中国开源权重人工智能模型正在全球范围内迅速崛起，但一项新研究表明，这些模型产生的大部分收入流向了第三方提供商，而非模型的开发公司。
China’s open-weight artificial intelligence models are taking off globally, but much of the money they generate is flowing to third-party providers rather than the companies that create them, according to a new study.
包括深度求索（DeepSeek）和月之暗面（Moonshot AI）在内的开发者，将其模型权重——即人工智能系统的“大脑”——免费开放供下载。这一策略在过去两年中加速了全球范围内的采用，使其他公司能够在自己的服务器上运行这些模型，并独立进行修改。
Developers including DeepSeek and Moonshot AI make their model weights – essentially the “brains” of their AI systems – freely available for download. This strategy, which has accelerated global adoption over the past two years, allows other companies to run the models on their own servers and modify them independently.
尽管这种方法提升了采用率，但也使开发者面临激烈的竞争，因为竞争对手平台将开源模型托管在自己的基础设施上，并以更低的价格提供访问服务。相比之下，美国的OpenAI和Anthropic等公司则保持其模型封闭，并对访问权限和定价进行严格控制。
While the approach has boosted adoption, it has also exposed developers to intense competition as rival platforms host the open models on their own infrastructure and offer access at lower prices. By contrast, companies such as OpenAI and Anthropic in the United States keep their models closed and maintain tight control over access and pricing.
例如，总部位于杭州的初创公司深度求索，根据美国研究机构Epoch AI周二发布的一份报告，仅获得了通过第三方聚合平台OpenRouter访问其模型的客户支出的约8%。
Hangzhou-based start-up DeepSeek, for example, captured only about 8 per cent of estimated customer spending on its models accessed through third-party aggregator OpenRouter, according to a report released on Tuesday by US research institute Epoch AI.
总部位于北京的智谱AI（Z.ai，又称Zhipu AI）也面临类似挑战。在智谱AI于8月发布其GLM-5.3 Flash模型权重后的一天内，就有12家第三方提供商通过OpenRouter提供该模型的访问服务。二十天后，这一数字上升至26家，其中八家提供商的收费低于智谱AI。
Beijing-based Z.ai, also known as Zhipu AI, faces a similar challenge. Within a day of the lab releasing the weights for its GLM-5.3 Flash model in August, 12 third-party providers were offering access to it through OpenRouter. Twenty days later, the number had risen to 26, with eight providers charging less than Z.ai.
在同一时期，智谱AI通过OpenRouter提供的GLM-5.3 Flash模型令牌（tokens）份额从88%骤降至22%。Epoch AI报告称，尽管该模型的整体使用量增长了17%，但智谱AI的每日令牌处理量却下降了70%。
Over the same period, Z.ai’s share of GLM-5.3 Flash tokens served through OpenRouter plunged from 88 per cent to 22 per cent. Its daily token volume dropped 70 per cent even as overall use of the model rose 17 per cent, Epoch AI reported.
这些数据表明，中国开发者在试图从直接访问其模型中获取收入时，面临着显著的不利局面。
The figures suggest Chinese developers face a significant disadvantage when attempting to capture revenue from direct access to their models.
Z.ai报告称，其应用编程接口平台在上半年的毛利率为24.6%，较2025年的18.9%有所上升。该平台允许客户在其自有应用程序和工具中使用Z.ai的AI模型。
Z.ai reported a gross margin of 24.6 per cent for its application programming interface platform in the first half of this year, up from 18.9 per cent in 2025. The platform lets customers use Z.ai’s AI models in their own applications and tools.
相比之下，Epoch AI估计，美国竞争对手Anthropic通过仅向其自家API提供封闭权重的Opus 4.8模型，实现了85%的毛利率。
By comparison, Epoch AI estimated that US rival Anthropic achieved a gross margin of 85 per cent by offering its closed-weight Opus 4.8 model exclusively through its own API.
