Mozilla CTO Argues Open Source AI Is About Leverage, Not Just Licences
Mozilla CTO Raffi Krikorian arrived at a pointed conclusion during a conversation with Tim O'Reilly at O'Reilly's recent AI Codecon: teams building on AI should try to become "owners," not "renters." The discussion has since been published by O'Reilly Radar as an interview.
The argument is less about open source licensing ideology than about plain dependency risk. Once a team places core workloads on a vendor-hosted endpoint, pricing adjustments, rate limits, silent model deprecations, and inference that cannot be audited can all reshape operating costs and compliance posture without warning. Krikorian's framing is ultimately a strategic one: decide which capabilities you need to be able to withdraw at any moment, and which you can afford to have withdrawn from you.
Mozilla itself serves as a case study in that logic. The nonprofit helped shape the open source movement in the late 1990s by releasing the Netscape browser's source code, and the Firefox browser it launched afterwards went on to challenge far larger platform players before being overtaken by better-funded rivals. Mozilla's long-standing reliance on search engine revenue share also shows what losing control of a critical platform layer actually costs.
One detail worth noting: the interview transcript was generated by Claude and lightly edited by Tim O'Reilly — a piece about AI tooling, written by an AI tool, which makes for a useful piece of self-referential evidence.
What this means for build-vs-buy decisions in Hong Kong
For teams weighing LLM build-versus-buy choices in Hong Kong, "rent vs. own" is not a binary but a spectrum that has to be divided deliberately:
- Workloads that lean toward ownership: inference pipelines touching core intellectual property, customer data, or regulated data. Relying on a third-party endpoint raises questions about data flow, retention location, and the vendor's own auditability; data processor arrangements, cross-border transfers, and data residency obligations under Hong Kong's Personal Data (Privacy) Ordinance all need to be settled before contracts are signed.
- Scenarios that lean toward renting: experimental, spiky, or non-critical tasks, where calling a commercial API is usually the cheaper option.
- The real failure mode: letting a "rented" solution quietly harden into load-bearing infrastructure without anyone making the decision — and only then discovering that pricing, model versions, and data policies are no longer in your hands.
Self-hosting costs are also routinely underestimated. Deploying open-weight models such as Qwen does deliver weight-level control, offline operation, and data that never leaves your jurisdiction, but the ongoing total cost of ownership — inference optimisation, observability, model updates, security patching, and staffing — tends to be far larger than the GPU bill alone suggests. Mature teams usually layer the spectrum: sensitive and core components kept in-house, long-tail demand rented out, and every new dependency explicitly recorded as either rented or owned.
Seen from that angle, "open" is being negotiated at the model layer for the first time — a shift distinct from earlier open source debates, which were fought over code and licences at the application layer. Whether a given model counts as open may matter less than whether the team holding it has the leverage to act if the terms change.
The full interview is available in the original article.
Mozilla CTO：開源 AI 的關鍵在於議價能力，不只是一紙授權
Mozilla 首席技術總監 Raffi Krikorian 在最近一場 O'Reilly AI Codecon 上與 Tim O'Reilly 對談時，得出一個耐人尋味的結論：在 AI 之上構建的團隊，應該努力成為「擁有者」（owners），而非「租用者」（renters）。這場討論其後由 O'Reilly Radar 以專訪形式刊出。
這個論點與其說是關於開源授權的意識形態，不如說是關於實實在在的依賴風險。一旦團隊將核心工作負載部署在供應商託管的 endpoint 上，價格調整、rate limit、悄然下架的模型版本，以及無法審計的 inference，都可能在毫無預警之下改變營運成本與合規狀態。Krikorian 的框架歸根到底是一個策略問題：哪些能力你必須能夠隨時抽走，哪些你又承受得起被對方抽走。
Mozilla 本身就是這套邏輯的案例。這家非營利機構在 1990 年代末透過公開 Netscape 瀏覽器的原始碼，塑造了開源運動的發展；其後推出的 Firefox 瀏覽器曾挑戰規模大得多的平台巨頭，直至被資金更雄厚的對手超越。Mozilla 長期依賴搜索引擎收入分成，亦正好說明失去一個關鍵平台層的控制權究竟要付出什麼代價。
有一個細節值得一提：專訪的文字稿由 Claude 生成，再由 Tim O'Reilly 作輕微編輯——一篇關於 AI 工具的文章，由 AI 工具撰寫，構成了一份頗為貼題的自證材料。
這對香港的自建還是外購決策意味著什麼
對於正在香港衡量 LLM 自建與外購選擇的團隊而言，「租還是擁有」並非非黑即白，而是一個需要刻意劃分的光譜：
- 傾向擁有的工作負載： 涉及核心知識產權、客戶資料或受規管資料的 inference pipeline。依賴第三方 endpoint 會引發資料流向、保存地點以及供應商本身可審計性的問題；在香港《個人資料（私隱）條例》之下，資料處理者安排、跨境轉移及資料所在地的責任，均須在簽署合約之前釐清。
- 傾向租用的場景： 實驗性、流量波動大或非關鍵的任務，這類任務調用商業 API 通常成本更低。
- 真正的失效模式： 一套「租用」方案在無人正式作決定的情況下，悄然硬化為支撐業務的基礎設施——到頭來才發現價格、模型版本及資料政策早已不在自己手中。
自行託管的成本同樣經常被低估。部署 Qwen 等 open-weight 模型固然可以帶來 weight 層面的控制、離線運行，以及資料永不離開自身司法管轄範圍的保障，但持續的總擁有成本——inference 優化、observability、模型更新、安全補丁及人手配置——往往遠非 GPU 帳單所能反映。成熟的團隊通常會為這個光譜分層：敏感及核心組件留在內部，長尾需求向外租用，並把每一個新依賴明確記錄為租用或擁有。
從這個角度觀之，「open」首次在 model 層面被重新談判——這一次轉變有別於早年的開源論戰，當年的爭議聚焦於應用層的 code 與 licence。某個模型是否算得上 open，或許已不如持有它的團隊在條款改變時是否擁有議價能力來得重要。
完整專訪可見於原文。