中国AI初创公司DeepSeek预测，如果所有用户都付费，其AI模型的成本利润率将达到545%。该公司在GitHub帖子中表示，其V3和R1模型在2月底的24小时内，每天的运行成本为87,072美元。如果按DeepSeek-R1的定价计费，每日收入将达到562,027美元。然而，目前只有部分服务实现了货币化。DeepSeek还应用了非高峰时段定价，这进一步降低了实际收入。尽管如此，这是该公司首次披露有关推理任务财务细节，引发了关于AI基础设施支出的讨论。

💰DeepSeek声称，在特定情况下，其在线服务的成本利润率可达545%，但这一数字基于所有用户都付费的前提，而目前只有部分服务实现了货币化。

⏱️DeepSeek通过在白天高峰时段将Nvidia H800芯片用于推理任务，并在夜间低需求时段重新分配用于研究和训练，从而降低成本，并为开发者提供夜间折扣。

🤔DeepSeek以较低成本推出强大AI模型，引发硅谷震动，并导致科技股抛售。该公司声称，其模型训练成本低于600万美元，引发了关于数十亿美元AI基础设施必要性的质疑。

🎤Google DeepMind CEO Demis Hassabis 认为DeepSeek的低成本模型被“夸大”了， 并表示DeepSeek声称仅以560万美元训练其AI模型可能仅占最终训练阶段的成本，而忽略了对开发和基础设施的更广泛投资。

DeepSeek said it would have a 545% cost-profit margin — under very specific circumstances.Perhaps the biggest disclaimer: it assumes everyone who uses its largely free AI models would pay.The Chinese startup said in a Saturday GitHub post that only a "subset" of its services are monetized.

Chinese AI startup DeepSeek has projected it would have a cost-profit margin of 545% for its AI models — provided everyone currently using them would pay.

In a GitHub post on Saturday, DeepSeek said its largely free V3 and R1 models had a daily running cost of $87,072 over a 24-hour period at the end of February, based on leasing Nvidia chips at $2 per hour.

If usage over that period was billed at DeepSeek-R1's pricing, it would generate daily revenue of $562,027, with a "cost profit margin of 545%," the post said.

Under those parameters, it would imply annual revenue exceeding $200 million.

However, DeepSeek said just a "subset" of its services are currently monetized, meaning its theoretical figures would require all users to switch to a paid plan — something that none of its rivals has achieved.

The assumption that everyone would pay isn't the only part of the post that brings the figure into doubt.

The startup said the cost-profit figure only applies R1 billing pricing and does not factor in the "significantly lower" V3 pricing.

Notably, DeepSeek also applies off-peak pricing, which adds a further asterisk to its theoretical financials.

During peak daytime hours, its Nvidia H800 chips focus on inference, which is when AI models conduct tasks. During lower-demand nighttime hours, it reallocates those chips to focus on research and training.

Because of this cost-saving mechanism, DeepSeek said it automatically applied nighttime discounts for developers, which would also bring down its actual revenue.

Despite the disclaimers, it's the first time the Hangzhou-based company has disclosed details about its financials from inference tasks.

It provides fresh insight into DeepSeek more than a month after it launched a powerful AI model at a seemingly lower cost, shocking Silicon Valley and causing a sell-off in tech stocks.

The market rout was partly fueled by the company's claim that it trained its models using less than $6 million worth of Nvidia H800 chips — a fraction of the spending by other US AI giants like OpenAI.

Since the H800 chips are less powerful than the chips used by many tech giants, DeepSeek's success raised further questions about the necessity of billion-dollar AI infrastructure expenditures.

However, DeepSeek's cost claims have drawn scrutiny from tech giants, researchers, and regulatory authorities.

Last month, Google DeepMind CEO Demis Hassabis called the buzz around DeepSeek's low-cost model "exaggerated." He said that DeepSeek's claim of training its AI model for just $5.6 million likely only accounts for the final training phase, overlooking the broader investment in development and infrastructure.