从不限量到自费上班 互联网公司们付不起Token账单了

This article covers the latest industry shift in the internet sector, where unlimited AI token usage policies are being replaced by quota-based management. Key takeaways are as follows:

1. Industry-wide token consumption has far outpaced initial projections, leaving major global and domestic tech firms facing severe cost overruns: Uber exhausted its entire 2026 AI budget in just four months; Meta spent over $100 million on token costs in 30 days; one domestic Chinese AI project burned through ¥2 million worth of tokens in a single night; if Tencent maintained unlimited access for all employees, annual token costs would hit ¥25.2 billion, accounting for nearly 30% of its projected 2025 R&D budget.

2. Adjustment measures adopted by firms: Tencent now dynamically allocates quotas based on work tasks, prioritizes free internal access to its Hunyuan large language model, sets large quota gaps across different roles, and allows employees to apply for additional quotas from supervisors. Most other major Chinese tech firms implement role-based differentiated quotas with approval processes for extra token purchases. Most small and medium-sized enterprises (SMEs) set fixed monthly quotas and promote more cost-effective domestic large models.

3. Industry trend: After a period of blindly encouraging token consumption that resulted in massive waste, with many firms failing to improve operational efficiency despite heavy investment, the industry is now moving beyond blind token worship and toward pragmatic, rational AI implementation.

This article outlines a new trend in enterprise AI adoption and offers actionable insights for brands looking to manage AI costs and implement AI effectively. Key takeaways are as follows:

1. A critical cost reminder: The actual cost of rolling out AI is far higher than most firms initially expect. Even deep-pocketed top internet giants cannot absorb the cost of unlimited token usage, and the entire industry is now tightening AI cost controls. Brands should draft detailed cost budgets in advance and reserve buffer space for adjustments to avoid cost overruns derailing overall budgets.

2. A reference for implementation strategy: The established industry best practice is dynamically allocating token quotas based on actual business value, prioritizing cost-effective domestic large models, and guaranteeing sufficient resources only for projects that deliver tangible value, to eliminate unnecessary waste.

3. A reference for performance measurement: Many brands previously used token consumption as a metric to evaluate AI adoption, but this approach has proven to encourage waste and fails to improve efficiency. Baidu CEO Robin Li's proposed DAA (Daily Active Agents) metric, which better reflects the actual value of AI implementation, is a useful framework for brands.

The shift to cost controls for enterprise AI use brings both risks and opportunities for sellers leveraging AI. Key insights are summarized below:

1. Risk warning: The actual cost of AI adoption is far higher than initial projections, and unlimited usage can quickly lead to major budget overruns. Many SMEs have already shifted from unlimited access to monthly quota systems, and some even require employees to pay for extra tokens out of pocket when they exceed their quotas. Sellers adopting AI to cut costs and boost efficiency must implement cost controls upfront to avoid runaway spending from blind investment.

2. Actionable implementation approaches: The mainstream industry cost control model, which dynamically allocates token quotas based on tasks, prioritizes cost-effective domestic large models, and reserves sufficient quotas for value-generating projects, controls costs without disrupting AI integration for core business. It is particularly well-suited for small and medium-sized sellers.

3. Opportunity outlook: The industry has shifted from irrational AI hype to pragmatic implementation. Domestic large models are gaining adoption across more firms thanks to their cost advantages. Sellers can capitalize on this trend to build custom AI efficiency tools aligned with their business at low cost, and reduce overall operating expenses.

This article outlines real challenges in enterprise AI transformation and offers key insights for factories advancing digital and AI transformation. Key takeaways are as follows:

1. Implement AI transformation with rational cost controls: Many internet firms initially encouraged unlimited employee AI access to tokens, only to end up with far higher costs than budgeted, leaving many SMEs unable to absorb overspending and forced to tighten quotas. Factories adopting AI for use cases including product design, production scheduling and customer service should calculate costs carefully upfront, avoid blindly pursuing full-stack AI adoption or premium overseas models, and prevent runaway investment.

