英伟达正计划进一步扩大软件业务，希望在销售人工智能芯片之外，通过企业软件许可获得更加稳定的经常性收入。英伟达首席执行官黄仁勋近日在德国柏林举行的GTC活动上谈及这一战略时表示，为运行人工智能代理而提供的软件，可以采用按GPU数量收费的授权模式，价格例如每块GPU每年约2000美元。这一数字引发了业界对英伟达软件业务潜在规模的关注，但目前还不能将其视为已经正式公布的统一收费标准。
据报道，黄仁勋于2026年10月22日在相关问答环节中谈到了软件授权业务。他表示，软件许可将日益成为英伟达业务的重要组成部分，尤其是为人工智能工厂中的计算设备提供运行人工智能代理所需的软件、管理工具和配套能力。
所谓人工智能代理，是指能够根据目标执行多步骤任务，并在授权范围内调用工具、访问数据或运行程序的人工智能系统。与传统聊天机器人相比，这类系统通常需要更复杂的运行环境、安全控制和持续监测机制。企业若要在数据中心大规模部署人工智能代理，不仅需要购买GPU，还需要解决软件部署、权限管理、运行监控和安全维护等问题。
英伟达正在试图将这些能力整合进更完整的企业级基础设施方案中。按照黄仁勋提出的示例价格，如果有500万块GPU采用每年2000美元的软件许可，理论上每年可产生100亿美元收入；如果连续五年维持这一规模，累计收入将达到500亿美元。不过，这只是根据假设计算出的情景，并非英伟达公布的收入预测。目前，公司尚未披露实际适用的GPU数量，也没有确认这一价格会覆盖哪些产品、何时实施以及具体合同条件。
值得注意的是，英伟达已经拥有成熟的企业软件订阅业务。其NVIDIA AI Enterprise平台面向企业数据中心，提供人工智能开发、部署和管理所需的软件组件。英伟达公布的许可价格显示，该平台的一年订阅费用为每块GPU 4500美元，包含相应支持服务；五年订阅价格为每块GPU 18000美元。
因此，黄仁勋提及的每年2000美元并不能简单理解为英伟达现有企业软件的降价。由于尚不清楚两种价格对应的软件范围、授权期限和服务内容是否相同，目前无法直接比较其实际成本。新价格可能针对不同的软件组合或特定的人工智能代理运行环境，具体情况仍有待英伟达进一步说明。
软件业务之所以受到英伟达重视，与其希望降低对单次硬件销售的依赖有关。GPU销售通常与数据中心建设周期、客户采购计划和芯片供应能力密切相关，而软件订阅则能够在客户持续使用相关产品期间形成重复收入。如果英伟达能够让企业长期依赖其软件平台来运行和管理人工智能工作负载，就有机会从已经部署的GPU中持续获得收入，而不仅仅是在新芯片售出时确认一次销售收入。
这一战略也与英伟达近期推进的人工智能代理安全架构有关。2026年10月21日，英伟达公布了OpenShell参考架构，旨在为企业运行人工智能代理提供更加完整的安全环境。该方案整合了英伟达的Vera Rubin计算基础设施、Vera CPU和BlueField-4数据处理器，并引入名为NVIDIA Sentry的监控机制。
OpenShell提供用于隔离人工智能代理的运行环境，可以限制代理能够访问的工具、数据和系统资源。NVIDIA Sentry则负责独立监控代理活动，并在发现代理试图越过既定权限边界时采取隔离措施。其设计思路是将部分安全控制放在底层基础设施中，而不是完全依赖人工智能模型自身的安全机制。
这一点对于企业部署人工智能代理尤其重要。代理可能被授权读取内部文件、访问业务系统，甚至执行具有实际后果的操作。如果权限配置不当，或者代理受到恶意输入诱导，就可能造成数据泄露、未经授权的操作或其他安全风险。英伟达希望通过运行环境隔离、权限限制和独立监控，为企业提供更容易管理的部署方案。
目前，其他企业也开始围绕这一架构开展合作。Anthropic正在将OpenShell和BlueField硬件与Claude Managed Agents进行整合，以便为企业代理提供额外的运行隔离和访问控制能力。Salesforce则开发了Slack Admin Bridge for OpenShell，让管理员能够直接在Slack中审核代理的访问请求、批准或拒绝权限，并监控相关活动。
这些合作显示，人工智能代理的竞争正在从模型性能逐步扩展到运行环境、安全管理和企业服务。随着代理开始承担更复杂的工作，企业不仅需要判断模型能否完成任务，也需要确保其在执行过程中遵守权限边界，并能够被持续监控和管理。英伟达希望将GPU、网络设备、软件平台和安全机制整合为一套完整方案，从而提高客户对其技术生态的依赖程度。
黄仁勋此前也多次强调，英伟达不应只被视为一家芯片制造商。公司的业务已经扩展到网络设备、整机系统、企业软件，以及医疗健康、汽车、通信和生命科学等领域。通过提供从计算硬件到软件管理的完整技术体系，英伟达希望在人工智能基础设施的多个环节中获得收入。
不过，软件授权模式能否成为英伟达新的重要收入来源，仍取决于企业客户的接受程度。部分客户可能愿意为稳定运行、安全保障和技术支持支付持续费用；另一些客户则可能选择开源软件、自行开发管理工具，或通过其他供应商提供的方案满足需求。按GPU数量收费也意味着，随着企业部署规模扩大，软件成本可能同步上升，因此客户会更加关注许可范围、合同期限以及是否存在重复付费。
目前，英伟达尚未正式公布这项每块GPU每年约2000美元的授权方案的完整条款，也没有给出具体实施日期。现阶段，这一数字更适合作为黄仁勋阐述软件业务战略时提出的价格示例，而不是已经全面生效的商业报价。
从长期来看，英伟达正在尝试建立一种新的收入结构：硬件销售帮助客户建设人工智能基础设施，软件许可则在基础设施投入使用后持续产生收入，而安全和管理工具进一步提高企业长期使用其平台的意愿。如果这一模式成功，英伟达未来的增长将不再完全取决于每年能够销售多少块GPU，也将越来越取决于客户愿意为运行在这些GPU上的软件和服务支付多少费用。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
