In 2026, the once booming artificial intelligence (AI) industry is facing a heavy financial blow. Various AI model developers have quietly ended significant subsidies for “Token” (referred to as “Terms” in China) calls, triggering a chain reaction of cost spikes. This reaction is spreading from Silicon Valley tech giants to the trading floors of Wall Street, forcing businesses and investors to confront the significant gap between high computational costs and actual investment returns.
According to reports, from Microsoft urgently halting internal incentive projects to Uber’s billion-dollar budget vanishing within months, the commercialization of AI is undergoing a severe “stress test”. The era of rampant Token consumption is coming to an end.
The direct trigger for this cost crisis is the sharp increase in the price of the basic unit for measuring AI input and output, the “Token”.
Between February and June this year, OpenAI, Anthropic, and GitHub each adjusted their pricing models, charging customers based on Token usage rather than using fixed rates.
A report by JingTong Finance on June 10 stated that over the past six months, the token pricing for high-quality inference services on cutting-edge models has increased by about 40%. This is due to the continued restriction on high-performance GPUs, a 15% to 20% rise in data center energy costs, and the explosive growth in demand.
For example, OpenAI’s recent release of GPT-5.5 has doubled the Token price to $5 per million input Tokens and $30 per output Token; Google’s new Gemini Flash 3.5 model is priced at 3 to 6 times that of its predecessor.
Although model providers have achieved about a twofold efficiency improvement in a year, the premium on Tokens during the same period has reached 40% to 50%. This has led to a significant increase of 20% to 30% in net costs for application-oriented enterprises relying on external APIs.
The unexpected cost peak has breached the internal budget defenses of tech giants. According to tech media The Verge, Microsoft made a rare decision in late May this year: to terminate the collective license for Claude Code by its internal “Experiences and Devices” department on June 30.
The pilot project, launched with great fanfare in December 2025, quickly collapsed due to the unexpectedly high Token consumption, forcing Microsoft to order engineers back to using the more controllable GitHub Copilot CLI.
Uber’s situation is a financial disaster. Its Chief Technology Officer, Praveen Neppalli Naga, recently admitted that the company’s $3.4 billion annual budget for AI was exhausted as early as April 2026.
After rolling out Claude Code to 5,000 engineers in the company, the monthly active usage soared to 85% to 95%, with monthly API call costs per engineer ranging from $500 to $2000.
Wall Street has quickly sounded the alarm. Ohsung Kwon, Chief Stock Strategist at Fidelity Bank, pointed out that the core of this storm is putting an end to the “Tokenmaxxing” trend among Silicon Valley engineers.
Many companies previously incorporated AI tool usage into their internal evaluation systems, even creating leaderboards to encourage employees to consume as many AI Tokens as possible as a measure of innovation capability. However, this strategy of blindly pursuing usage has evolved into a severe waste of resources.
Kwon warned that if AI demand starts to stabilize, it would be a major bearish sign for AI transactions. Based on this assessment, Fidelity Bank shifted its overall position from “bullish” in April to a “firm neutral” stance.
Bryan Catanzaro, Vice President of Applied Deep Learning at Nvidia, admitted in an interview to widespread industry anxiety: “In the team I lead, computational costs have far exceeded personnel costs.”
If at the beginning of winter in 2026 companies were feasting on “AI all you can eat,” then summer marks the start of calorie counting.
According to Business Insider, between February and June this year, OpenAI, Anthropic, and GitHub each adjusted their pricing models to charge customers based on Token usage rather than fixed rates.
“The era of cheap ‘AI all you can eat’ is over,” said a senior software engineer at Deloitte, commenting on the shift in GitHub’s pricing model causing chaos in work expectations. He estimated that under a pay-as-you-go system, detailed prompts that may require models to work for a few hours could cost over $100 per occurrence.
Mario Rodriguez, Chief Product Officer at GitHub, explained that under the old model, the cost for a casual chat question was the same as extended autonomous code work, and this subsidy was “no longer sustainable.”
Faced with sudden changes in billing rules, the business world quickly changed direction. Walmart set usage caps on internal programming tools, and Amazon closed its internal “Tokenmaxxing” leaderboard in May because they found employees were artificially inflating computational costs for unnecessary operations to increase their score.
Some companies were even forced to implement strict quota systems. A senior executive at cryptocurrency exchange Coinbase indicated that since the launch of Claude Opus 4.6 in February, internal usage had skyrocketed. Consequently, they established a complex weekly cost ceiling system, setting upper limits ranging from $500 to $5000 based on employee rank.
He gave an extreme example: if using the most advanced models to scan all company code for vulnerabilities could cost $50,000 to $100,000 each time, “if a hundred people each independently did this, you’ll spend $10 million.”