2. Guidance on implementation paths: The industry's standard pragmatic implementation model, which allocates quotas based on actual business needs, prioritizes cost-effective domestic large models, and reserves resources only for value-generating projects, is well-suited for factory AI transformation. Factories can start with simple, easy-to-implement use cases to verify value before scaling up investment gradually, avoiding excessive one-time spending.

3. New business opportunities: As more firms shift to domestic large models, demand for customized AI applications tailored to specific industries is growing rapidly. Eligible factories can explore partnerships with domestic AI platforms to develop exclusive AI tools customized for manufacturing production and design needs, and unlock new business growth opportunities.

This article identifies the core pain points of enterprise AI clients and outlines emerging industry trends, offering valuable insights for AI service providers. Key takeaways are as follows:

1. Core client pain points: Enterprise demand for AI has shifted from early-stage experimentation to practical implementation, and the most pressing pain point is prohibitively high AI usage costs: unlimited access models easily lead to budget overruns. The market also faces a clear capability gap: premium overseas models are too costly, while low-cost domestic models cannot handle complex tasks. Many firms even require employees to pay for extra tokens out of pocket when quotas run out, creating a clear opportunity for service providers to enter the market.

2. Industry development trends: Enterprises are no longer using token consumption as the core metric for AI adoption, and are instead focusing on actual output value. Robin Li's proposed DAA (Daily Active Agents) metric will likely become the new standard for measuring AI value, as the entire industry shifts from AI hype to rational, practical implementation.

3. Directions for solution development: Service providers can launch tiered token pricing packages tailored to the needs of different roles and projects. They can also accelerate improvements to the complex task processing capabilities of domestic large models to meet enterprises' cost reduction goals. Additionally, they can explore new pay-for-output billing models aligned with enterprises' current performance measurement frameworks.

This article outlines the latest shifts in enterprise demand for large AI models and offers guidance for platform operators on management and risk mitigation. Key takeaways are as follows:

1. Updated client demand: Enterprise demand for large AI models has shifted from unlimited usage to controlled-cost targeted access. A growing number of enterprises now prioritize cost-effective domestic models, and demand for features including differentiated quota management and project-based additional quota applications has risen sharply. Platforms need to quickly adapt to clients' cost control needs and develop corresponding functional modules.

2. Guidance on operations and business development: Large enterprises now require employees to prioritize internal models and block competing external models, while small and medium clients prefer lower-cost models. Platforms can develop differentiated products for different client segments: offer customized quota allocation services for large enterprise clients, and launch cost-effective bundled packages for small and medium clients to attract more users to the platform.

3. Risk mitigation guidance: The previous industry focus on token consumption led to massive unnecessary waste, and the sector is now moving past this hype. Platforms should avoid overmarketing the concept of token volume, and instead guide clients to focus on actual AI output. They can also adopt the DAA metric to measure platform ecosystem health, to mitigate the risk of falling demand as the industry matures.

This article documents the latest developments and core challenges in AI industry implementation, providing first-hand data for industrial research. Key takeaways are as follows:

1. Latest industry developments: The global AI industry is shifting from early-stage blind expansion to rational implementation. From global and domestic top giants to small and medium-sized internet firms, all have adjusted their AI token supply strategies: from previously encouraging maximum employee token usage (even linking consumption to promotions and salary increases) to dynamic quota allocation based on tasks, and promoting domestic large models as replacements for premium overseas models. The entire industry is now tightening AI cost controls, and blind "token worship" is being abandoned.

2. Core new industry challenges: The current cost structure of large AI models is unreasonable, and the actual cost of enterprise AI adoption is far higher than expected. Widespread token waste has left many firms failing to improve operational efficiency despite heavy investment, even slowing down core product development. The market also faces clear supply-demand mismatches: low-cost domestic models lack sufficient capability, while premium models are too expensive to meet enterprises' complex needs, which is the core contradiction the industry must resolve today.

3. New research directions: Robin Li has proposed replacing token consumption with DAA (Daily Active Agents) as the metric for measuring the prosperity of AI platform ecosystems, offering a new framework for industrial research and business model evaluation that merits further in-depth study.