DeepSeek模型降价75% 难解行业百倍token消耗问题

This article discusses why a 75% price cut for the DeepSeek model has failed to resolve the high-cost problem plaguing the AI industry, with key takeaways as follows:

1. The core contradiction has upended the industry's long-held consensus. It was widely believed that, like traditional software, AI infrastructure costs would decline year-over-year, and inference costs would eventually become negligible. In practice, however, the growth rate of token consumption has far outpaced the decline in per-token pricing — even a 75% price cut is not enough to close this consumption gap.

2. Cost structures vary dramatically across different AI applications. For basic single-turn chatbots, the ratio of input tokens to output tokens is just 1:5. For multi-step AI agents that serve multiple scenarios, this ratio can reach as high as 1:700 or more. Even simple queries can require tens of thousands of tokens, pushing monthly costs for one million monthly queries into the six-figure range.

3. The industry has already developed mature cost-reduction frameworks. Companies can maintain healthy gross margins by implementing five key cost-control practices. In the long run, businesses with clear cost management strategies will survive, rather than those that rely solely on adopting low-cost models.

As AI adoption becomes widespread among brands, uncontrolled costs have emerged as the top challenge, with key takeaways as follows:

1. When brands deploy AI agents for internal and customer-facing use cases such as customer support, sales operations and legal review, token consumption far outpaces that of traditional AI applications, making costs very easy to spiral out of control. The traditional seat-based SaaS subscription model is no longer viable: inference costs for heavy users can easily exceed their monthly subscription fees, leading to negative gross margins for the corresponding business line.

2. Multiple industry-validated cost-reduction strategies are already available for reference. These include cost-aware routing to assign requests across different models, leveraging cache discounts from major model providers, pruning redundant context to eliminate unnecessary consumption, and securing steep discounts via bulk pre-purchasing, all of which effectively cut inference costs.

3. For brands implementing AI, inference costs must be treated as a core management KPI. Setting cost alerts and regularly pruning redundant prompts are required to ensure AI deployments remain economically viable and avoid dragging down overall profits.

The AI agent sector now faces clear cost risks, alongside defined paths for mitigation, with key takeaways as follows:

1. Risk warning: Model price cuts do not translate to lower overall AI costs. The growth of token consumption for multi-scenario AI agents far outpaces the rate of model price cuts, and the traditional seat-based subscription model, the dominant business model for AI SaaS, is no longer economically viable. Even Salesforce has faced gaps between its marketing demonstrations and actual delivered functionality, as the promised features are not economically feasible to deliver. Multiple sellers already face negative gross margins for lines of business serving heavy users, meaning the risk has shifted from a theoretical concern to real bottom-line pressure.

2. Opportunities and mitigation: Sellers can control costs and maintain healthy gross margins through five key actions: track inference costs as a top-level KPI broken down by dimension, set cost caps per thousand queries that trigger alerts, build routing systems as core infrastructure, audit and prune system prompts quarterly, and negotiate steep discounts via bulk pre-purchasing.

3. Opportunity outlook: Future winners in the AI space will be players with strong cost control. Building out cost management capabilities early creates sustainable differentiated competitive advantage.

Factories can draw the following insights from current industry trends as they advance digital transformation and deploy AI applications, with key takeaways as follows:

1. When factories deploy AI for multi-scenario, multi-step agent use cases including production management, finance, legal work and customer support, token consumption is far higher than for basic chatbot AI applications. As a result, costs can easily exceed original budgets. Factories should not rely solely on the public list prices of large models, and must pre-calculate actual token consumption costs to avoid overspending on transformation initiatives.

2. Factories can adopt mature industry-proven cost-reduction strategies: using cost-aware routing to assign model requests, leveraging cache discounts from major model providers, pruning redundant context to eliminate unnecessary consumption, applying speculative decoding to improve GPU utilization, and building an orchestration layer to enable granular cost accounting. IBM data shows that organizations with an orchestration layer deliver six times the productivity gains compared to those that only implement compliance-focused solutions.

3. Factories must include AI inference costs as a core cost control KPI and regularly audit for redundant content. This ensures that AI-powered digital transformation remains economically viable, and delivers on the core promise of AI-driven cost reduction and efficiency gains.

The core pain point for the AI services industry today is margin compression driven by the widespread adoption of AI agents, and the sector is shifting toward a new paradigm where cost control is the core competitive advantage, with key takeaways as follows:

1. Current core customer pain point: Price cuts from large model providers cannot offset the growth in token consumption. The traditional seat-based SaaS subscription pricing model, the industry standard, is no longer suited for AI agent use cases. Inference costs for heavy users now exceed the subscription fees they pay, and multiple AI service providers already face negative gross margins for these lines of business. Cost pressure has shifted from a theoretical risk to a major material headwind for profitability.

2. The technology sector has already reached a consensus on mature cost-reduction solutions: cost-aware routing can cut inference costs by 60% without reducing output quality; top large model providers offer 75% to 90% discounts for cached prompt prefixes; context constraints eliminate unnecessary consumption; speculative decoding increases throughput by 2x to 3x; and an orchestration layer enables granular cost accounting, delivering six times the productivity gains of standard approaches.

3. Future industry winners will be service providers that prioritize cost control. Building cost management as a core capability is required to maintain healthy gross margins.

The AI industry now has new demands for platform operators, and traditional models are no longer suited to current needs. Platforms need to adjust their operating strategies to adapt to these changes, with key takeaways as follows:

1. Core market needs and problems: After enterprises deploy AI agent applications, token consumption has grown exponentially. A 75% cut in per-token pricing from large models cannot close the 100x consumption gap. Traditional per-token pricing for models and seat-based subscription pricing for service providers no longer meet customer demand for cost-effectiveness, and uncontrolled costs are now a widespread industry-wide problem. This is the first time in the cloud era that architectural decisions have a direct, real-time impact on financial results.

2. Leading platforms already have established best practices for reference: Top platforms including Anthropic, OpenAI and Google already offer 75% to 90% discounts for cached prompt prefixes, and have launched prepaid bulk plans similar to cloud reserved instances that offer far deeper discounts than public list prices, helping customers cut overall costs.

3. Risk mitigation directions for platform operations: Platforms need to guide customers to implement granular cost control, and can launch dedicated cost management tools to serve customers, helping them maintain healthy gross margins. This approach supports the stable and healthy development of the platform ecosystem, and attracts more enterprise customers.

New developments and challenges have emerged in the AI industry, upending long-held assumptions about industry growth logic and business models, and offering new directions for industrial research, with key takeaways as follows:

1. New industry developments and challenges: The long-held growth consensus of the AI industry has been broken. It was previously assumed that AI would follow the same development trajectory as traditional software, with infrastructure costs declining annually and inference costs eventually becoming negligible operating expenses. This assumption has now been invalidated: the token amplification effect grows much faster than the decline in per-unit inference costs, and even a 75% per-token price cut cannot close the 100x consumption gap.

2. New changes in business models: The traditional dominant seat-based SaaS business model for AI companies is no longer viable for AI agent workflows. It is common for inference costs from heavy users to exceed subscription revenue, leading to negative gross margins for the corresponding business, and this problem has shifted from a theoretical risk to a material profitability headwind.

3. New directions for industrial research: This is the first time in the cloud era that architectural decisions have a direct, real-time impact on financial results. The core competitive advantage in the AI industry has already shifted from model performance to granular cost management capabilities, which opens up an entirely new direction for research on AI industry competition and business model evolution.