TOKENMAXXING
A recent technology business news article by the Guardian said that while many companies like Amazon, Meta, Uber and Walmart have no limit on staff use of AI, in fact some encourage it like Sequoia Capital does for instance, others like Atlassian have introduced a limit on AI spending.
“Atlassian tightens tracking of staff AI use as other technology firms encourage ‘tokenmaxxing’.
Now I know who Atlassian is, and Amazon, Meta, Uber, Walmart and Sequoia, and I know what AI is, but what on earth is “tokenmaxxing”? I had no idea. Never heard of it.
Back to basics.
Initially a software program, like a DOS operating system or a programming language like BASIC or a game like Zork, could be bought. You owned it. Forever. The products were bug free. Otherwise their product life cycle would be short as word got out. Admittedly they were small programs compared with today, but the product was finished, done and dusted, and sold. I could probably pull out an old rusty PC, boot it up even today, write a Pong game or type a letter and print it out, and if the modem still worked, transmit it over a copper phone line to somewhere. If I pulled out one of my later computers I doubt Windows 3.14 would still work. Actually I bet it would and I would be wondering why I was forced to buy some ten updated versions after that. (Zork by the way used a natural language interface, back in 1977!)
As software products became more sophisticated and complex, how to sell them became equally complex.
Computers, or devices as they tend to be known now; software programs, or apps as they tend to be known now, and more recently the Internet, and even more recently the Cloud, and yet more recently the AI revolution, have become so useful, so convenient, so ubiquitous, that the industry has exploded, all summed up so far as what we call the Information Age. The progression is not finished yet, proven with the advent of AI. There was the digital age, the computer age, the information age, the artificial intelligence age. Something will come next. The Industrial Revolution began around 1712 with Newcomen’s engine, but it wasn’t until a century later that the expression appeared. Even the mine water pump of Newcomen was preceded a century earlier by Jerónimo. No, not that Geronimo, although they look similar, but Jerónimo de Ayanz Beaumont, a Spaniard who really was the one to invent the steam engine to pump mine water. The point is it takes time, so who knows what we will eventually call what is happening now. Time will tell. But just as simple boiler steam pumps became steam trains, were replaced by internal combustion engines and electric motors, telegraphs by telephones, gas by electricity, log books and slide rules by computers, our current age has seen similar albeit much more rapid evolution.
Unlike the Industrial Revolution era, transitioning from millennia of an agrarian economy, we are in the midst of a global market economy. Manufacturing and selling stuff are what we do. So the competition inherent in our current revolution has created more and more products, bigger and bigger, greater and greater sophistication, wider and wider applications, and along with it, competing ways to sell the products and services.
Creating a new technology, and a product or a related service is tough enough, but most people would be surprise at just how problematic it is to put a price on something new. A whole scientific industry sits around trying to predict what a customer will pay for something.
It’s a challenge. Instead of selling a program on a floppy disk, how do you sell a giant program, so large it is never really ever bug-free or finished, and it takes a bank of computers to transfer it online. And a method for keeping it running? Well the answer to that was solved, sell licences that come with maintenance contracts. That turned out to be a cool alternative, because you end up reselling the same product basically each time an annual licence is renewed. Then the next challenge, how to sell not a software program, but a software serviced that you retain and store in your own warehouse. But with professional SaaS or even just high-end computer games the scale can become something so complex it needs a team of experts to keep it running and another team to tell the user how to use it? Again the answer is straight forward, a subscription, a bit like a licence, but with built in usage counters. You can use it until your time is up. Then came Cloud services such as storage or analytics. In that case the solution was actually not to send the customer anything, keep it at home and allow them access. And then finally comes this last challenge of making money from selling the skills of an artificial intelligence.
That’s where we get back to tokens. 
The current challenge is how to charge for the use of an AI system. Accessing an AI system is like accessing someone’s brain. That’s more true for those believing that AI is approaching human-like abilities. There is a similarity there. When you pay for the brain of say a doctor or a psychiatrist, or an engineer or architect, you are accessing their intelligence. You are taking advantage of all those years of training they undertook. Those specialists know how to charge. Often by time, sometimes by procedure, sometimes throughput, like a freelance journalist being paid by the word. Between $0.25 and $0,50 per word.
