Daybreak - Why Big Tech is tokenmaxxed out

Episode Date: June 4, 2026

Amazon built a leaderboard to track how much AI its engineers were using. Employees gamed it. Costs exploded. Last week, the leaderboard was gone.Uber burned through its entire annual AI budget in four months — after telling staff to use AI "as much as possible." Microsoft cancelled most of its Claude Code licences six months after rolling them out.Three companies, the same couple months, the same lesson: that measuring AI adoption is turning out to be a very different thing from measuring AI productivity.Tune in.Daybreak is produced from the newsroom of The Ken, India’s first subscriber-only business news platform. Subscribe for more exclusive, deeply-reported, and analytical business stories.

AI skeptics everywhere are having a major, I told you so moment. Over the last couple weeks, we've been seeing how one of the biggest trends in AI adoption is backfiring badly, generating big bills for big tech. The trend I'm talking about and what you must have already heard of is token maxing. Just a few weeks ago, it seemed like it was the thing to be doing inside many tech companies. The basic idea was that to track real innovation and productivity, one could track an employee's token usage. Tokens, by the way, are the units of data that AI models process. And while different companies measure tokens differently,

it's generally considered to be equivalent to about a word and a half in the English language text. All the famous tech companies like Meta, Amazon, Open AI, and others had even set up formal or informal leaderboards to record token. usage. They were even encouraging engineers and developers to compete with each other, to see who could use the most tokens. But all of that has started falling apart. Just last Friday, Amazon announced that it was scraping its internal leaderboard called KiroRank. Financial Times reported that employees were trying to boost their scores artificially through unnecessary activity, and that was driving up Amazon's compute costs. Even Microsoft has cancelled most of its

its internal clod code licenses in its experiences and devices division. This move is supposed to be effective from June 30th, 2026 onwards. The reason is pretty much the same. The token-based billing system consume the company's annual AI budget far ahead of schedule, and so the company is redirecting its engineers to its own AI coding assistant called GitHub co-pilot CLI. Now, the billions worth of spend on AI, specifically about $725 billion, is based on a promise. That number, by the way, is the combined amount that Amazon, Microsoft, Google and Meta are all spending just this year on AI infrastructure. Things like data centers, chips, compute power.

And the promise, the very large bet on the table is that AI will be so productive, so transatlanticer. transformative that the returns will justify the cause. That AI will eventually be cheaper and faster and better than human labor. But this month, at least three companies have shown us whether that bet is paying off. And the answer, if we're being direct, is no one is sure yet. Welcome to Daybreak, a business podcast from the Ken. I'm your host, Tresia Virgis, and every day of the week, my co-host, Nikita Sharma and I will bring you one new story that is worth. understanding and worth your time.

Today is Friday, the 5th of June. Now, Amazon's official claim is that the leader board was not a formal one. The company had reportedly also told its employees that AI token statistics would not be tied to performance evaluations. But that did not stop employees from believing that managers were still monitoring the data. A current employee told Financial Times mint last month that he believes the managers who track the numbers are creating perverse. incentives, causing some people to be very competitive about the usage.

One employee even admitted to 404 media that they cheated their way up the leader board after they were told in a performance review that they're not using AI enough at work. Another claim that the pressure to use AI tools was just so much and was the reason why several employees were using AI agents to maximize their token usage. Dave Dreadwell, a senior vice president at the company, maintained that while the leader board had been built with good intentions, the token maxing was costing Amazon. He reportedly told stuff, please don't use AI just for the sake of using AI. Now, Amazon is not the only company to have suffered through this kind of an arc.

