A year ago, Silicon Valley was still buzzing about a term: Tokenmaxxing.

Employees competed over who consumed more tokens, with internal leaderboards at some companies where high usage was assumed to mean high efficiency. At GTC, NVIDIA CEO Jensen Huang called tokens "the fundamental currency of the AI era," implying that the more you use, the more you produce—as natural as can be.

But by November 2026, the winds had shifted faster than many expected.

Apple was reportedly said to have quietly lowered the daily AI spending cap for some engineers from $300 to $200, with some teams potentially dropping further to $100, and non-engineering teams facing caps as low as $25. Meanwhile, Apple's fiscal 2026 R&D spending rose 32% to $45.8 billion, and operating expenses jumped 22% to $75.5 billion.

Cutting employee token quotas on one hand while ramping up AI R&D on the other—Apple's "split personality" is precisely the most worthwhile sample to dissect in this AI cost shift.

What Apple Is Cutting, and What It's Adding

First, let's clarify a fact that is easily misread: what Apple is cutting is not strategic AI investment, but the "retail cost" of employees' daily calls to external models.

Apple employees can use Anthropic's Claude, OpenAI's ChatGPT, and Google's Gemini. Some teams have already been asked to switch to cheaper models or to use AI running locally on devices. In other words, what Apple is doing is turning "using AI" from "swiping the company card freely" into "every penny spent must justify the output."

On genuine strategic investment, Apple has not held back. Fiscal 2026 R&D spending rose to $45.8 billion, up 32%, and the CFO explicitly stated that the company will "continue to significantly increase overall AI investment." But Apple's capital expenditure actually fell 29% to only about $9 billion for the year—in stark contrast to Amazon and Alphabet, whose latest quarterly capex surged 122% and 114% respectively.

Behind this is a clear financial logic: Apple is unwilling to bear the heavy-asset risk of AI infrastructure itself. It outsources the heavy lifting to Google Cloud, uses NVIDIA chips to handle complex requests, and focuses on end-user experience and R&D efficiency. eMarketer analyst Jacob Bourne's assessment hit the nail on the head: this "lukewarm AI infrastructure spending" helps Apple preserve cash flow while maintaining AI capabilities through strategic partnerships without taking on the financial risks its peers do.

So Apple's logic is not "not doing AI," but "not paying a premium for the intermediate links of AI." The tokens employees burn daily are, in Apple's eyes, shifting from "efficiency investment" to "operating costs" that require meticulous management.

This Isn't Just Apple's Story

Apple's adjustment is worth magnifying because it lands on a node of industry-wide resonance.

Uber exhausted its entire annual AI budget in the first four months of 2026, with its COO publicly admitting it is "increasingly difficult to justify internal AI costs." Tesla set a $200 weekly cap on employee AI spending; Uber limited monthly quotas for individual AI coding tools to $1,500; Meta introduced a token budget system and encourages employees to switch to internally developed tools; and Microsoft is also tightening internal AI spending quotas. Accenture even compressed centrally borne Claude quotas to $29 per employee per month.

A UBS survey of more than a dozen enterprise IT executives offers a more structural judgment: about 60% of enterprises are setting "guardrails" in some form to curb AI spending. UBS analysts used restrained language, calling it a "mild headwind" rather than an "alarm"—no one is stopping AI deployment, but the carefree Tokenmaxxing era is indeed over.

Chinese companies have pivoted almost in sync. Tencent's non-R&D employees had monthly token quotas exceeding 1,500 yuan at the start of the year, but by early June most had been cut by more than 60%; Xiaomi shifted from encouraging unlimited use of its self-developed models to setting individual caps. Tencent, ByteDance, Alibaba, JD.com and other leading internet companies have successively introduced new rules tightening employee AI token permissions, with the only difference being the approach: some differentiate by role, others apply two sets of standards for self-developed versus external models.

This is not one company's cost anxiety, but an industry switching from "experimental mode" to "operational mode."

Why Now, of All Times

A natural question: aren't AI inference unit prices falling? Why is the total bill spiraling out of control?

