In the first few months of 2026, Silicon Valley appeared to have stumbled upon the key to maximizing AI gains, especially within software engineering. Called tokenmaxxing, it was rapidly embraced by the biggest names in technology, from Microsoft and Meta to Salesforce. It relied on a surprisingly simple idea: to get the biggest rewards from AI, companies only needed to give their employees free rein to use as much AI (measured in AI tokens) as they wanted.

The underlying assumption was: the more tokens a company uses, the more work it is getting done by AI, and the greater the productivity boost it can achieve.

But tokenmaxxing's popularity was short-lived. By the middle of the year, it had already given way to a new approach: 'valuemaxxing'. Favoured by companies like IBM, valuemaxixing is a more considered strategy for maximizing AI gains, based on closely monitoring the business value and ROI of individual use cases.

Why did tokenmaxxing take off in the first place? What led to its rapid demise? And why is valuemaxxing now becoming the preferred approach?

Why tokenmaxxing?

If one of the main reasons a company isn't seeing gains from AI is that it simply isn't finding enough ways to apply it, then tokenmaxxing can absolutely make sense. This very likely explains why the likes of Meta, Microsoft and Salesforce were incentivizing and rewarding higher token consumption at a time when they needed to encourage development teams to experiment more with AI.

In fact, many companies were almost glorifying higher AI token consumption. They were creating leaderboards that ranked and rewarded individual developers for the number of tokens they were using.

Employees that didn't embrace valuemaxxing were openly criticized. The CEO of Nvidia was famously quoted as saying that if a $500,000 engineer that worked for him did not consume at least $250,000 worth of tokens, he would be "deeply alarmed".

Employee token allocations were even being promoted within the company compensation package to attract candidates for developer job openings.

The pull-back from tokenmaxxing

While it seemed like a good way to encourage employees to test out and find new ways to integrate AI into their work, tokenmaxxing also ratcheted up costs. Which meant many companies began to pull back from it.

For example, Microsoft was reported to have cut its Claude Code subscriptions to reduce costs. Uber admitted that it had burned through its entire 2026 AI budget in four months, with the COO remarking that "it's hard to draw a connection between the company's rising use of Claude Code and innovations meant to serve consumers".

That, of course, is the root of the problem. Just because employees are using more AI doesn't automatically mean they will realize productivity or innovation gains. Tokenmaxxing only considers the inputs, not the outputs that come out the other side.

That, of course, is the root of the problem. Just because employees are using more AI doesn't automatically mean they will realize productivity or innovation gains. Tokenmaxxing only considers the inputs, not the outputs that come out the other side.

The rise of agentic AI is another factor. Rather than using AI to generate code, developers have evolved to using AI agents to plan and coordinate complex workflows. Agents can analyze code repositories, call tools, test solutions and more and in the process consume even more tokens, as well as more compute power, memory and other resources. So, continuing to encourage unrestricted token usage imposes even higher costs.

No wonder that Gartner is predicting that AI coding costs will overtake the average developer's salary by 2028.

Why valuemaxxing?

Instead of measuring success by how many AI tokens are burned, valuemaxxing advocates for optimizing the measurable business value and ROI from each token or unit of AI consumed.

Under this strategy, the number of tokens staff consume is not the most important question. It's more about measuring the output or value produced per AI token. How many tasks were completed? How much developer time was saved? How many bugs were fixed? How much rework was avoided?

IBM's approach to valuemaxxing

IBM never went in for tokenmaxxing, or requiring people to use as much AI as possible as quickly as possible. Instead, Neel Sundaresan, GM of automation and AI at IBM, suggests the company achieved broad AI adoption with over 95% engagement by giving employees the power to choose whether or not to use AI. The average developer at IBM doesn't use more than $150 per month, although some "power users" might use much more.

Rather than focusing on raw AI token numbers, IBM's approach to valuemaxxing requires that AI usage is always tied to a specific process and a specific outcome.

Neil Dhar, Senior Vice President at IBM, says that every use case should connect to a specific workflow and a measurable outcome. Performance should be tracked every quarter and if a use case is not able to deliver a 2.5x to 3x return, either through time savings, improved customer and employee experiences, or increased revenue, then the organization should pass on it.

AI model rotation

Another part of IBM's approach to valuemaxxing is the recommendation that AI models should be rotated depending on the use case.

In addition to the big frontier AI models, there are now a growing number of more affordable open-source models and task-specific small language models that use fewer AI tokens. Rather than relying on the same powerful large language model for every requirement, companies need to orchestrate between models so that the most cost-effective model is chosen for each task.

Ideally, a company's AI infrastructure should be designed so it can evaluate incoming requests and automatically route them to the model best suited to the job, or combine models to produce the best outcome most cost-effectively.

The IBM Bob AI-powered software development partner, for example, orchestrates across multiple models, automatically optimizing for cost, quality, and performance. IBM Consulting Advantage, an AI-powered delivery platform, also includes a model router that directs each task to the appropriate model.

In fact, IDC predicts that by 2028, 70% of leading AI-driven enterprises will use architectures to manage model routing across multiple models, orchestrating complex processes.

Enabling accountability

Another important aspect of valuemaxxing in software engineering, according to IBM, is accepting that both developers and technology leaders need to be held accountable for how AI is used.

Developers must consider what they are using AI for. Is it a legitimate use case? What would be the most efficient way to apply it, including providing the right amount of context to ensure minimal rework?

Leaders need to ensure the right tools are available to track AI usage, costs and outcomes, and to incentivize and reward efficiency rather than raw token usage.

For example, tools like IBM Bob provide the transparency, administrative controls and usage tracking with analytics needed to help leaders connect AI consumption to business outcomes.

Similarly, IBM Apptio AI Value & ROI, provides a single view of AI spending, including token consumption costs and its connection to measurable business results.

Tokenmaxxing was the right strategy for the experimentation phase

It would be easy to be critical of tokenmaxxing. Why would promoting unrestricted consumption of a costly resource ever make sense? But that probably does it a disservice. Tokenmaxxing was probably the right approach for a very brief early phase in an enterprise's AI lifecycle: the phase when they were simply not experimenting with AI enough or working hard enough to actively find ways to incorporate it into their processes and workflows. Once that phase is passed, IT and business leaders quickly realized that unrestricted AI consumption was no longer fit for purpose. Valuemaxxing is replacing it as the smarter, more sustainable approach.

This article originally featured on the IBM Community.