Chris Reed, Senior Director of Finance at Priceline IT, described the industry’s growing concern as an “addiction epidemic to crack cocaine”. He bluntly stated, “They let you try it, get you addicted, and then you’re hooked.”
Why are companies not seeing the expected returns despite substantial investment in AI? An in-depth investigation by Economic Information Daily revealed the other side of companies rushing to embrace AI.
An employee at a tech giant named Wang Hao (alias) complained, “A department of twenty people, spending $50,000 in Tokens per month, without achieving anything.” This $50,000 turns into intangible costs because each team member selected different tools (such as open-source Hermes Agent or third-party tools), creating disconnected “AI islands” that ultimately lead to starting over.
Vitaly Gordon, CEO of engineering operations platform Faros AI, shared an extreme case: a CTO discovered that an engineer burned through $40,000 in Tokens in a month, but was unsure whether to stop or encourage this behavior.
Research data from engineering management platform Jellyfish further quantified this contradiction: driven by agent functions, the Token consumption for each developer increased about 18.6 times in nine months; the productivity of engineers using the most Tokens was approximately twice that of low users, but the amount of Tokens they consumed was ten times higher.
A report released by consulting firm Bain in June revealed a harsh commercial reality: among enterprises capable of quantifying the cost-saving effects of AI, up to 40% achieved actual cost reductions of 10% or less. Initially, 37% of companies set their cost reduction goals in the 11% to 20% range, but only 31% achieved this target.
Analyzing the situation, Zhang Yi, CEO of iMedia Consulting, stated that many companies fell into the trap of making high-stakes bets based on “fear of missing out” (FOMO), only calculating explicit API call fees but completely ignoring significant hidden engineering costs like prompt engineering, output verification, and data governance.
As the “era of magical thinking” comes to an end and pragmatic utilitarianism begins to dominate the summer of 2026, business executives are starting to view Token waste as financially irresponsible.
A reassessment of AI costs is quietly unfolding within the corporate world.
Economic Information Daily learned that Tencent has recently adjusted the Token allocation mechanism for employees, moving away from a shared pool approach to dynamic allocation by department managers based on job functions. Tencent internally stated that the measure of AI effectiveness should be based on efficiency and value, not merely Token consumption.
Parker Harris, CTO of Salesforce, noted that due to Token expenditures far exceeding plans in the 2026 fiscal year, the company is introducing a metric called “Effective Output score” to predict returns and control expenditures.
Meanwhile, the search for cheaper “substitute models” has become a new trend. Companies like Coinbase have started shifting basic work to lightweight models from China. A code agent startup, Command Code, revealed a surge in demand for affordable models, attracting 10,000 new customers within 30 days.
Trevor Stuart, Senior VP at software startup Harness, aptly compared this shift: “Using cutting-edge AI models for basic text summarization work is like using a Ferrari to go grocery shopping.”
To establish standards at a macro level, a new market and regulatory organization has emerged. The Linux Foundation announced the formation of the “Tokenomics” Foundation in July this year, supported by giants like IBM, Oracle, and JPMorgan.
J.R. Storment, Executive Director of the FinOps Foundation, pointed out that tracking cloud costs is a monthly data set of tens of millions of rows, while tracking Token costs is a “problem of trillions of rows of data per month.” This foundation aims to establish new metrics like “cost per unit of intelligence” and “Token per watt” to incorporate AI expenses into financial discipline similar to cloud computing.
Facing the cost crisis, the industry’s focus naturally turns to next-generation hardware. However, despite Nvidia’s acquisition of chip startup Groq and AMD, Intel, and others redesigning AI accelerators to lower per-Token costs, most hardware releases are not expected until the second half of this year. Mass deployment to alleviate supply-demand imbalances is not expected until early to mid-2027. A distant remedy to an urgent situation.
Even if hardware costs eventually decline, the exponential growth of AI agents could offset these benefits. Jensen Huang, CEO of Nvidia, once envisioned a grand scenario of “a hundred AI agents working alongside each employee”. Goldman Sachs predicts that by 2030, global Token usage will surge 24 times to reach 120 quadrillion per month.
However, Gartner’s Director Analyst Will Sommer cautioned that although the inference cost of large language models will be nearly 90% cheaper by 2030 than in 2025, “Chief Product Officers should not conflate consumer Token tightening with the democratization of cutting-edge inference.” Because the amount of Tokens required for agent-based models to complete tasks is significantly higher than standard models, the rate of increase in consumption could potentially outpace the decrease in unit costs.
Is the Token cost crisis of 2026 a sign of the AI bubble bursting, or a necessary growing pain for its commercial model maturity?
An executive at an AI company stated that when the speed of cost escalation overwhelms all efficiency improvements, the industry is no longer about whether AI can change the world but about how companies find balance between technical enthusiasm and commercial reality.
This “pressure test of authenticity” will determine who can truly survive and profit in the future AI wave.