巴克莱关于人工智能计算和代理人工智能：虽然计算能力表面上似乎足够，但结构性挑...

network

Barclays on AI Compute and Agentic AI: While computing power appears sufficient on the surface, structural challenges remain.

1) Industry Inference on Capacity: By 2025, approximately 15.7 million AI accelerators (GPU/TPU/ASIC, etc.) will be online globally, with 40% (around 6.3 million) dedicated to inference. Among these, about half (3.1 million) will be specifically allocated for agent/chatbot services.

2) Supporting a Large User Base: Depending on the computational demands of different models, the existing computing capacity can support between 1.5 billion to 22 billion AI agents—sufficient to meet the needs of over 100 million white-collar workers in the U.S. and EU, as well as more than 1 billion enterprise software licenses.

3) Application of More Efficient Models: If more efficient models (such as DeepSeek R1) are used instead of cutting-edge but costly models (like OpenAI’s o1), industry capacity could increase by up to 15 times. Notably, companies like Salesforce have already been observed adopting open-source models (e.g., Mistral) instead of the most expensive proprietary models.

While computing power appears sufficient on the surface, the industry faces structural challenges.

The AI sector must shift from "meaningless benchmark testing" to the deployment of truly useful agent-based products.

----------------------------------------

On Agentic AI:

If agent products truly take off and prove highly valuable to both consumers and enterprise users, we may need:

1) Cheaper, smaller, yet equally high-performing foundational models (similar to DeepSeek);

2) Increased deployment of inference chips; and

3) Potential repurposing of installed training GPUs for inference.
由
翻译自 英语
巴克莱关于人工智能计算和代理人工智能：虽然计算能力表面上似乎足够，但结构性挑战仍然存在。

1）行业对容量的推断：到 2025 年，全球将有约 1570 万个 AI 加速器（GPU/TPU/ASIC 等）上线，其中 40%（约 630 万个）专用于推理。这其中，约一半（310 万个）将专门分配给代理/聊天机器人服务。

2）支撑庞大的用户群：根据不同模型的计算需求，现有的算力可以支撑15亿至220亿个AI代理，足够满足美国和欧盟1亿多白领人群，以及10多亿个企业软件授权的需求。

3）应用更高效的模型：如果使用更高效的模型（如 DeepSeek R1）代替尖端但昂贵的模型（如 OpenAI 的 o1），行业产能可以提高 15 倍。值得注意的是，已经观察到 Salesforce 等公司采用开源模型（例如 Mistral）代替最昂贵的专有模型。

虽然计算能力表面上看起来足够，但该行业面临着结构性挑战。

人工智能领域必须从“毫无意义的基准测试”转向部署真正有用的基于代理的产品。

----------------------------------------

关于 Agentic AI：

如果代理产品真正起飞，并证明对消费者和企业用户都具有极高的价值，我们可能需要：

1）更便宜、更小但性能同样高的基础模型（类似于DeepSeek）；

2）增加推理芯片的部署；

3）可能重新利用已安装的训练 GPU 进行推理。
翻译得准确吗？请提供反馈，以便我们加以改进：  
下午6:37 · 2025年3月27日
·
1万
查看

network

Barclays on AI Compute and Agentic AI: While computing power appears sufficient on the surface, structural challenges remain.1) Industry Inference on Capacity: By 2025, approximately 15.7 million AI accelerators (GPU/TPU/ASIC, etc.) will be online globally, with 40% (around 6.3 million) dedicated to inference. Among these, about half (3.1 million) will be specifically allocated for agent/chatbot services.2) Supporting a Large User Base: Depending on the computational demands of different models, the existing computing capacity can support between 1.5 billion to 22 billion AI agents—sufficient to meet the needs of over 100 million white-collar workers in the U.S. and EU, as well as more than 1 billion enterprise software licenses.3) Application of More Efficient Models: If more efficient models (such as DeepSeek R1) are used instead of cutting-edge but costly models (like OpenAI’s o1), industry capacity could increase by up to 15 times. Notably, companies like Salesforce have already been observed adopting open-source models (e.g., Mistral) instead of the most expensive proprietary models.While computing power appears sufficient on the surface, the industry faces structural challenges. The AI sector must shift from "meaningless benchmark testing" to the deployment of truly useful agent-based products. ----------------------------------------On Agentic AI: If agent products truly take off and prove highly valuable to both consumers and enterprise users, we may need:1) Cheaper, smaller, yet equally high-performing foundational models (similar to DeepSeek);2) Increased deployment of inference chips; and3) Potential repurposing of installed training GPUs for inference.

由翻译自 英语

巴克莱关于人工智能计算和代理人工智能：虽然计算能力表面上似乎足够，但结构性挑战仍然存在。1）行业对容量的推断：到 2025 年，全球将有约 1570 万个 AI 加速器（GPU/TPU/ASIC 等）上线，其中 40%（约 630 万个）专用于推理。这其中，约一半（310 万个）将专门分配给代理/聊天机器人服务。2）支撑庞大的用户群：根据不同模型的计算需求，现有的算力可以支撑15亿至220亿个AI代理，足够满足美国和欧盟1亿多白领人群，以及10多亿个企业软件授权的需求。3）应用更高效的模型：如果使用更高效的模型（如 DeepSeek R1）代替尖端但昂贵的模型（如 OpenAI 的 o1），行业产能可以提高 15 倍。值得注意的是，已经观察到 Salesforce 等公司采用开源模型（例如 Mistral）代替最昂贵的专有模型。虽然计算能力表面上看起来足够，但该行业面临着结构性挑战。人工智能领域必须从“毫无意义的基准测试”转向部署真正有用的基于代理的产品。----------------------------------------关于 Agentic AI：如果代理产品真正起飞，并证明对消费者和企业用户都具有极高的价值，我们可能需要：1）更便宜、更小但性能同样高的基础模型（类似于DeepSeek）；2）增加推理芯片的部署；3）可能重新利用已安装的训练 GPU 进行推理。

翻译得准确吗？请提供反馈，以便我们加以改进：
















下午6:37 · 2025年3月27日
·
1万
查看