Tokenmaxxing is over. It was a flawed way to measure a company's ROI from AI.

Hello and welcome to Eye on AI. It’s Jeremy here, filling in for Sharon who is on vacation. In this edition…CNN sues Perplexity…IBM and RedHat form $5 billion bug patching project…Snowflake signs a $6 billion deal with AWS…and the White House gives U.S. intelligence agencies $9 billion to build their own AI chip cluster.

Just a few weeks ago, it seemed that ‘tokenmaxxing’ was all the rage inside many companies. The idea was: if you wanted to find out which employees were being most innovative in deploying AI agents, you should track their token usage. (Tokens are the units of data that AI models process; a token is equivalent to about a word-and-a-half of English language text.) The more tokens expended, the more productive that employee’s AI agents were, or at least, the more AI-forward and innovative that employee was trying to be. That was the idea anyway. Meta, Amazon, OpenAI, and many other companies even established formal or informal leaderboards of token usage and encouraged engineers and developers to compete to see who could use the most tokens in a given period of time.

Of course, Goodhart’s Law still holds (it posits that any measure that becomes a target, ceases to be a good measure) and tokenmaxxing had some predictably perverse results. At Amazon, the Financial Times reported, some employees spun up AI agents to complete wholly meaningless or unnecessary tasks just to keep up their token usage stats, which were now being used by managers to assess employee performance.

Also, all those tokens are hardly free, and some companies have gotten sticker shock from their Anthropic and OpenAI bills. So, now many companies seem to be pulling back from the tokenmaxxing ethos and even limiting which employees can use third party AI agents, at least those that use the most advanced AI models as the “brains” inside the agentic harnesses. Meta took down the informal tokemaxxing leaderboard its employees had created. Microsoft has cancelled Claude Code subscriptions for employees in several key product divisions, according to reporting from The Verge. Uber said it had burned through its entire 2026 “token budget” in just the first four months of the year, in part due to high usage of Claude Code. Meanwhile, Salesforce CEO Marc Benioff has said his company’s Anthropic bill will be about $300 million this year and that he wished there were a “smart router” that could determine which queries actually required the most capable, and most expensive, models and which could be handled by smaller, less-capable-but-capable enough, cheaper alternatives.

Many executives are also saying token spending isn’t translating into firm-wide return on investment. Uber Chief Operating Officer Andrew Macdonald told a podcast last week that the ride-hailing firm has been struggling to connect the boost in the productivity of some workers with any company-wide impact. “If you‘re not actually able to draw a direct line to how much useful features and functionality you’re shipping to your users,” he said. “[The token costs are] harder to justify.” The net result is that the days of tokenmaxxing are over.

Why AI spend is still not producing ROI

But that still leaves the broader question of why this disconnect exists between AI spend and ROI? Certainly explicitly rewarding tokenmaxxing doesn’t help, since it fails to align employee incentives with company goals (see that Amazon example). Azeem Azahar, the author of the Exponential View newsletter, who is as good a thinker on the economic and business impact of AI as anyone, argues that the current AI productivity paradox may simply be the expected “productivity J-curve” one would expect with any new, general purpose technology.

Unlike with a technology designed to make a particular process better, which can often have immediate positive productivity impacts, it often takes considerable time for people to figure out how best to deploy a general purpose technology. During this “figuring it out” period, productivity can actually fall rather than increase. This is because companies need to spend time and money experimenting with how to use the new technology, often without seeing a positive bottom line impact. Only later, once people figure out the optimal ways to redesign business processes around the new tech, does productivity experience a sudden acceleration.

The classic example of this that Azhar goes into some depth on is the invention of electricity and its impact on manufacturing. The first thing factories did with electricity was to replace gas lighting with electric lighting. That was a cost savings, but didn’t really change much in terms of the firm’s output. (And there was some cost in installing the lights and wiring the factory, which even muted those savings.) The physics of steam meant that pre-electric factories were built with a central engine that powered many, or even all, of the factory’s equipment off a single drive shaft. So, the second thing factories did was replace the large central steam engine with large electric motors, which they still used to run clusters of machines off central drive shafts. This was cheaper than trying to reconfigure the whole factory. But it turned out to not be very efficient or operationally cost-effective. Productivity gains in one part of the production floor often simply caused bottlenecks elsewhere on the assembly line, and overall the factory saw little gain. It was only when companies began electrifying individual machines and reorganizing the entire layout of factories, that firms saw big productivity boosts.

Very few firms are getting to Stage 3

Azhar predicts that the same thing will happen with AI, but that most firms are sort of stuck in stage one or stage two of this evolution. I think he’s probably right. Tokenmaxxing is easy. Redesigning workflows is hard. Harder still—and something which Azhar doesn’t talk about—is rethinking entire business lines, i.e. what products or services the firm sells, and even business models. This gets at the fundamental purpose of the company. This is where the really big value from AI is. It’s about reinvention, not redesign. But most companies are still not thinking big enough.

Because most existing businesses are being too small minded about how they use AI, AI-native firms have a great opportunity right now. They will be able to move faster and to steal significant market share from incumbents before the legacy companies can effectively respond. It’s much easier to invent a new business from the ground up than it is to try to gut-renovate an existing one. (This is also why it may be more difficult than many private equity firms hope to simply add a dash of AI to their portfolio investments and hope to flip the businesses at higher valuations.)

Ok, with that, here’s more AI news.