In Weida is planning to further expand the software business in the hope of obtaining more stable recurrent revenue through enterprise software licences, in addition to the sale of artificial intelligence chips. Commenting on this strategy at the recent GTC event in Berlin, Germany, CEO Hwang In-hoon indicated that software provided for the operation of artificial intelligence agents could be based on an authorized model based on the volume of GPUs at a price of about US$ 2,000 per GPU per year. This figure raises industry concerns about the potential size of the software operation in Weida, but it cannot be considered a formally published flat rate at this time.
On 22 October 2026, Huang In-hoon reportedly addressed the software authorization business during the relevant question and answer session. He stated that software licences would increasingly become an important part of the British business, particularly in providing the software, management tools and supporting capabilities needed to run artificial intelligence agents for computing devices in artificial intelligence plants.
Artificial intelligence agents are artificial intelligence systems capable of carrying out multi-step tasks according to the target and of accessing tools, access to data or running programs within the mandate. Such systems typically require more complex operating environments, security controls and continuous monitoring mechanisms than traditional chat robots. The large-scale deployment of artificial intelligence agents by the enterprise in the data centre requires not only the purchase of GPUs but also the resolution of software deployment, permission management, operation monitoring and security maintenance.
Weeda is trying to integrate these capabilities into a more complete enterprise-level infrastructure programme. According to the illustrative price offered by Huang In-hoon, if 5 million GPUs were licensed with $2,000 a year, it would theoretically generate $10 billion a year; if this size were maintained for five consecutive years, the cumulative income would be $50 billion. However, this is only a scenario based on assumptions and not a revenue projection published by Inverda. At present, the company has not disclosed the quantities of GPUs actually applied, nor has it confirmed which products would be covered by this price, when they would be implemented and the specific terms of the contract.
It is worth noting that Ingweida already has mature corporate software subscriptions. Its NVIDIA AI Enterprise platform is oriented towards enterprise data centres and provides software components required for artificial intelligence development, deployment and management. The licence price published by Yin Weida shows that the one-year subscription fee for the platform is $4,500 per GPU, including corresponding support services; the five-year subscription price is $18,000 per GPU.
