The artificial intelligence (AI) sector, after years of explosive growth, is now confronting a harsh financial reality in 2026. As major AI labs quietly end substantial subsidies for token usage costs, a chain reaction triggered by soaring expenses is spreading from Silicon Valley's codebases to Wall Street trading floors. From Microsoft (MSFT) abruptly halting internal AI projects to Uber (UBER) seeing hundreds of millions in budget evaporate within a quarter, and Wells Fargo strategists issuing stark warnings of an end to the "exuberant rally," the AI commercialization journey is undergoing a severe stress test.
Wall Street Sounds the Alarm on the Most Pressing Bearish Argument
Wells Fargo's chief equity strategist, Ohsung Kwon, stated in an interview this Tuesday that the AI-fueled market rally faces an "imminent" threat, with the core issue being the end of the "Tokenmaxxing" era. Kwon described the recent sell-off in tech stocks as a "wake-up call for investors," emphasizing that "nothing goes up every single day forever." He noted that the initial sell-off stemmed from position adjustments rather than deteriorating fundamentals, but the most compelling bearish argument now is that AI labs are no longer subsidizing costs, leading to a sharp spike in token prices. This has already prompted major corporations like Walmart and Uber to sound alarms, reflecting that their AI budgets are on the brink of depletion in just a few months.
Kwon warned, "This is essentially a direct read on AI demand. If demand starts to at least flatten out, that would be a significant negative for the AI trade." Based on this assessment, Wells Fargo has decisively shifted its overall stance from bullish in April to "firmly neutral." The firm is not advising clients to liquidate all AI assets but rather to immediately establish hedges through strategies like buying put options or selling call options. For defensive positioning, Kwon specifically mentioned the notably lagging healthcare sector, suggesting it could benefit if AI-driven volatility persists.
Exploding Token Bills: From Code Assistants to Sky-High Consumption
The immediate trigger for this cost crisis is the sharp increase in the price of tokens, the unit of measurement for AI API calls. Statistics show that over the past six months, pricing for high-quality inference services using cutting-edge models has risen by approximately 40% cumulatively. This is the result of a combination of factors: persistent constraints on high-performance GPUs, a 15%-20% increase in data center energy costs, and explosive growth in demand.
Although model providers have achieved about a 2x efficiency improvement through optimization techniques within a year, token premiums have surged 40%-50% over the same period. This has led to a net cost increase of 20%-30% for application-layer companies reliant on external APIs. Further reports have pointed to potential price collusion and "addictive" testing in the industry.
Recently, OpenAI doubled token prices with the release of GPT-5.5, charging $5 per million input tokens and $30 per million output tokens. Alphabet (GOOGL) followed suit, with its new Gemini Flash 3.5 model priced 3 to 6 times higher than its predecessor. More critically, intelligent agent tools capable of performing multi-step complex tasks consume tokens tens of times faster than standard chatbots. This level of consumption can instantly overwhelm corporate IT budgets.
Analysis suggests AI companies are attempting to justify high prices by anchoring to the logic of "replacing human labor costs": "Having AI do the work costs about $30 per hour, while hiring an employee costs $40 per hour plus benefits. Under the narrative that AI is cheaper than humans, future AI pricing may not be based on tokens but directly marketed as 'the cost to replace a full-time employee.'"
Compounding the problem is that major model developers are currently operating at significant losses, leaving no profit margin for competitive price cuts.
Tech Giants' Budget Meltdown: Microsoft's Retreat and Uber's $34 Billion Lesson
The unexpected surge in costs first breached the internal budget defenses of tech giants. According to reports, Microsoft made a rare decision earlier: after just six months, it announced it would terminate its internal "Experiences & Devices" division's collective license for Claude Code by June 30th of this year. The pilot project, launched with fanfare in December 2025, quickly collapsed as token consumption led to bills far exceeding expectations. Microsoft has now been forced to mandate its engineers revert to using GitHub Copilot CLI.
Similarly, Uber's experience resembles an AI fiscal disaster. Uber's CTO, Paraveen Neppalli Naga, recently admitted that the company's $34 billion annual budget provision for AI in 2026 was completely exhausted by April of this year. After rolling out the Claude Code assistant to its 5,000 engineers, monthly active usage soared to 85%-95%. The uncontrolled intensity of use resulted in catastrophic bills: monthly API call costs averaged between $500 and $2,000 per engineer. This "unlimited" consumption rate caught management off guard.
NVIDIA's (NVDA) Vice President of Applied Deep Learning, Bryan Catanzaro, also acknowledged widespread industry anxiety in a recent interview: "In the team I lead, compute costs have far exceeded people costs." In the court of public opinion, debates about an AI bubble have been reignited, with one industry observer commenting, "The cost issue is now the elephant in the room. Everyone claims to have tracking capabilities, but almost no one is actually watching the bill."
IPO Market Polarization: Application Layer Squeezed, Infrastructure Providers Reap Rewards
The soaring cost of tokens poses a severe threat to the valuation of star AI companies planning IPOs by the end of 2026, such as Anthropic and Perplexity. The risk is that, unlike the private market with its higher tolerance for losses, public market investors will scrutinize gross margins and profitability paths extremely harshly. When core input costs are increasing at a rate of 40% every six months, it becomes exceedingly difficult for portfolio companies to prove that revenue growth can outpace cost inflation.
The industry landscape is undergoing intense polarization: infrastructure providers controlling the lifeline of computing power are enjoying the红利. Microsoft Azure, Google Cloud, Amazon.com (AMZN) AWS, and GPU giant NVIDIA maintain absolute pricing power amid supply constraints. Meanwhile, application-layer companies are mired in a profit trough. Consumer-facing AI apps relying on freemium models face the collapse of their unit economics; even enterprise-facing companies with stronger bargaining power struggle to easily pass on costs in a fiercely competitive environment with low switching costs.
To survive, some companies are adopting "model distillation" strategies, routing everyday tasks to cheaper, smaller models and only invoking expensive cutting-edge models for complex, high-end queries. Others are exploring on-premise deployment options for enterprise clients, shifting infrastructure costs to the users. Simultaneously, consolidation is looming, as deep-pocketed giants can absorb costs to squeeze out independent developers. Microsoft's deep integration with OpenAI represents a formidable cost moat.
Hardware Solutions Are a Distant Prospect as AI Enters a "Separating Wheat from Chaff" Moment
Amidst the raging cost crisis, the next-generation hardware that could fundamentally reduce inference costs still requires a lengthy wait. Reports indicate that, despite NVIDIA's acquisition of chip startup Groq and efforts by AMD (AMD), Intel (INTC), and Amazon.com's AWS to redesign AI accelerators specifically for lowering per-token costs, most hardware is not expected to launch until the second half of this year, with large-scale deployment to alleviate supply-demand pressures likely not until early to mid-2027.
In this crisis directed by supply bottlenecks, rising energy prices, and unchecked consumption, the industry has formally entered a "separating wheat from chaff" phase, moving away from野蛮 growth and confronting the reality of profitability. As one AI company executive stated, when the speed of cost increases overwhelms all efficiency gains, the question facing the AI industry is no longer whether AI can change the world, but whether its current fragile business model is standing atop a bubble ready to burst.
Comments