引導式測驗 · 中等 等級
人工智慧業務測驗中的毛利率和推理成本
測試您對 AI 毛利率為何落後於經典 SaaS、推理成本如何隨使用量變化以及路由和快取如何保護利潤的理解。
問題 1 的 8
根據指南,為什麼對 AI 產品而言,服務額外用戶的成本通常比傳統 SaaS 更高?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- 根據指南,為什麼對 AI 產品而言,服務額外用戶的成本通常比傳統 SaaS 更高?
- AI products must buy a separate software license for each user
- AI products need far larger customer databases
- Every prompt consumes GPU time, a variable cost that grows with usage
- AI products require more salespeople per customer
- 模型路由有什麼作用?
- Sends easy requests to small, cheap models and keeps large models for hard ones
- Sends every request to the most capable model to guarantee quality
- Spreads requests evenly across several providers to get volume discounts
- Stores model outputs so identical questions are never answered twice
- Andreessen Horowitz 在 2020 年發表的文章「人工智慧的新商業模式」("The New Business of AI")提出了什麼論點?
- AI companies would soon have higher gross margins than any SaaS business
- Inference costs were too small to affect AI company finances
- AI companies should avoid using cloud providers entirely
- Many AI companies looked like a mix of software and services, with lower margins from compute and humans in the loop
- 為什麼 token 價格下降不會自動改善 AI 的毛利率?
- Token prices have not actually fallen for any capability level
- Products often use up the savings by moving to newer models, longer contexts and agentic workflows
- Gross margin does not include compute costs
- Providers raise prices again every time they cut them
- 為了充分利用供應商的 prompt 快取,提示詞(prompt)應該如何結構化?
- Put the user's question first so the model reads it before anything else
- Randomize the order of instructions on each call to avoid stale answers
- Put static instructions, tool definitions and reference documents first, and variable user content last
- Split every prompt into many small calls so each one is cached separately
- Why does the guide recommend tracking contribution margin by customer cohort instead of relying on averages?
- Averages overstate the cost of light users
- Cohort tracking is required by accounting standards
- Averages make routing impossible
- Averages hide the power users whose heavy usage can wipe out profit
- How does a routing cascade work?
- It tries a small model first and escalates to a larger one when a validation check or confidence signal fails
- It sends each request to every available model and picks the longest answer
- It always starts with the largest model and falls back to smaller ones when budgets run out
- It stores each answer and replays it to future users with different questions
- Why should cost be modeled per task rather than per API call?
- Providers bill by task, not by token
- An agent may make many calls for a single user action, so per-call figures understate true cost
- Per-call costs are always zero for cached prompts
- Tasks and calls always cost exactly the same