In March 2025, NVIDIA CEO Jensen Huang proposed a bold vision at the GTC conference: companies should allocate a Token budget for each engineer, equivalent to half of their salary.
In a subsequent podcast, he explained this idea more bluntly: Suppose an engineer with an annual salary of $500,000 tells me at the end of the year that they haven't even burned through $250,000 worth of Tokens, I would be deeply shocked; what if they only spent $5,000? I might be so furious I'd be jumping up and down right there.
Huang's logic is straightforward and enticing. Tokens are the lines of code in the new era: the more you burn, the more you innovate, and the more you produce. This formula was quickly embraced by Silicon Valley and domestic internet tech giants, evolving within a year into an internal management competition: Token consumption was incorporated into KPIs, probation evaluations, and promotion defenses, with some companies even directly introducing internal "Token consumption leaderboards."
However, this competition has suddenly hit the brakes across the board.
From Leaderboard Frenzy to Quota Contraction: Tech Giants Collectively Slash Employee Tokens
On April 15, 2026, Uber's CTO publicly disclosed that the company's full-year 2026 AI Token budget was entirely exhausted in just four months, urgently cutting the Token quota for the R&D department across the company by 70% and simultaneously launching a tiered approval mechanism for quotas.
On May 28, 2026, Amazon took the lead in taking down its internal "Employee Token Consumption Leaderboard," with executives issuing a public notice: deliberately consuming Tokens to climb the rankings is prohibited, and internal AI models may only be invoked for scenarios with genuine business needs.
On June 9, 2026, Meta followed suit, sending an internal memo to approximately 6,000 employees, announcing a cap on Token usage and building a real-time tracking platform to curb cost inflation.
In May and June, Microsoft massively revoked the Claude Code authorizations of thousands of developers, requiring teams to migrate back to its in-house GitHub Copilot CLI by June 30; at the same time, it set daily/weekly Token quota limits for employees, built an internal AI usage monitoring platform, and required business approval for exceeding quotas, eliminating unregulated calls that drive up computing costs.
Even more dramatically, Tesla, starting July 6, 2026, capped the weekly external AI tool usage fee for each employee at $200.
In addition to the aforementioned overseas tech giants, domestic internet tech giants have also successively rolled out internal Token control measures:
On April 8, 2026, ByteDance issued an internal memo, setting an annual reimbursement cap for external AI Tokens for R&D and operations employees, relying on the TRAE platform's usage alerts to manage excess costs.
On May 20, 2026, Alibaba issued an internal notice, setting monthly Token limits for third-party models for product and R&D employees, and tracking the group's computing power consumption in real time through the Alibaba Token Hub.
On June 3, 2026, Baidu introduced control rules to limit the annual quota for third-party large models for technical positions, relying on the Qianfan backend to monitor Token usage in real time.
On June 5, 2026, Tencent released a company-wide memo, canceling the unified high Token quota, implementing role-based differentiated allocation, and building departmental monitoring dashboards to curb ineffective consumption.
These actions occurred intensively between April and July 2026, with highly overlapping timelines and highly consistent approaches: setting caps, slashing quotas, shutting down leaderboards, and prohibiting volume padding. Global tech giants are collectively hitting the emergency brakes on the narrative that "burning Tokens equals innovation."
Tech giants are not short of money. In 2026, the combined AI capital expenditures of Microsoft, Amazon, Google, and Meta approached $700 billion. But spending money to build infrastructure is one thing; letting employees let Tokens flow like tap water is quite another.
When the "internal bills" for AI have become so large that even giants like Meta are crying out in pain, things have changed.
Tech giants are paying the bill for the narrative of "burning Tokens for innovation," and the numbers on the bill are larger than anyone anticipated.
How to Understand This Reversal?
To understand this reversal, we must first review what happened a year ago.
Starting in the second half of 2025, "Token consumption" rapidly evolved from a backend technical metric into a management tool at domestic and overseas tech giants.
