The significance of deploying cost-effective open-source models and reselling tokens for CSP (cloud service providers) lies in: 1) reshaping the CSP business model; 2) Get rid of the dilemma of AI commercialization only focusing on closed source cutting-edge models ARR (CSP performance is equally important). Yesterday, Coinbase's AI cost engineering instance used engineering methods to suppress cost expenditures - routing to open source models, increasing cache hit rates from 5% to 60%, streamlining context, and moving towards "using more at a cheaper price" rather than "using less". It is bound to drive many companies to follow suit, of course, Coinbase itself is a technology enterprise and has the ability to configure its own internal AI operating system. However, most enterprises do not actually have this capability and ultimately rely on cloud providers. Deploying open-source models with high cost-effectiveness in CSP can significantly reduce its own marginal costs and improve the profitability of AI related businesses. This is precisely one of the core business logics of current cloud vendors (AWS, Azure, Google Cloud, as well as domestic Alibaba Cloud, Tencent Cloud, etc.) vigorously promoting open source models to the cloud. 1. Why is the cost lower? Compared with reselling closed source cutting-edge models (selling Claude, GPT, Gemini through channels such as Bedrock, Azure OpenAI, Vertex AI, etc.), deploying open source models (Llama, DeepSeek, Qwen, Mistral, etc.) has significant marginal cost advantages: 1) When reselling closed source models without model authorization/sharing costs, CSP receives a limited sharing ratio (usually 20-50%, depending on the contract) and also bears the pricing pressure of the model provider. Self hosted open source model resale, with almost zero licensing costs for open source models, CSP only needs to bear their own computing power, electricity, and operation costs. CSP takes almost all of the markup (after deducting computing power costs). Because pricing can refer to the actual cost of the open source community plus a reasonable premium, the space is larger. 2) Hardware and inference optimization control rights - CSP can deeply optimize the serving of open source models (quantization, batch processing, speculative decoding, vLLM/TensorRT-LLM, etc.) using its own customized chips (AWS Inferentia/Trainium, Google TPU, Alibaba Cloud self-developed chips, etc.). The actual inference cost of open source models on the same hardware is usually only 1/5 to 1/10 or even lower than that of closed source models. 3) Scale effect and utilization improvement - CSP has a massive GPU/accelerator cluster, making it easier to achieve high utilization (especially for low to medium complexity and high-frequency tasks) after deploying open source models. High utilization directly dilutes fixed costs. 2. This will bring practical commercial effects 1) CSP can provide customers with "sufficient" models at lower prices while maintaining high gross margins. 2) This helps attract a large number of small and medium-sized enterprises and cost sensitive large customers, and enhance the overall AI revenue scale and stickiness of the platform. 3) In multi model routing scenarios, CSPs can also default route simple tasks to open source models through their own Gateway, further improving the overall profit margin of their own platform. Individual enterprise AI spending has decreased (such as Coinbase), but more enterprises have adopted AI (with a larger total plate) Several cloud providers are already doing this: AWS Bedrock has launched a large number of open-source models such as DeepSeek and Llama; Google Vertex AI Model Garden、 Azure AI Studio also vigorously promotes open source options. This is not charity, but profit driven. It also reshaped the cost structure of CSP. 3. Greater significance 1) Previously, the AI commercialization market first looked at the ARR of cutting-edge closed source large model manufacturers, especially Authropic's ARR. In terms of its impact on the stock market, the weight of the ARR model is higher than the cloud business growth rate of CSP; 2) However, Coinbase's engineering practice also demonstrates the role of open-source models in high-frequency, low value scenarios. From the perspective of cloud providers, a large amount of data center and computing power investment is dominated by CSP. If they cannot dominate in commercialization, then the input-output ratio is indeed disproportionate. 3) Deploying open-source models for CSP, occupying high-frequency, medium low value business scenarios, and changing the cost structure, is a great benefit for CSP. The commercialization return of CSP is better, and naturally their huge capital expenditure ROI is even more beautiful. This will also greatly enhance the overall commercialization process of AI and accelerate the pace of enterprise adoption of AI. In the future, when it comes to the commercialization of AI, we cannot only focus on big model manufacturers, and the performance of CSP is equally important, not just the second weight. In general Deploying cost-effective open-source models using CSP and reselling tokens is currently one of the most effective ways to increase the profit margin of AI businesses. It essentially commodifies the 'model layer', and CSP transfers more value to the infrastructure+platform service layer (hosting, optimization, routing, governance, security). This is highly consistent with yesterday's discussion on Coinbase's "Enterprise AI Adoption Enters Cost Engineering Stage" - CSP is becoming a key "middle layer" in enterprise multi model architectures, helping customers reduce costs while earning a higher proportion of profits on their own. The future trend is very clear: Closed source cutting-edge models will continue to follow the high-end premium route, while open source models+CSP hosting will become the main force in mid to low end and high-frequency scenarios. A CSP that can achieve extreme low cost and high reliability in open source model serving will gain significant profit margin advantages in the AI era.