The answer lies in how usage patterns have changed. Gartner predicts that by 2030, the cost of inference on trillion-parameter large models will drop by more than 90% compared to 2025. But at the same time, enterprises are turning AI from a "Q&A tool" into "agentic execution"—letting intelligent agents autonomously write code, debug, and reason continuously for hours or even tens of hours.

Cognitive scientist Gary Marcus offered an intuitive order of magnitude: agentic tasks issue a large number of internal queries on the way to an answer, with token consumption "sometimes 500 times, even 1,000 times that of ordinary interactions." Writing an app can consume hundreds of millions or even billions of tokens, while a typical conversation uses only a few hundred to a few thousand.

This leads to the most counterintuitive tension in the token economy: unit prices are plummeting, but total spending is expanding exponentially. Over the past three years, token unit prices have fallen by more than 90%, yet China's daily average token calls surged from 100 billion in early 2024 to 140 trillion in March 2026—a thousandfold increase. In economics, this is called the "Jevons paradox"—improved resource efficiency actually stimulates expansion of total consumption.

When tokens go from "a nearly negligible variable cost" to hard spending where "a single R&D engineer consumes at least 20,000 yuan a month, and leading companies rack up monthly bills in the hundreds of millions," no CFO can continue to leave it in the footnotes of the budget sheet. Uber burning through its entire annual budget in four months is an extreme footnote to this structural mismatch.

What Is Really Being Repriced

If you understand this only as "big companies saving money," you'll miss the more important signals.

First, differentiation among model providers will accelerate. UBS explicitly noted that OpenAI and Anthropic are most vulnerable in the short term to enterprise cost-cutting, while open-source models like DeepSeek and Chinese models are potential beneficiaries, especially for non-coding tasks. When enterprises start "choosing models by task" rather than "choosing models by brand," the pricing power and customer stickiness of the model layer will be reassessed. Apple is already asking some teams to switch to cheaper models—itself a price signal from the demand side.

Second, the logic of AI procurement is shifting from "capability first" to "output per unit first." Over the past two years, enterprises worried about "will we fall behind if we don't use AI"; now the question has become "how much measurable output does each dollar of tokens buy back." A quote from a UBS interview captures it accurately: "The question is not whether to use tokens, but how to use them efficiently. Optimization is becoming a continuous engineering discipline, not a reaction to a budget crisis." This means AI management is shifting from a one-time "adopt the tool" exercise to an ongoing "cost discipline."

Third, the infrastructure investment narrative will come under pressure, but won't be overturned. Apple's strategy—more R&D, restrained capex, outsourcing the heavy lifting—is an extreme case, but the question it raises is universal: when inference demand explodes while per-token prices keep falling, how exactly should the payback period of heavy-asset investment be calculated? Amazon and Alphabet's capex is still doubling, showing that the long-term logic of compute scarcity has not been falsified; but the tightening of enterprise token budgets will at least subject the pace of "building data centers at all costs" to stricter scrutiny.

The End of the Money-Burning Era Doesn't Mean the AI Era Is Slowing Down

The news of Apple cutting employees' daily AI quotas is easily read as yet another piece of evidence for an "AI retreat." But a more accurate reading may be this: AI is moving from a "faith narrative" into an "accounting narrative."

When a technology is still in the "use it or not" stage, encouraging more use and tolerating waste is a reasonable strategy. But when it's embedded in core workflows, costs run into hundreds of millions, and every expenditure can be tracked and attributed, "can you do the math" becomes a more important capability than "dare you burn."

Apple's approach may not suit every company. It has $136.7 billion in free cash flow, a moat in its hardware ecosystem, and can outsource heavy assets while staying in the game. But for the vast majority of enterprises, Apple's "split personality" conveys a more universal signal: AI competition is shifting from "who adopts the strongest model first" to "who can convert tokens into the most real value under given cost constraints."

The Tokenmaxxing era is over. But the ledger of the token economy has only just begun to be seriously opened.