Thus, the US$ 2,000 per year mentioned by Hoang In-hoon cannot simply be understood as a reduction in the cost of software for the existing enterprise in Weida. As it is not clear whether the scope of the software, the duration of the authorizations and the content of the services are the same for the two prices, it is not possible to compare its actual costs directly. The new price may be specific to different software packages or specific AI environments, the details of which are yet to be further clarified by the UK.
The focus on software operations is linked to the desire to reduce reliance on single hardware sales. GPU sales are usually closely related to the data centre construction cycle, customer procurement plans and chip supply capacity, while software subscriptions generate duplicate income during the continuous use of the relevant products by clients. If Ying Weida were to allow businesses to rely on their software platforms for long periods of time to operate and manage artificially intelligent work loads, there would be an opportunity to obtain sustained revenue from deployed GPUs, not just to recognize a sales income when new chips are sold.
This strategy is also linked to the recently promoted artificial intelligence agency security architecture in Weida. On 21 October 2026, Inverda published the OpenShell reference architecture, which aims to provide a more complete security environment for the operation of artificial intelligence agents in enterprises. The programme integrates Vera Rubin computing infrastructure, Vera CPU and Bluefield-4 data-processing devices in Wevda and introduces a monitoring mechanism called NVIDIA Stanley.
OpenShell provides an operating environment that isolates artificial intelligence agents and limits the tools, data and system resources that agents can access. NVIDIA Centry is responsible for independent monitoring of agency activities and for taking isolation measures when agents are found attempting to cross the established permission border. It is designed to place part of the security control in the base infrastructure, rather than relying entirely on artificial intelligence models for their own safety mechanisms.
This is particularly important for enterprises to deploy artificial intelligence agents. Agents may be authorized to read internal documents, access business systems or even perform operations with practical consequences. If permission is not properly configured or the agent is induced by malicious input, there may be data leakage, unauthorized operations or other security risks. Weeda hopes to provide enterprises with more manageable deployment programmes through operational isolation, permission restrictions and independent monitoring.
Other enterprises are now working together around this structure. Anthropic is integrating OpenShell and Bluefield hardware with Claude Managed Agencies to provide additional operational isolation and access control capabilities for business agents. For its part, Salesforce developed Slack Admin Bridge for OpenShell to enable managers to review agency requests for access, approve or reject permission directly in Slack and monitor related activities.
These collaborations show that competition for artificial intelligence agents is gradually expanding from model performance to operational environments, safety management and business services. As agents begin to undertake more complex tasks, enterprises need not only to judge whether the model is capable of accomplishing its tasks, but also to ensure that it complies with the permission boundary in its implementation and is able to be continuously monitored and managed. It is the hope of the United Kingdom that the GPU, network equipment, software platforms and security mechanisms will be integrated into a complete package, thereby increasing client dependence on its technological ecology.
Wong In-hoon has also stressed on previous occasions that Wei Da should not be regarded as a chip manufacturer alone. The company ' s operations have been extended to areas such as network equipment, machine systems, enterprise software, medical health, automobiles, communications and life sciences. By providing a complete technology system, from computing hardware to software management, YVD wants to earn income in multiple segments of the AI infrastructure.
However, whether or not the software authorization model is an important new source of revenue for Britain will depend on the degree of acceptance of the enterprise ' s clients. Some clients may be willing to pay ongoing costs for stable operations, safety and security, and technical support; others may choose open source software, develop their own management tools, or meet demand through programmes provided by other suppliers. A GPU charge also means that as the enterprise deploys on a larger scale, software costs may rise simultaneously, so that clients pay more attention to the scope of the licence, the duration of the contract and the existence of duplicate paid.
At present, the full terms of the mandated programme of approximately US$ 2,000 per GPU per year have not been officially published in Weida, nor has a specific implementation date been given. At this stage, this figure is more appropriate as an example of the price presented by Wong In-hoon in elaborating the business strategy for software, rather than a commercial offer that has become fully effective.
In the long run, Weida is trying to create a new revenue structure: hardware sales help clients build artificial intelligence infrastructure, software licences generate sustained revenue after the infrastructure is put into operation, and security and management tools further increase the willingness of enterprises to use their platforms on a permanent basis. If this model succeeds, the future growth of Britain will no longer depend entirely on how many GPUs can be sold each year, and increasingly on how much customers are willing to pay for running the software and services on these GPUs.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.