In April this year, Praveen Nepali Naga, the CEO of Uber confirmed that Uber had burned through its AI budget for an entire year in just four months. And guess what? This was apparently after the company Encourage staff to use AI as much as possible. The information also reported that the company even had a similar leaderboard that was tracking internal usage. Just this week though, Bloomberg reported that the company

now has a spending limit per employee, $1,500 per month per tool. And not just that. Andrew McDonald, the company C.O. express some doubt in the more AI use means more productivity hype in a recent podcast. He said, and I'm quoting here, maybe implicitly there's more that is getting shipped. But it's very hard to draw a line between one of those stats and say,

okay, now we're actually producing like 25% more useful consumer features. Just imagine. This is a COO of a company that blew through a year's budget in four months. is forced to set a spending limit on its employees and now he's still struggling to justify the cost. What both Amazon and Uber both illustrate here is something that analysts are now pointing out explicitly. They call it Goodhart's Law. It basically says that when a measure becomes a target, it stops being a good measure. So the moment token consumption became tied to something employees could see and compete about,

it stopped measuring productivity and started measuring something else. Something more like anxiety. Specifically, the anxiety of being seen as someone who isn't using AI enough. With Microsoft, though, the story goes a little beyond just employee usage exploding because of anxiety. Stay tuned. In mid-May, Microsoft began canceling Claudecone licenses for its experiences and devices division. This is a group that builds Windows, Microsoft 365, outlook and teams, by the way.

The company's engineers were told to switch to GitHub co-pilot CLI by June 30th, which is the last day of Microsoft's fiscal year. And this is just six months after the company had even opened up access to ClaudeCode for his employees. The Wurge reported that canceling ClaudeCode licenses before this date was basically an easy way for the company to cut some operating expenses for when the new financial year starts in July. But the official internal memo from Rakeesh Jha, the company's EVP, said something a little different.

He wrote, Claude Code was an important part of learning. At the same time, co-pilot CLI has given us something especially important. A product we can help shape directly with GitHub for Microsoft's repos, workflows, security expectations and engineering needs. So basically, what that means is that engineers have now learned what they need to from using tools like Claude, and now they need to focus on unifying internal tool chains. Just one small hitch in that plan though. The engineers preferred ClaudeCode over co-pilot's AI. The adoption was quick and its rate was high,

and that combined with the token-based pricing meant that the cost escalated much sooner than the company's budget projections had anticipated. Another reason is that the rising internal popularity of Claude was spelling a branding problem. You see, Microsoft can't possibly sell co-pilot with any kind of credibility to its enterprise clients while its own engineers are abandoning it for something else. And now, the larger irony in this whole story is that usage prices are actually increasing

even as individual token prices themselves are going down. But that's not the same as AI getting cheaper to use, at least not for enterprise clients. Actually, a Gartner report published recently put it quite plainly. Cheaper tokens don't really translate to cheaper enterprise AI. And that's because agentic models require far more tokens per task than standard models. Plus, increased consumption can actually outpace the following unit costs. And AI providers won't fully pass on those cost reductions to enterprise clients. Even a pure-reviewed paper published just this April noted that agentic tasks actually

consume up to thousand times more tokens than standard code chats too. What's worse is that the study found that higher token usage does not really mean higher accuracy. And apparently, accuracy peaks at intermediate costs and saturates at higher costs, which means you can spend more and not necessarily get a better result. And that explains why not everyone was buying into this trend. Take Rabi Kumar as the CEO of Cognisant, the American IT company. for example. He called token maxing a vanity metric just this Monday. His argument was basically

that knowing when to use AI and when not do is the actual skill. And the companies that figure that out will outperform the ones who are gamifying adoption with, say, leaderboards. Now, the promise of AI has always been that it would be cheaper than human labor. What this month showed is that things aren't all that straightforward and that AI is in fact, becoming more expensive because of how companies have chosen to incentivize its adoption by measuring the wrong thing, rewarding quantity over quality, and introducing competition and anxiety where it need not have been. Daybreak is produced from the newsroom of the Ken India's first subscriber-focused business news platform.

What you're listening to is just a small sample of our subscriber-only offerings. A full subscription offers daily long-form feature stories, newsletters, and a whole bunch of premium podcasts. To subscribe, head to the ken.com and click on the red subscribe button on the top of the Ken website. Today's episode was hosted and produced by my colleague Rachel Vargis

and edited by Rajiv Sien.