A Xiaohongshu blogger named "Slow Hand" shared: "I was at the top of the Token consumption leaderboard in March and was specifically praised by my boss. But what I didn't dare to say is that over half of my Tokens were spent on organizing personal data and posting notes."
According to his description, a certain internet tech giant he works for recently implemented a new management measure. The department head published a March Token consumption leaderboard in the group chat and explicitly stated that future employee probation confirmations, annual KPI assessments, and promotion evaluations would all refer to this data, with employees with lower consumption potentially facing position adjustments.
In Silicon Valley, engineers even coined a term for this behavior: tokenmaxxing (Token maximization). One OpenAI engineer processed 210 billion Tokens in a single week, equivalent to the total text volume of 33 Wikipedias. Another engineer using Claude Code incurred a monthly Token cost of up to $150,000, far exceeding their own salary.
The most extreme example comes from Meta. According to internal data obtained by The Information, Meta once launched an internal leaderboard called "Claudeonomics," ranking the top 250 employees by Token usage and incorporating "AI-driven work outcomes" into core performance requirements. Under such rules, employees were forced to farm Tokens to climb the leaderboard. After the leaderboard ran for 30 days, the company's total Token consumption skyrocketed from 6 trillion to 73.7 trillion, an increase of over 12 times. The leaderboard was subsequently taken down urgently.
Meta's CTO had already sounded the alarm in an April internal memo: "No one should use AI just for the sake of using AI. Token usage itself is not a measure of impact in any sense."
Domestic Low-Cost Models: Has the Replacement Window Really Arrived?
On May 22, 2026, DeepSeek announced that the API price for its flagship model V4-Pro would be permanently reduced to a quarter of its original price, a 75% drop. Specifically: input (cache hit) is only 0.025 RMB per million Tokens, input (cache miss) is 3 RMB per million Tokens, and output is 6 RMB per million Tokens.
Evaluations by third-party institutions show that, under the same task complexity, the average execution cost of DeepSeek V4-Pro is about one-tenth that of GPT-5.5 and one-eleventh that of Claude Opus 4.8.
If we draw a price comparison chart, the gap is suffocating:
Data source: Official API pricing pages of various models, compiled from public materials
The price gap between domestic models and international flagships has widened to 25-35 times. Against the broader backdrop of "converging capabilities," this price cliff will inevitably trigger substitution logic.
However, not within the tech giants themselves. The current response strategies of internet tech giants generally exhibit the characteristic of "in-house development first." The primary beneficiaries of cutting external Token budgets are not cheaper external alternatives, but internal in-house solutions.
Tech giants have in-house alternatives; small and medium-sized enterprises do not. When AI Tokens become a basic productivity tool, a 25-fold price difference is a highly decisive variable. For a startup replacing Claude Sonnet with DeepSeek V4 Pro for a customer service Agent, the perceived capability gap is minimal, while the perceived cost gap is massive.
Figure: Market share of mainstream large models on the OpenRouter platform, with DeepSeek in dark green; Data source: OpenRouter
According to data from OpenRouter, the world's largest neutral AI model API aggregation platform, on the OpenRouter platform, the share of text requests for DeepSeek models rapidly increased from 14.3% in the week of May 25, 2026, to 20.1% in the week of July 6.
What Remains After the Tide Recedes?
The collective tightening of Token budgets by tech giants has punctured the false prosperity of "burning money for innovation," but this is not the end of the AI industry; rather, it is an inevitable transformation from "blind frenzy" to "lean production."
The vanity leaderboards and ineffective "volume padding" that were previously stuffed into KPIs have been thoroughly purged, replaced by stringent ROI (Return on Investment) audits. AI computing power is comprehensively shifting towards essential businesses that can truly reduce costs and increase efficiency, in-house solutions, and cost-effective alternative models. What the quota storm has taken away is the computing power bubble and management anxiety; what it leaves behind is a healthier and more rational industrial ecosystem.
This receding tide is not an abandonment of AI by Silicon Valley and internet tech giants, but rather the second half of the game where hard tech innovation steps into refined operations and pragmatic implementation.