DeepSeek在开放权重开发者中被称为“异常值”，据报告称，该公司因推理成本低廉，估计毛利率在70%至80%之间。尽管如此，第三方主机仍然占据了DeepSeek模型的大部分终端用户需求。
DeepSeek was an “outlier” among open-weight developers, with estimated gross margins of between 70 and 80 per cent because of the company’s low inference costs, according to the report. Even so, third-party hosts still captured much of the end-user demand for DeepSeek’s models.
与此同时，美国公司通过其封闭模型的高价订阅计划，直接从用户那里获取更多收入。
US firms, meanwhile, have been able to capture more revenue directly from users through lucrative subscription plans for their closed models.
例如，据星期一由安德烈森霍洛维茨投资公司发布的报告称，Anthropic“积极”地对用户进行了商业化变现。该报告称，约有7.3%的Claude付费用户使用该公司的最贵个人套餐，该套餐起始价为100美元/月。美中AI开发者之间的商业化变现差距仍然很大。
Anthropic, for example, had “aggressively” monetised users, according to a report on Monday by venture capital firm Andreessen Horowitz. About 7.3 per cent of Claude’s payingrs were on the company’s most expensive individual plan, which starts at US$100 a month, the report said. The monetisation gap between US and Chinese AI developers remains wide.
Epoch AI估计，截至9月，中国六大AI开发公司——字节跳动、阿里巴巴集团、Z.ai、月之暗面、DeepSeek和MiniMax——共同产生的AI收入约为OpenAI与Anthropic合计的10%。阿里巴巴拥有《南华早报》。
China’s six leading AI developers – ByteDance, Alibaba Group Holding, Z.ai, Moonshot AI, DeepSeek and MiniMax – collectively generated about 10 per cent as much AI revenue as OpenAI and Anthropic combined as of September, Epoch AI estimated. Alibaba owns the South China Morning Post.
中国开发者正在试验新的许可协议和商业伙伴关系，这些协议可能使他们能够获取更多第三方使用其模型所产生的收入。
Chinese developers are experimenting with new licensing agreements and commercial partnerships that could allow them to capture more of the revenue generated by third parties using their models.
今年7月，北京的月之暗面以一种许可协议发布了其Kimi K3模型，该协议要求一些将该模型作为服务提供的大型供应商签订收入分成协议。
In July, Beijing-based Moonshot AI released its Kimi K3 model under a licence requiring some large providers that offer the model as a service to enter revenue-sharing agreements.
据路透社八月份报道，月之暗面（Moonshot）还曾寻求获取主要云平台上与Kimi K3相关服务所产生收入的最高30%。Kimi K3已于上个月在亚马逊云平台上架。
Moonshot had also sought up to 30 per cent of the revenue generated by Kimi K3-related services on major cloud platforms, including Amazon Web Services, Microsoft Azure and Google Cloud, Reuters reported in August. Kimi K3 was launched on Amazon last month.
智谱AI（Z.ai）正采取类似策略。周一，亚马逊将GLM-5.3模型加入其Bedrock平台，使企业能够通过这家美国云巨头的服务器访问该模型。中国科技媒体《星市场日报》报道称，该协议包含与智谱AI按使用量进行收入分成的条款。
Z.ai is adopting a similar strategy. On Monday, Amazon added GLM-5.3 to its Bedrock platform, giving businesses access to it through the US cloud giant’s servers. Chinese tech media outlet Star Market Daily reported that the arrangement included usage-based revenue sharing with Z.ai.
Epoch AI在报告中指出，此类协议可能为中国AI开发者提供一条路径，使其既能从开放模型的分布优势中获益，又能保留其产生收入的更大份额，但目前尚难判断这些新安排能带来多少收入。
Such agreements could give Chinese AI developers a way to benefit from the distribution advantages of open models while retaining a greater share of the revenue they generate, Epoch AI said in the report, though it remains too early to judge how much revenue the new arrangements could produce.