Recently, cutting-edge modeling companies such as OpenAI have started to purchase or use Mac mini in large quantities, which means that modeling companies are rapidly transitioning from selling tokens to selling deliverable work results. This is a shift from AI-as-a-Service to Work-as-a-Service. (Yes, WaaS will kill SaaS faster) In the past, when users purchased AI, they were essentially buying "intelligent capabilities": inputting Prompt, the model generated tokens, and as for how these tokens were converted into work results, it was still mainly done by users themselves. As the capabilities of Agent and Computer Use mature, this pattern is changing. In the future, users are increasingly likely to no longer purchase 'how many tokens', but directly purchase' how much work has been completed '. Therefore, the AI business model may go through three steps: Sell Token → Sell Model+Compute Environment → Sell Outcome. Token will gradually retire to the backend and become an internal cost indicator for the model company. The truly user oriented units of measurement may become Agent hour, Completed Task, Resolved Ticket, or even Human equivalent Work Hour. The core efficiency indicator of the model company will also shift from "how many tokens are generated per dollar" to "how much effective work is completed per dollar". That's also why cutting-edge model companies are starting to purchase large quantities of Mac mini. An agent that can truly complete its work, having only a model as its' brain 'is not enough. It also requires CPU, memory SSD、 Browser, file system, operating system, network, software, and permissions. These resources together constitute the "execution environment" of AI. Therefore, in the future, AI infrastructure may shift from two layers to three layers: Training Compute → Inference/Reasoning Compute → Environment/Action Compute. GPU is responsible for 'thinking', while CPU, Mac, Windows, Linux VM and other environments are responsible for 'doing'. The significance of Mac mini is that it is a complete, inexpensive, and scalable real computer environment. The model can open software, click, input, edit files, run code, observe results, and continue actions within it. For reinforcement learning, a computer itself can become an RL environment. If this model is effective, the model company will have the motivation to continue to integrate downwards. They not only purchase GPUs, but may also heavily purchase CPUs, servers, Macs, and other execution resources, selling models and execution environments together to customers. Enterprises no longer need to prepare their own computers or set up workflows, they only need to set goals. This means that model companies will gradually shift from "providing the brains of digital employees" to "directly providing complete digital employees". And this will change the way users use AI. Last year, people designed Prompt; At the beginning of this year, human design skills; Later on, I set up Workflow. Next, the instructions given by the user to the AI will continue to move up to higher levels of abstraction: Prompt → Skill → Workflow → Goal → Outcome. Future users may only need to define four things: goals, constraints, resource budgets, and acceptance criteria. This means that Workflow will transform from a pre designed process by users to an execution plan dynamically generated by AI at runtime. AI will plan, execute, identify errors, re plan, and reallocate resources on its own. The user's role will also shift from Workflow Designer to Objective Designer, from assembly line manager to company department manager or even CEO. 、 The truly important ability for the future is Objective Engineering: the ability to propose valuable goals, define reasonable constraints, and define what results are considered well achieved. This is actually a continuation of the continuous upward movement of the software abstraction layer. It may eventually become: People are responsible for defining deliverables, while AI is responsible for allocating Intelligence, Compute, and Actions. And once this is achieved, there will be a new order of magnitude increase in AI usage. Today, AI consumption is still subject to an important limitation: human management of AI bandwidth. There is still a limit to how many workflows a person can set up for supervision and how many agents they can manage. But if in the future humans are only responsible for setting goals, and AI disassembles tasks, manages agents, and schedules resources on its own, then one person can simultaneously make dozens or even hundreds of agents on Mac Mini work at the same time. This could be thousands of agents. At this point, the measurement of personal computing ability may also shift from 'how fast my computer is' to: How many Autonomous Work Streams can I run simultaneously. So the demand for AI computing power will shift from simple Token Scaling to Digital Labor Scaling. In the long run, computers themselves may also undergo changes as a result: if the primary user becomes an agent, future machines may not necessarily require screens and keyboards, but will place more emphasis on large memory, high-speed SSDs, networking, virtualization, remote management Snapshot、 Permission isolation and 24-hour operation. Mac mini has come close to being an early AI Worker Box to some extent. The software industry, especially SaaS, will also be changed. Computer Use allows AI to immediately operate existing GUI such as Salesforce, SAP, Excel, etc. This is an important compatibility layer, but it will accelerate SaaS from front-end and back-end software to back-end software; SaaS customers who can still survive will also shift from Human Seat to Human+Agent, and the charging model may further shift from Seat to Agent, Transaction, and Outcome. Meanwhile, when the agent can operate emails, code CRM、 When it comes to cloud services and financial accounts, new Agent Identity, Permission, Audit, Spending Limit, and Security infrastructure will also be generated. Ultimately, this industrial evolution chain may be: Token Economy → Agent Economy → Digital Labor Economy → Outcome Economy. The data center may also evolve from today's Token Factory to a Digital Labor Factory: Input electricity GPU、CPU、 Software and data output written programs, completed research, and processed orders.