# Canva付出三分之一營收，買到一堂全棧自建課

- URL: https://justfly.idv.tw/canva%e4%bb%98%e5%87%ba%e4%b8%89%e5%88%86%e4%b9%8b%e4%b8%80%e7%87%9f%e6%94%b6%ef%bc%8c%e8%b2%b7%e5%88%b0%e4%b8%80%e5%a0%82%e5%85%a8%e6%a3%a7%e8%87%aa%e5%bb%ba%e8%aa%b2/
- 日期: 2026-09-01
- 分類: 我知故我在
- 標籤: anthropic, app, claude, 技術與AI

![Canva付出三分之一營收，買到一堂全棧自建課]
2026年，Canva執行長Melanie Perkins向股東坦承：「多個自研模型尚未準備好發布，定價、使用量模型與使用控制未能跟上需求爆炸。」營收預測因此下修三分之一。這句話的意思翻譯成工程語言就是：核心功能外包給別家API，使用量失控，帳單追上來時已經來不及。

##### 外包算力的天花板

過去幾年，「API先行」是最合理的產品決策。OpenAI、Anthropic、Google的前沿模型能力強、接入快，產品團隊三個月就能推出AI功能，VC也喜歡看到這種速度。問題出在[邊際成本](https://zh.wikipedia.org/wiki/%E8%BE%B9%E9%99%85%E6%88%90%E6%9C%AC)結構：每個用戶請求都是直接燒向第三方帳單的一筆費用，使用量越高、虧得越快。Canva撞上的就是這道牆。

Snapchat稍早也發出同樣的信號，停止推薦純AI生成影片，原因和Canva一致——成本結構撐不住。同樣的架構，在同樣的負載下斷裂。

##### 做AI的公司，早就開始自己造晶片了

Anthropic在2026年8月正式確認組建自研晶片團隊，目標是為Claude「共同設計硬體與模型」，讓推理在客戶需要的規模下跑得更快、更便宜。洽談代工對象是三星，職缺薪資開到年薪32萬至48.5萬美元，要求應徵者「親自參與過晶片量產流程」——要真的把矽片做出來。

同一時間，OpenAI的自研晶片「Jalapeño」已於2026年6月與Broadcom合作推出，Meta自研晶片預計量產。六大AI玩家——OpenAI、Anthropic、Meta、Google、Amazon、Microsoft——全部進入自研晶片賽道，方向一致：減少對NVIDIA的依賴，把算力成本結構掌握在自己手裡。

Anthropic本身在2026年4月達到300億美元年化營收，同時是Google Cloud最大的TPU客戶，2027年起計畫動用數百億瓦等級的計算力。年化營收300億美元的公司還在外面租算力、同時自己造晶片——這兩件事同時發生，代表未來的成本結構必須自己掌控。

##### 張一鳴下令不准蒸餾，代表什麼

ByteDance創辦人張一鳴禁止員工蒸餾競爭對手模型，包括OpenAI與Anthropic，同時宣告啟動中國規模最大的AI訓練計畫，對標Anthropic Mythos等級的前沿模型。「蒸餾」本來是追趕者的捷徑——用強模型的輸出去訓練弱模型，快速獲得接近的能力。張一鳴選擇放棄這條路，押注從零構建。

這個決定在商業上意味著更高的訓練成本、更長的週期，但換來的是一件事：模型能力的自主性。蒸餾出來的模型，能力上限永遠被原始模型框住；從零訓練，才有可能在某個維度超越。

禁令本身也是一種聲明：ByteDance不打算再做二線跟隨者。

##### 垂直整合就是生存邏輯

從晶片到模型到應用，這條鏈條上每一層都外包，意味著每一層都有成本洩漏點、都有能力天花板、都受制於上游的定價決策。Canva的案例清楚展示了最下游那層的脆弱：產品做得越好、用戶越多，賠得越快。

NVIDIA宣布與Apollo、BlackRock、Blackstone等六大金融機構合作，動員逾5000億美元建構AI算力基礎設施。這個規模的資本在移動，說明算力本身已經是基礎設施投資，不是軟體公司可以用API費用線性外包的消耗品。

