Tech giants who once encouraged employees to spend tokens wildly are now collectively hitting the brakes, as a wave of "token cost-cutting" sweeping giants like AT&T, Meta, and Uber is fully underway.

At the core of this shift is runaway cost pressures. Companies with the highest AI usage intensity spend as much as $7,500 per employee per month, and Uber even spent its entire budget in April.

Big companies have realized that AI is too costly. Previously, they created "consumption rankings" to encourage innovation, but now they set caps to save money. But this also creates a dilemma: while forcibly limiting AI usage can save money, it may also cut the productivity gains AI brings. Companies are caught in the dilemma of "efficiency or wallet."

| From "Mindless Money Burning" to "Extreme Cost-Cutting"

The root cause of cost runaway lies in the widespread adoption of agent tools.

These tools repeatedly and automatically call models during task execution, causing the company's overall bill to triple compared to before. Faced with sky-high bills, giants have shifted from "token maxxing" to "token minizing." Amazon has directly abolished its internal consumption rankings, and Uber and Walmart have also set monthly usage caps.

However, not all businesses choose to tighten. Box executives stated that the company has never gone astray, while Databricks insists on not setting a budget cap.

This divergence in response reveals the inherent tension of token-cutting policies: controlling usage can certainly reduce costs, but it may also simultaneously cut the productivity gains AI originally promised—precisely the main justification companies made for this expenditure.

| Infrastructure is on the rise: "Model routing" has become a core necessity

On the other side of the "token cost-cutting" wave is the explosive structural demand for cost control infrastructure.

More and more companies are migrating simple tasks from high-cost, cutting-edge models to cheaper or open-source alternative models to control costs without cutting actual usage.

微软、Databricks等巨头正迅速推出“网关”工具，英伟达投资的Factory也发布了新款模型路由器，旨在将低复杂度任务自动分配给低成本模型。

微软CEO纳德拉公开主张AI模型应像商品一样可替换，避免价值被少数模型垄断。与此同时，微软正推出“座位费+消耗量”的新计费模式，并测试用开源模型替代昂贵的Anthropic模型。

在“Token节流”时代，如何在维持产品竞争力的同时缓解客户的成本焦虑，已成为企业软件市场新一轮竞争的核心命题。

| Looking for companies in the era of "token cost-cutting."

As corporate AI budgets enter a stage of refined control, the era of simply pursuing the strongest models is over. In the future, whoever can complete the same tasks at a lower cost will win the favor of corporate clients.

Under cost pressures, "gateway" tools that can automatically monitor AI usage, enforce spending caps, and dynamically match optimal models are expected to experience explosive growth. Focus on infrastructure providers such as Microsoft, Databricks, and Factory who are positioning themselves in this sector.

The trend of enterprises migrating simple tasks to cheap models is irreversible. AI companies that can provide high-quality open-source model fine-tuning services and achieve "alternative" services in specific scenarios will gain huge incremental market growth.

JCT's Perspective | Focus on enterprise-level SaaS with a "cost-controllable" architecture

Enterprise AI spending has shifted from "unlimited use" to "meticulous budgeting," marking the official entry of the AI industry into the era of "token economics."

Cost control tools and cost-effective open-source models will bring structural dividends; However, the real winners are platform companies that can build "model routing" capabilities and find the perfect balance between intelligence and cost.

For investors, closely monitoring those who help big companies "save money" is currently the most certain investment theme.

Future enterprise-level AI products must have the gene of "controllable costs." Software vendors that support multi-model switching and offer pay-as-you-go options will hold an absolute advantage in the next round of enterprise software procurement cycles.