, and local governments are already imposing far stricter requirements on new data center developments due to the enormous strain on the power grid\. ## 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\.