對那些正在用第三方前沿API構建核心功能的產品團隊來說，Canva的數字是[先行指標](https://zh.wikipedia.org/wiki/%E5%85%88%E8%A1%8C%E6%8C%87%E6%A8%99)。使用量控制機制、自研模型的時間節點、API依賴比例的上限——這些工程決策現在要做，不是等帳單來了才想。

Anthropic造晶片，ByteDance從零訓練，Canva下修三分之一營收。這三件事發生在同一個月，同一個結構在三個位置同時顯形。

— 邱柏宇

### Canva Lost a Third of Its Revenue Trusting Someone Else’s AI

In August 2026, Canva cut its revenue forecast by a third. CEO Melanie Perkins told shareholders that “multiple in-house models weren’t ready to ship, and pricing, usage models, and usage controls couldn’t keep up with the explosion in demand.” Translated into engineering terms: the company outsourced its core AI capability to third-party APIs, usage scaled faster than expected, and the bills caught up before any internal alternative was ready.

##### The API Ceiling

For a few years, “API-first” was the rational product call. OpenAI, Anthropic, and Google offered frontier model capability with fast integration — a product team could ship an AI feature in months, and investors rewarded that speed. The structural problem is [marginal cost](https://en.wikipedia.org/wiki/Marginal_cost): every user request is a direct line item to a third-party invoice. The more usage, the faster the burn. Canva hit that ceiling hard. Snapchat hit it earlier, quietly stopping recommendations of purely AI-generated video for the same reason.

Same architecture breaking at the same load.

##### AI Companies Are Building Their Own Chips

Anthropic confirmed in August 2026 that it is assembling an in-house silicon team, with the goal of “co-designing hardware and models” so Claude runs faster and more efficiently at the scale customers need. The company is in talks with Samsung for manufacturing — not TSMC. Job postings for chip engineers list salaries between $320,000 and $485,000 annually, requiring candidates who have “personally shipped silicon.”

OpenAI already launched its own chip “Jalapeño” in partnership with Broadcom in June 2026. Meta’s custom chip is expected to begin volume production. All six major AI players — OpenAI, Anthropic, Meta, Google, Amazon, Microsoft — are now on the custom silicon track. The unified direction: reduce dependence on NVIDIA and own the cost structure of inference.

Anthropic itself reached $30 billion in annualized revenue as of April 2026, while simultaneously being Google Cloud’s largest TPU customer, and planning to deploy multiple gigawatts of compute starting 2027. A company generating that kind of revenue is still renting compute externally while building its own chips. Future cost structure must be controlled internally.

##### Why Zhang Yiming Banned Distillation

ByteDance founder Zhang Yiming ordered employees to stop distilling competitor models — including those from OpenAI and Anthropic — while announcing what the company describes as the largest AI training run in China’s history, targeting the scale of Anthropic’s Mythos-tier frontier models. Distillation is the fast-follower’s shortcut: use a strong model’s outputs to train a cheaper one. Abandoning it means higher training costs and longer timelines, but also means the resulting model isn’t capped by whatever the original model could do. A distilled model’s ceiling is someone else’s floor.

The ban is also a posture: ByteDance is not positioning itself as a second-tier follower.

##### Vertical Integration as Survival Logic

From chip to model to application, outsourcing every layer means a cost leak at every layer, a capability ceiling at every layer, and pricing exposure to every upstream decision. Canva demonstrated what happens at the bottom of that stack: the better the product performs, the faster the losses accumulate.

NVIDIA announced partnerships with Apollo, BlackRock, Blackstone, and other major financial institutions to mobilize over $500 billion in AI compute infrastructure. Capital at that scale moving into compute means inference is no longer a consumption line item that software companies can linearly outsource — it’s [infrastructure](https://en.wikipedia.org/wiki/Infrastructure).

For any product team currently relying on third-party frontier APIs for core functionality, Canva’s one-third revenue cut is a leading indicator. Usage controls, the timeline for internal model development, the acceptable ceiling for API dependency — those are engineering decisions that have to be made now. The bill doesn’t announce itself before it arrives.

Anthropic builds chips. ByteDance trains from scratch. Canva revises its forecast down by a third. All three events happened within weeks of each other. Same structural force, three different points of impact.

— 邱柏宇