AI 模型公司按兆瓦批量采购电力，再将其转售为认知工作。
“算力的基础成本通常在每兆瓦 1000 万、1300 万或 1500 万美元左右。以 Anthropic 为例，其收入已高达每兆瓦 5000 万美元。这现在让他们能够做到的是——嘿，如果我在推理能力上花 10 美元，我实际上能产生 50 美元的收入。然后我可以转身把所有这些利润增量地投入到训练上。”
—— Dylan Patel，SemiAnalysis，出自 Dwarkesh Podcast¹
以电力为函数衡量的毛利润证明，模型公司可以在贡献毛利的基础上实现盈利。
Anthropic 2024 年的毛利率为 −94%：每 1 美元收入对应 1.94 美元的算力成本。²
到 2025 年，拐点到来，Anthropic 从 −94% 一举扭转为 40–50% 的毛利率。
2026 年，收入超过成本：每兆瓦 5000 万美元收入，对应 1000 万至 1500 万美元成本。Anthropic 录得首个盈利季度，收入 109 亿美元，营业利润 5.59 亿美元。
5% 的营业利润率远低于兆瓦数学所隐含的 70% 至 80% 毛利率。训练运行和人员规模吃掉了其中的差额。
每兆瓦毛利润不仅关乎智能水平，也关乎效率。一个以极小部分算力提供同等智能的模型，即使定价更低，也能在每兆瓦上产生更多利润。
GLM-5.3-Flash 在 Artificial Analysis Intelligence Index 上得分 57，与 Claude Opus 4.8 完全持平。³ 但它仅用 180 亿激活参数就做到了这一点，成本降低 90% 至 97%。更少的激活参数和更少的注意力计算意味着每兆瓦能产出更多 token。
前沿本身也在一波波地持续攀升。三次 3 分或以上的跃升，贡献了从 37 到 63 这 25 分涨幅的一半。高效模型追逐的是一个每个季度都会重置的目标。
这正是每兆瓦利润所买来的东西。Patel 的最后一句话是关键转折：“转身把所有这些利润增量地投入到训练上。” 利润率支撑着这座工厂。
Nvidia 为此支付 60 亿美元收购 Poolside，并追加投资 10 亿美元。模型构建正在变成一种工业化流程：在一个搜索空间里进行数千次实验，而非手工艺式的精雕细琢。
“模型是产出。真正意义上的创新，在于持续打造出更优模型的能力，且每一次都更快、更高效。模型工厂才是那个能不断复利的资产。”
—— 杰森·沃纳（Jason Warner），Poolside
Laguna S 2.1 从项目启动到正式发布，仅用了 52 天。
推理环节的利润，为工厂提供了资金，使得下一代模型的构建成本更低、运行效率更高。
-
迪伦·帕特尔（Dylan Patel）在 Dwarkesh 播客中提到，“到 2028 年，Anthropic 和 OpenAI 将拥有全球大部分算力”（2026 年 8 月）。Apple Podcasts ↩︎
-
The Information 报道，“Anthropic 的毛利率引发对 AI 长期盈利能力的质疑。” The Information ↩︎
-
Artificial Analysis 智能指数。Artificial Analysis ↩︎
-
杰森·沃纳（Jason Warner），Poolside。X ↩︎
An AI model company buys wholesale electricity by the megawatt & resells it as cognitive work.
“The base cost of compute tends to be around 10 or 13 or $15 million per megawatt. In the case of Anthropic, the revenue has gone as high as $50 million per megawatt. And what that now enables them to do is, hey, if I spend 10 bucks on inference capacity, I actually generate 50 bucks of revenue. And then I can turn around and incrementally spend all of that profit on training.”
— Dylan Patel, SemiAnalysis, on the Dwarkesh Podcast1
Gross profit measured as a function of electricity proves model companies can be profitable on a contribution basis.
Anthropic’s gross margin was −94% in 2024 : $1.94 of compute for every $1 of revenue.2
By 2025 the corner turned, & Anthropic swung from −94% to a 40-50% gross margin.
In 2026 revenue passed cost : $50m per megawatt against a $10-15m cost. Anthropic booked its first profitable quarter, $10.9b of revenue & $559m of operating profit.
That 5% operating margin sits well below the 70 to 80% gross margin the megawatt math implies. Training runs & headcount consume the difference.
Gross profit per megawatt is not just about intelligence, but also efficiency. A model that serves the same intelligence at a fraction of the compute generates more profit per megawatt, even at a lower price.
GLM-5.3-Flash scores 57 on the Artificial Analysis Intelligence Index : identical to Claude Opus 4.8.3 But it does it on 18 billion active parameters, at a 90 to 97% reduction in cost. Fewer active parameters & less attention compute means more tokens per megawatt.
The frontier itself keeps climbing in waves. Three jumps of 3 points or more carry half of the 25-point gain from 37 to 63. The efficient models chase a target that resets every quarter.
Which is what the profit per megawatt buys. Patel’s last clause is the hinge : “turn around and incrementally spend all of that profit on training.” The margin funds the factory.
Nvidia paid $6b to acquire Poolside & invested another $1b on that premise. Model building becomes an industrial process : thousands of experiments across a search space, not artisanal hand-tuning.
“The model is the output. The ability to keep building better models, faster & more efficiently each time is the actual innovation. The Model Factory is the compounding asset.”
— Jason Warner, Poolside4
Laguna S 2.1 went from kickoff to release in 52 days.
Inference margin funds the factory that makes the next model cheaper to build & more efficient to run.
-
Dylan Patel on the Dwarkesh Podcast, “Anthropic & OpenAI will have most of the world’s compute by 2028” (Aug 2026). Apple Podcasts ↩︎
-
The Information, “Anthropic’s Gross Margin Flags Long-Term AI Profit Questions.” The Information ↩︎
-
Artificial Analysis Intelligence Index. Artificial Analysis ↩︎