Trending Topics
Trending
The era of Tokenmaxxing is ending: why reckless data burning is giving way to harsh economic reality
This article is part of our Opinions section, where we invite industry professionals to share their views on the most pressing technology questions of our time.
We have just lived through a fascinating, wild, and incredibly expensive two years with artificial intelligence. The corporate world fell into a collective frenzy known in tech circles as “tokenmaxxing”. This phase of reckless experimentation, during which management gave employees unrestricted access to advanced language models for passive consumption – such as uploading documents and asking questions – only to later stare incredulously at astronomical bills, is finally hitting a wall.
The true financial reckoning is just arriving. As we enter an era of sobering up and financial discipline, a new and far larger cost driver is emerging: Agentic AI. Unlike human-in-the-loop workflows, these autonomous systems generate their own context in loops and execute multi-step tasks, consuming orders of magnitude more tokens per task. The CFO crackdown is already underway, shifting the focus to precision optimisation before this next wave of autonomous waste spirals out of control.
Imagine a scenario that has played out in dozens of corporations across Europe and America over the past few months. An eager data analyst or project manager takes a hundred-page annual report, uploads it into GPT-5.5 or Claude Fable 5, and asks a basic question about what the net profit was in the third quarter. The model obediently responds with a single sentence, and the task is accomplished.
What the employee does not see, however, is the backend mechanics. Every single time, the model had to reread the massive context window, consuming tens of thousands of input tokens. As a result, the company paid vastly more for that single query than if it had just looked it up in a standard corporate database. Providers are now addressing some of this waste at the API level through prompt caching, which makes repeated context cheaper, but the rest comes down to the internal architecture of the company itself.
Multiply this behaviour by thousands of employees and hundreds of thousands of queries a month, and you get the phenomenon we now define as tokenmaxxing. Its core was not just waste, because token consumption actually became a reporting metric in many companies. Management tracked it as an indicator of innovation and productivity, and some firms even created internal leaderboards to see who could burn the most tokens.
However, when a metric becomes a target, it ceases to be a good metric. Higher token usage does not mean more work done; it simply means a higher bill. It was an era that ignored efficiency, where the sole goal was to be part of the trend, demonstrate activity to investors, and inject artificial intelligence into every conceivable workflow regardless of whether it made economic sense. This wild west of blind experimentation is coming to an end as chief financial officers worldwide take the reins and demand transparency.
Illusory abundance and the Uber effect in AI infrastructure
The reason tokenmaxxing happened in the first place does not lie solely in corporate naivety. The entire generative artificial intelligence market is currently going through a phase reminiscent of the early days of ridesharing or food delivery apps. We are witnessing a brutal battle for market share. Model labs like OpenAI, Anthropic, and Google, which are backed by tech giants and venture capital, are pouring tens of billions of dollars into infrastructure. They keep token prices artificially low to lock in enterprise customers before their competitors can, since the race is on to become the new operating system for the enterprise world.
Yet, focusing solely on potential future price hikes misses the real crisis unfolding right now. While it is true that current API pricing for frontier models is heavily subsidised by investor capital, shareholder pressure will eventually force a realignment with actual operational costs. The immediate financial pain is thus driven by sheer volume.
The data tells a striking story: the average cost per million tokens across major providers actually plummeted from roughly $10 to $2.50 in a single year. Yet, enterprise AI bills have skyrocketed by an estimated 320%, with the average enterprise AI budget ballooning from $1.2 million in 2024 to $7 million in 2026. The true crisis isn’t just that prices might eventually rise; it’s that consumption is exploding exponentially faster than unit prices are falling. Open-ended use and unmonitored workflows are driving a volume surge that completely wipes out any efficiency gains from cheaper infrastructure.
The rise of financial discipline and capped budgets
It is precisely this threat of future price increases, combined with the current shock of uncontrolled spending, that is forcing a paradigm shift in the enterprise sector. The days when every developer or product manager had a blank check to call third-party APIs are gone. Strict financial controls, cost allocation to specific business units through AI FinOps, and strictly capped budgets are entering the game.
