AI Bubble Fears, Why Tokenmaxxing Finally Caught Up
The bill for unchecked AI spending just came due
A culture of internal leaderboards and token-burning agents pushed AI adoption to extremes. This week, markets started asking what it actually bought.
If you’re hearing about AI bubble fears for the first time this week, here’s the short version: a culture Silicon Valley insiders call “Tokenmaxxing” just collided with the bill that comes with it. For months, some of the biggest tech companies pushed employees to use AI as aggressively as possible, tracking usage like a competitive sport. Now investors are asking whether any of it actually paid off.
The reckoning showed up fast. On June 23, the Nasdaq dropped 2.2% in a single session, while South Korea’s KOSPI plunged 10% and triggered its first trading halt since March — driven largely by losses at chipmakers SK Hynix and Samsung. This isn’t really about AI failing. It’s about a market that spent the better part of 2026 rewarding spending now demanding evidence of returns.
Here’s what’s actually driving the sell-off, what Tokenmaxxing has to do with it, and what it means for the AI trade going forward.
It’s not whether AI can do the job —
it’s who can actually afford the bill
led by chipmakers and AI names
breaker since March 2026
capex across four tech giants
high-AI-adoption environments
Tokenmaxxing: how the spending got out of hand
Root causeEarlier this year, some Silicon Valley giants reportedly created internal leaderboards ranking employees by how many AI tokens they consumed. The intent was to push AI adoption as far as it would go. The result was a culture where workers threw massive, unchecked projects at agentic AI just to watch their usage numbers climb.
An AI agent doesn’t just answer a question — it runs an ongoing loop, writing code, testing it, hitting errors, and trying again. According to industry analysis, instructing an AI agent to perform a task like tracking a competitor’s announcements can burn nearly 100,000 tokens before it even produces an answer. Multiply that across an entire workforce optimizing for volume, and budgets disappear fast.
The revenue growth AI companies have reported partly reflects this token-burning behavior, not necessarily sustainable demand.
The hangover: code churn instead of productivity
Unexpected costHeavy token consumption hasn’t translated into the productivity gains companies expected. Faros AI reported an 800%-plus increase in code churn — lines of code deleted versus added — in environments with high AI adoption. Instead of saving engineering time, teams are spending their days cleaning up broken, AI-generated code.
Box CEO Aaron Levie has described this pattern as a kind of “AI psychosis” among executives — the belief that more AI usage is automatically better, regardless of whether it produces working output. The mismatch between spending and results is now showing up in earnings calls and investor scrutiny.
If usage metrics don’t translate to real output, the revenue story behind AI valuations gets harder to defend.
Companies are already pulling back
Course correctionThe spending backlash has been swift. Reports indicate that early adopters including Uber and Microsoft have scaled back some autonomous agent licenses after costs became unmanageable. Fortune 500 executives have quietly admitted to burning through token budgets far faster than projected.
This doesn’t mean companies are abandoning AI. It means the unrestricted, leaderboard-driven phase of adoption is giving way to something more deliberate — measuring actual ROI before scaling usage further.
A slower, more selective adoption curve is healthier long-term, even if it looks like a slowdown in the short term.
The market reaction: why chips fell hardest
Market impactBy mid-2026, AI-related companies accounted for roughly 80% of gains in the U.S. stock market, with the five largest tech companies representing about 30% of the S&P 500’s value — the highest concentration in fifty years. That concentration meant any crack in confidence hit hard and spread fast.
On June 23, the sell-off intensified across Asia: South Korea’s KOSPI dropped 10% and triggered a trading halt, with SK Hynix and Samsung Electronics each losing more than 12%. Japan’s Nikkei fell 3.6%, and SoftBank dropped 15%. Combined hyperscaler capital expenditures — over $452 billion across Microsoft, Alphabet, Amazon, and Meta for 2026 alone — became the figure investors started questioning out loud.
When a handful of stocks drive most of the market’s gains, any doubt about their AI bets has an outsized ripple effect.
This looks like a correction,
not a bubble bursting — yet
Talent is moving too, and that’s its own signal
Talent shiftsAdding to the unease, high-profile AI researchers have been switching teams at a notable pace. Google DeepMind’s Nobel Prize-winning researcher John Jumper left for Anthropic, while Gemini co-lead Noam Shazeer departed for OpenAI. These moves raise questions about whether established players can maintain their competitive edge while spending scrutiny intensifies.
Talent migration during a period of financial doubt isn’t unusual — it often reflects researchers betting on where they think the next phase of AI development will actually happen, rather than where the biggest balance sheet currently sits.
Watching where top researchers go next is often a better signal than watching quarterly earnings calls alone.
Some economists have explicitly compared this moment to the dot-com bubble of the early 2000s — a period where massive capital flowed into internet companies regardless of whether they had a path to profit. The comparison isn’t unreasonable: AI-related enterprises drove roughly 80% of U.S. stock market gains by mid-2026, an unusually narrow source of growth for an entire market.
But there’s an important difference. The dot-com bubble was largely funded by venture capital chasing speculative startups with no revenue. This AI buildout is being funded by some of the most profitable companies in history — Microsoft, Alphabet, Amazon, and Meta — using their own cash flow and balance sheets, not just investor enthusiasm. That doesn’t make the spending immune to scrutiny, but it does mean the companies involved can absorb a slowdown more easily than dot-com-era startups could.
What’s more likely than a full collapse is what one analyst described as the end of the phase where “VCs throw a hundred million bucks at any pitch deck with the words large language model.” The technology doesn’t disappear when the financial enthusiasm cools — it shifts from a “look what this can do” phase into a harder, more accountable “prove the ROI” phase. That’s a maturing market, not necessarily a dying one.