🪙🤖 The corporate world is hitting a sobering economic ceiling as the high cost of running enterprise AI systems is proving hard to justify. In some instances, it’s even becoming more expensive than simply hiring human labor. Only months ago, executives around the world were eager to replace expensive human staff with digital agents, believing that agentic AI technology offered a cheaper, faster alternative.

Now, running complex multi-agent workflows at scale has resulted in budget blowouts for those very same companies. Execs are realizing that using expensive models for small, everyday tasks is a financial black hole with no clear link to increased value creation, forcing a desperate reassessment of their aggressive AI-first agenda known as “tokenmaxxing.”

Tokens are the basic units of data processed by LLMs. “Tokenmaxxing is the pop culture term for adopting AI at any cost and being willing to double down and spend heavily in order to push the limits of what’s possible,” says Omar Khashaba, managing partner at Algebra Ventures.

It all started with some good ol’ boardroom panic

Fearing they would be left behind, management teams across sectors — particularly in software engineering — scrambled to embed LLMs into daily workflows wherever they could. Many enterprise decision-makers declared that tokenmaxxing was a matter of sheer competitive survival. “It is existential for us. We are in the most competitive space that has ever existed and will ever exist,” May Habib, CEO and co-founder of enterprise AI startup Writer, told The Wall Street Journal.

Everyone has been experimenting for the past year and a half, Kashaba tells us. “Model capabilities obviously grew quite exponentially during that time and boards were sold the narrative that this was going to be a transformational technology, and we couldn’t afford to be left behind. They authorized significant budgets, which in a lot of cases would get spent within the first month or two of the year,” he says.

… Which quickly led to an over-reliance on some iffy KPIs.

To justify these expensive mandates, management introduced performance metrics that sparked a meaningless mad dash for token use. Lacking clear ways to measure actual productivity, companies turned “tokens” into a key performance indicator. At Amazon, internal leaderboards actively ranked almost 80% of developers based on their token usage, according to Business Insider. Engineers quickly learned to game the system by deliberately bloating prompts with massive, irrelevant baseline data to artificially inflate their scores, The New York Times reports. Even firms like Accenture had warned that senior staff could miss out on chances for promotion if they failed to use AI.

The cost breakdown

But the infrastructure costs started adding up pretty rapidly, exposing the friction of scaling these models. Anthropic’s Claude 3.5 Sonnet, for instance, costs just USD 3 per mn input tokens and USD 15 per mn output tokens. Scaled across thousands of multi-step, agentic coding requests, those costs start compounding pretty quickly. Earlier this year, Uber rolled out agentic AI coding tools like Claude Code and Cursor to its engineering team. Within just four months, the company managed to burn through its entire 2026 AI budget. Per-employee token costs were running between USD 500-2k a month — ultimately leading to the decision to place hard caps on employee AI use.

The illusion of high productivity generated by these inflated token metrics has already triggered real-world losses. Watching their corporate dashboards light up with huge token use stats, executives falsely assumed their human workforces had successfully automated their workloads. This misunderstanding earlier this year fueled a wave of aggressive lay-offs, particularly targeting junior and mid-level developers. Firms are now discovering that they traded skilled engineers for expensive “empty compute,” leaving gutted teams struggling to manage increasingly unstable, AI-generated systems, Business Insider reports.

Distorting the wider market

One of the larger consequences of this phenomenon has been the inaccurate signals sent to foundation model providers who watched their API traffic soar. Many of whom interpreted the data as a sign of organic demand, instead of the corporate mandate it actually was.

The physical infrastructure required to feed this manufactured demand might have to be revised. Tech giants like Amazon and Meta had committed to pouring hundreds of bns USD into constructing massive, energy-intensive data centers to support what they believed was an exponential curve of real-world AI demand. Now they might have to scale back their plans.

Where do we go from here?

“There’s been a little bit of sobering up,” says Kashaba regarding the broader market. As far as Egyptian companies are concerned, we kind of have an advantage because we never fully jumped on the tokenmaxxing bandwagon. “I think we've always been pruned by frugality, trained in monetary disasters, so we’re always careful. I think we sort of leap-frogged to where Silicon Valley is now,” adds Khashaba.

It seems like the big productivity push promised by AI has still not transpired. Only 30% of CEOs are confident about revenue growth in 2026 as most struggle to turn AI investment into tangible returns, according to PwC’s 2026 Global CEO Survey.

In the short term, companies are slamming on the brakes. The era of the blank check for enterprise AI is abruptly closing as businesses scramble to implement strict monthly quotas on employee prompts. The corporate directive has reversed overnight from “tokenmaxxing” to “token minimizing,” according to The New York Times. Managers who once demanded high-volume usage are now training employees to write ultra-concise queries to limit what once seemed like unlimited spending.

What can Om El Donia’s entrepreneurs learn from this? Simply put, Egypt has the benefit of seeing the fumblings happen elsewhere first — course-correcting accordingly. An AI optimist, Khashaba tells us that what makes the AI wave exciting is that it’s not “centralized in Silicon Valley,” arguing that the capital and capabilities once needed to build, say, an e-commerce company or a social network are no longer gatekept. “AI allows you to have that same level of capability; you get to do things on a local level,” he says.