The first batch of "burn-resistant" token giants has appeared.

Author: Little Bear Cookies, BitpushNews

Over the past two years, global tech giants have been throwing money around without blinking an eye to grab a “first-class ticket” in the AI era. But this nearly blind frenzy is being dragged back to reality by a series of cold, hard bills.

The first to publicly tear off this shameful cover was a core veteran giant in Silicon Valley: Uber.

This past weekend, Uber’s Chief Operating Officer (COO) Andrew Macdonald poured out his frustrations in an interview.

The core message is simple: the company has been spending wildly on AI tools for engineers, only to find that this money not only failed to turn into new features perceived by users, but also exhausted the entire annual budget.

Macdonald shared a set of impressive comparisons in the interview:

The penetration rate of engineers using AI programming tool Claude Code has skyrocketed from about one-third in February to 84% in March, with the company's monthly bill per engineer usually between $150 and $250, and heavy users reaching $500 to $2,000.

What’s even more exaggerated is that he himself, for a technical demo, ran for two hours, and $1,200 worth of tokens evaporated.

Macdonald pointed out: “Engineers who don’t responsible for paying invoices think of AI tools as free and unlimited, like tap water.”

What’s more crushing for management is: despite burning so many tokens, from the user’s perspective, the Uber app hasn’t become more user-friendly.

“Greater token consumption hasn’t translated into more useful features for users,” Macdonald said frankly, “It’s hard to draw a direct line between these two data points and say ‘Look, we now provide 25% more features to users.’”

In 2025, Uber spent $3.4 billion on R&D, a 9% increase year-over-year. CEO Dara Khosrowshahi has explicitly stated that the company is slowing down hiring to offset the ever-expanding costs of AI investments. In other words, the saved labor costs are being redirected to AI — but the output remains unmeasurable.

This phenomenon is not unique to Uber. Microsoft has already begun canceling most Claude Code licenses, requiring engineers to switch back to their own Copilot CLI.

Macdonald highlighted a key issue: “If you can’t establish a direct link between token consumption and user features, then the trade-off between AI spending and labor costs becomes increasingly hard to justify.”

From Frenzy to Pragmatism: Internal “AI Retreat” in Big Tech

Managerial anxiety is being transmitted to frontline employees through performance evaluations. But different companies’ responses are visibly diverging.

The most notable shift comes from Duolingo.

Last year, this edtech company announced an “AI-first” strategy, incorporating AI usage frequency into employee performance evaluations.

However, just a year later, in April 2026, during the “Silicon Valley Girl” podcast, the CEO expressed regret and announced the full withdrawal of this policy, stating that AI usage would no longer be considered in performance reviews. The new company stance is: “Doing your job well is the priority; use AI if it helps, but don’t force it if it doesn’t.”

The famous “Squid Factory” Meta is even more dramatic. At the end of March, the company launched a leaderboard called “Claudeonomics,” tracking token usage of over 85k employees, awarding titles like “Token Legend” to the top 250. According to Reuters, employees consumed about 60 trillion tokens in 30 days, with the highest individual usage reaching 85k tokens.

But this trend quickly spiraled out of control: the top employee’s monthly token consumption hit nearly $500k, far exceeding company expectations. Soon after, Meta quietly shut down the leaderboard — the official reason was data leakage, but employees widely suspected it was because “the costs of abnormal competition had become unsustainable.”

Behind this collective backlash is a workplace resistance to the “pseudo-efficiency” of AI.

According to a recent global survey jointly released by SAP and WalkMe, over one-third of white-collar workers actively and consciously “skip” AI tasks assigned by their companies in daily work. The reason is simple: the hallucinations and fragmented operations of large models frequently disrupt the original coherent workflow. In many fine-grained scenarios, spending half a day tinkering with AI only to manually correct results actually makes the process slower than pure manual work.

In this “AI KPI suppression war,” only a few small-to-medium e-commerce automation platforms like Omnisend still insist on “pushing hard,” trying to boost morale by giving “AI-skilled workers” a 2% to 4% salary increase. But even they have shifted their evaluation metrics from initially vague “usage time” to three cold financial indicators: how much time and money was saved, and how many colleagues reused the AI workflows you trained.

Meanwhile, regulators are also starting to act against the negative social effects of AI.

Five days ago, California — the global hub of AI technology — signed an unprecedented executive order calling on state agencies, labor experts, and universities to comprehensively address the potential drastic impacts of AI on the job market.

This is the third major AI regulatory measure in California over the past five months: earlier, the SB 53 transparency law took effect in January, requiring leading AI model developers to publish risk management frameworks and report major safety incidents within 15 days; in March, Executive Order N-5-26 set AI safety thresholds for state procurement.

The background is that the speed of replacing entry-level jobs is visibly accelerating. Data from the employment monitoring platform Layoffs.fyi shows that in the first 18 weeks of 2026, employers disclosed over 113k layoffs involving tech workers. According to the latest report from Challenger, Gray & Christmas on May 7, in April 2026, U.S. employers announced 83,387 layoffs, up 38% from March. Although this is 21% lower than the same period last year, it remains the third-highest April since 2009 — only behind April 2025 and April 2020.

When signing the order, Governor Newsom told reporters: “We don’t want to wait until thousands are laid off before we start thinking about solutions.”

New highs in the US stock market, underpinned by a “new structural divergence”

Meanwhile, as top executives worry about token costs and California’s legislation slows down, the US stock market continues to hit new highs driven by strong sectors.

On May 26, the S&P 500 closed up 0.61% at 7,519.61 points, and the Nasdaq surged 1.19% to 26,643.45 points. The semiconductor sector led the rally: Micron Technology soared over 19%, with its market cap surpassing $1 trillion for the first time; the Philadelphia Semiconductor Index jumped 4.6%, hitting new highs again. JPMorgan analysts even raised the target for the S&P 500 to 9,000 points, with the core logic that AI capital expenditure will directly boost U.S. GDP growth.

The logic of the capital markets is extremely pragmatic: since mid-tier giants like Uber are frantically burning money to buy tokens, this expensive technical service fee will ultimately turn into revenue and profit for infrastructure providers like Nvidia, Microsoft, Amazon, and top model providers such as Anthropic and OpenAI. The upstream companies are raking in huge profits, while the mid- and downstream application layers are struggling with ROI.

The question Macdonald raised at the beginning still has no convincing answer. Today’s market continues to rise, but beneath the rally, the cracks are widening.