Optimisation is no longer a dirty word but a core competency. Companies are starting to realise that they do not need the biggest, hungriest, and most expensive model on the market to solve 80% of their daily tasks.
Enter the concept of precision engineering. If a company needs to extract data from invoices, there is no need to ask for an all-encompassing frontier model. The open-source landscape has matured dramatically. DeepSeek’s arrival in early 2025 shattered the long-held assumption that capability and cost were inseparable. Today, deploying open-source powerhouses like Llama 4, Mistral, or DeepSeek hosted securely on internal cloud infrastructure represents a genuine architectural choice, not a budget compromise. It is tailored for the task and orders of magnitude cheaper.
However, relying purely on market forces to lower your enterprise AI bill is a dangerous gamble. Falling inference costs are by no means guaranteed to reach your budget. As Gartner recently warned, AI and cloud providers are under immense pressure to deliver profitability; they are highly likely to use cheaper infrastructure to improve their own margins first, rather than passing those savings down to reduce customer bills. True precision engineering means taking control of your own infrastructure choices.
The ideal business window: from chaos to maturity
Some might argue that stricter controls and budget caps will put the brakes on corporate development, but the opposite is true. Right now, the most exciting business window in years is opening up because both the technology itself and internal corporate know-how have finally reached maturity. We are no longer reinventing the wheel; we know what works.
In the first wave between 2023 and 2025, companies built proof of concept projects on the fly, often resulting in fragile code, vendor lock-in, and massive resource waste. Today, however, engineering teams possess advanced tools for model orchestration, semantic caching to pull identical queries from fast memory rather than resending them to the language model, and intelligent query routing.
The math behind the savings is simple. Take a company with one hundred thousand internal queries per month. If all of them are routed to an expensive frontier model, the average query costs a few cents, bringing the monthly bill to tens of thousands of euros. If we introduce semantic caching, which catches roughly 10% of repeated queries virtually for free, and routing, which directs another 30% to a model that is ten times cheaper, the bill drops by about a third, saving around two hundred thousand Czech korunas (approx £7,000). This is achieved without any loss in output quality because complex queries are still handled by the best model.
How properly managed enterprise projects deliver savings
Shifting from tokenmaxxing to a managed architecture yields immediate, tangible results. A well-designed enterprise project can reduce operating costs by 30 to 40% without compromising output quality or slowing down innovation. The strategy rests on three core pillars.
First, semantic caching addresses the fact that employees often ask very similar questions within internal systems. By implementing a layer that recognises when a semantically identical query was answered recently, the system can serve the user a pre-cached response. The cost of such a query is virtually zero, and the response time is instant.
Second, dynamic model routing ensures that every user query first passes through a very cheap classifier, often a small open-source model, that assesses its complexity. A simple translation or text extraction is routed to a lightweight model like Claude Haiku or Mistral Small. A complex strategic analysis requiring deep logical reasoning is sent to an expensive frontier model only when necessary. The result is a drastic reduction in the average cost per thousand tokens.
Third, context window optimisation and retrieval-augmented generation replace the reckless dumping of massive documents. These systems search the corporate database to find only the specific paragraphs relevant to the answer, presenting the model with a surgically precise minimum of data so that token consumption drops dramatically.
The future belongs to architects, not consumers
The end of tokenmaxxing does not mean we should dial back artificial intelligence in the workplace. It simply means the era of romantic amateurism is over. The winners will be the companies that stop looking at AI as a magical black box where you pour in money and data and hope for a miracle.
Artificial intelligence is becoming a standard infrastructural commodity, and like any commodity, we must learn to handle it efficiently, economically, and strategically. The business window opening up right now favours those who can combine technological vision with uncompromising financial discipline. Those who successfully navigate this transition will gain a massive competitive advantage backed by substantial bottom-line savings. Much like other maxxing trends born on social media, this one was toxic, addictive, and, above all, unsustainable.
Related articles
About The Author
