【AI 大客戶的月帳單,一年半從 3 萬美元變成 222 萬】🧾
追蹤企業軟體支出的資料公司 YipitData 統計,Anthropic 前 1% 客戶的月支出中位數,2024 年 11 月還在 3 萬美元上下,2026 年 8 月來到 222 萬美元,平均值 283 萬美元。
中位數和平均值的距離很近,比絕對金額更有訊息。如果是一兩個巨型客戶把數字拉上去,平均值會遠高於中位數;222 萬對 283 萬,代表整排大帳戶一起長大。同一份資料顯示,年支出超過 100 萬美元的企業帳戶近期翻倍到 1,000 個以上。
資料來自 1,300 多家中大型企業組成的支出樣本,不是公司財報,也不含個人訂閱。所以它量的是企業把多少預算放進 AI,不是有多少人在用 AI。
企業軟體的收入通常靠大量中型客戶堆出來,跑掉一個不痛不癢。目前 AI 模型供應商的形狀比較接近雲端運算早期:少數用量極大的客戶決定整條營收曲線,而這些大客戶本身多半也是 AI 公司,有能力自己訓練模型或換供應商。
所以這張圖真正在說的是,黏著度還沒被測試過。帳單金額愈大,客戶自建或轉單的誘因也愈大,而這兩件事在圖上是同一條線。
資料來源: YipitData 企業支出樣本、PYMNTS (圖片來自 @zerohedge)
The Biggest AI Customers Went From $30,000 a Month to $2.2 Million
YipitData, which tracks enterprise software spending, puts the median monthly spend of Anthropic's top 1% of customers at around $30,000 in November 2024 and $2.22 million in August 2026. The average for that group is $2.83 million.
That the median and the average sit so close together carries more information than the headline number. If one or two enormous accounts were doing the work, the average would tower over the median. At $2.22 million against $2.83 million, a whole row of large accounts is growing at once. The same data shows enterprise accounts spending over $1 million a year have recently doubled to more than a thousand.
The figures come from a spend panel of more than 1,300 mid-market and enterprise companies rather than from company accounts, and they exclude consumer subscriptions. What they measure is how much corporate budget has moved into AI, not how many people are using it.
Enterprise software revenue is normally built from a large base of mid-sized customers, where losing any one of them barely registers. AI model providers currently look more like cloud computing in its early years, where a handful of very heavy users shape the entire revenue curve. Many of those heavy users are themselves AI companies, able to train their own models or switch suppliers.
So what the chart is really showing is that the stickiness test has not happened yet. The larger the bill, the stronger the incentive to build in-house or move elsewhere, and both of those live on the same line.
Source: YipitData enterprise spend panel and PYMNTS (Image: @zerohedge)