Jeremy [email protected]@jeremyakahn

FORTUNE ON AI

Exclusive: Geordie AI raises $30 million Series A to be ‘air traffic control’ for your company’s AI agents—by Jeremy Kahn

Exclusive: Orbital Industries, startup using AI to discover exotic new materials, raises $50 million Series B funding round—by Jeremy Kahn

Boos, AI-washing, and ‘low-value human capital’: The psychological traps CEOs are falling into when they botch their AI messaging—by Claire Zillman

America’s new AI map shows something surprising: ‘A lot of normal people are adopting AI’—by Nick Lichtenberg

AI IN THE NEWS

CNN sues Perplexity for copyright infringement. The news network has sued the AI company, alleging Perplexity’s AI “answer engine” scraped more than 17,000 CNN stories, photos, videos, and other content to provide data for its AI-generated outputs. The suit contends that after negotiations over a licensing deal broke down in 2025, Perplexity continued to appropriate CNN content and falsely implied a commercial relationship with the network that does not exist. CNN is seeking unspecified monetary damages and an injunction blocking further infringement, while Perplexity has pushed back with a terse response from its spokesperson: “You can’t copyright facts.” This is the first time CNN has sued an AI company. Read more from CNN here.Report: Trump appoints former AG Bondi to White House AI panel. President Trump has appointed former Attorney General Pam Bondi to the Presidential Council of Advisors on Science and Technology (PCAST), a White House advisory panel that is influential on AI policy, Axios reports, citing unnamed sources familiar with the decision. The panel is chaired by former AI czar David Sacks as well as current White House science adviser Michael Kratsios, and also includes tech heavyweights such as Nvidia CEO Jensen Huang, Meta CEO Mark Zuckerberg, and Oracle CEO Larry Ellison. Bondi, who was ousted as AG last month, will be tasked with facilitating coordination between the government and the tech executives on the panel, and will also take on a newly created advisory role focused on national infrastructure. The appointment comes as Bondi is recovering from thyroid cancer, which she was diagnosed with shortly after departing the Justice Department, Axios said, again citing unnamed sources.

IBM and Red Hat announce $5 billion project to patch open source code. The initiative, which IBM is calling Project Lightwell, comes as advanced AI models, such as Anthropic’s Mythos, discover more and more critical vulnerabilities in code bases. The project will see IBM and Red Hat deploy 20,000 AI-assisted engineers to create a trusted enterprise clearinghouse designed to identify, test, and patch security vulnerabilities in open-source software which is heavily-used by the majority of large corporations for many critical functions. Enterprises will access the service through commercial subscriptions, receiving validated, production-ready patches they can plug directly into their software supply chains. A cohort of major financial institutions—including Bank of America, Citi, Goldman Sachs, Morgan Stanley, Visa, and Wells Fargo—are already participating as early adopters. You can read more from the Wall Street Journal here.

Snowflake inks $6 billion deal to use AWS chips. The Wall Street Journal reports that data management giant Snowflake has signed a $6 billion, five-year deal to use Amazon Web Services’ Graviton CPU chips, making Snowflake one of AWS’s largest CPU-based computing customers alongside Meta and Apple. The deal reflects a broader surge in demand for CPUs driven by the rise of AI agents, which require large numbers of the processors to orchestrate and sequence their computing tasks. CPU makers including Intel, AMD, and Arm Holdings have all seen rising sales and share prices in recent months as agentic AI has gone mainstream.

Robinhood rolls out agentic AI trading features. Robinhood has unveiled two new products—Agentic Trading and an Agentic Credit Card—that allow customers to connect third-party AI assistants, such as Anthropic’s Claude or the coding agent Cursor, to carry out investing strategies or spending tasks with minimal human involvement. For trading, customers can establish a dedicated agentic account entirely separate from their main portfolio, directing the AI to build a diversified portfolio from scratch or rebalance holdings as opportunities arise. For spending, agents can be given access to a virtual Robinhood Gold credit card to make automatic purchases such as snagging concert tickets or buying products when prices drop below a set threshold. Safety guardrails include isolated accounts with limited funds, spending caps, real-time activity feeds, and a one-tap kill switch—though Robinhood cautions that AI agents can err or behave unexpectedly, and that users bear responsibility for monitoring their accounts. Read more here from CNBC.

EYE ON AI NUMBERS

$9 billion

That’s the amount of money the White House is giving U.S. intelligence agencies to help them establish their own computing clusters of sophisticated Grace Blackwell superchips from Nvidia. The chips are needed so that U.S. intelligence agencies can run their own copies of frontier AI models, such as OpenAI’s GPT-5.5, and possibly Anthropic’s Mythos, as well as future AI models, on their own classified networks. These state-of-the-art models require a large number of specialized AI chips to run or to fine-tune. The Pentagon has recently signed deals with OpenAI, Google, and xAI that allow their AI models to be used in classified networks. The National Security Agency is also believed to be using many of these models as well as those from Anthropic, which the Trump administration has sought to bar from being used by government agencies after the company refused to accede to the Pentagon’s insistence that it allow its models to be used for “any lawful purpose.” The NSA is reportedly still working on some kind of arrangement that will enable it to continue to use Anthropic’s model. Although the full terms of all the contracts are not public, it is believed that in some cases the companies are providing versions of these models to the government that contain fewer guardrails than the version they release to the general public. Read more from the New York Times here.