Well that is what a token is. Not exactly a word, but a set of letters, usually around 4. Sometimes a word if it is no more than 4 letters.
So like a Cloud service, which relies on the use of your supplier’s hardware sitting in a giant warehouse on some remote location, banks of thousands of computers and servers, allowing you to use unlimited computing power and memory storage, current AI systems are similarly giant hardware based centres, electricity needs equivalent to a town and water cooling needs similar to a nuclear reactor. Clearly a customer cannot buy a “Cloud” only a Cloud service, and same with using AI. You pay for access.
So what kind of access and what kind of payment strategy?
AI companies use all of the aforementioned methods of charging. Subscriptions, licences, resource usage, access charges. Not to be outdone though, they have invented a new method. They charge for the actual text flow back and forth.
It’s a bit like that journalist payment scheme mentioned earlier. Except instead of paying for each word in a final proof, sometimes the payment is based on every version sent through and every response. So you ask AI a question, it replies, always with a prompt for more, and now you know why, then you ask for clarification or a different question, and it comes back, and you pay for everything transpiring back and forth. The text string is broken into units of four characters, and each unit is called a token. There it is. What a coup for the AI companies.
In order to ensure staff made full use of AI, companies instituted a protocol of rewarding staff for the amount of AI usage they recorded. Token maximisation. Which becomes tokenmaxixing.
But any fool can see that this is an open ended argument. It reminds me of the person who took a job as a salesperson at a sports and camping store. With no experience the company was reticent to hire him, but gave him a chance. After the store closed, the boss asked, “How many sales today?” The man says, “One.” The boss says, “One? We average 20 to 30 sales a day. How much was the sale for?” The man says, “$101,237.64.” Boss says, “$101,237.64? What the hell did you sell?” “Well first I sold him a small fish hook. I asked him what size fish he planned to catch and he did not know, so then I suggested he buy a medium fish hook and a larger fish hook. Then I sold him a new fishing rod. I asked him where he was going fishing and he said ‘at the coast; so I told him he was gonna need a boat. We went down to the boat department and I sold him that twin engine. He said he didn’t think his Honda Civic would pull it, so I took him down to the automotive department and sold him that 4X4 Blazer.” The boss said, “A guy came in here to buy a fish hook and you sold him a boat and truck?” Man says, “No sir, he came in here to buy a box of tampons for his wife and I said, ‘Well, your weekend’s shot, you might as well go fishing”.
AI companies love tokenmaxxing and do everything to encourage it. But finally the end user companies promoting tokenmaxxing have had second thoughts. That was what the Guardian article was all about.
‘Software firm Atlassian has sought to tighten trafficking of its staff’s AI spending by introducing “wallets” with monthly caps of up to $2,000 for each employee, amid an explosion in costs at other tech companies.
‘The move by Atlassian, which recently cited AU as part of the reason behind cutting 1,600 staff, bucks the trend of others in the tech sector who encourage employees to use as much of the technology as possible, dubbed “Tokenmaxxing”. Some companies have reportedly introduced leaderboards for employees who used the most AI in their work.
‘Tokens refer to the measurement of a response AI gives to a prompt. OpenAI has said one token is about four characters, and something like the US Declaration of Independence amounts to 1,695 tokens. OpenAI’s flagship model, GPT-5.6, has a charge of US$5 for every 1m tokens, while Anthropic’s Claude Fable and Mythos models are US$10 for every 1m tokens.
‘Under tokenmaxxing, the costs quickly add up.
‘Uber reportedly blew through its AI budget in four months, and Amazon has reportedly told employees to stop using AI just for the sake of AI.’
This clearly will become a factor in the current moves in Australia to charge AI companies for the material used to train AI systems. Clearly if an AI company is charging its customer to access the text of say a lecture, or a novel, or a news report, they are using someone else’s creative content to make money. And if tokenmaxxing is any indication, lots of money.
Speaking at the Andrew Olle Media Lecture, Reuters editor-in-chief Alessandra Galloni said AI could support journalism, but warned original reporting must be protected and paid for.
Australia does not have a standalone AI law; instead, it applies existing laws, including copyright and privacy regulations, to AI technologies. New mandatory Australian AI Standards are expected to be established by early 2027, which will include specific obligations for AI companies regarding copyright protections and data usage.
Maybe journalists and authors and news agencies could think of tokenmaxxing AI companies?