白宫在 Nvidia 加入会谈之际，将前沿人工智能审查规则保密

Why it matters

The framework could decide which labs, infrastructure providers and foreign partners see powerful models early. Keeping its rules private gives participants clarity and access that excluded developers cannot evaluate or challenge.

Jensen Huang's Nvidia had staff inside the White House's August 4th industry meetings on frontier AI security, while developers excluded from the talks were left without the framework that could determine how the government reviews powerful models before release.

白宫在8月4日就前沿人工智能安全举行的业界会议中有来自 Jensen Huang 的 Nvidia 员工参与，而被排除在这些会谈之外的开发者则无法获得可能决定政府在发布前如何审查强大模型的框架。

The White House does not plan to publish the completed framework, three people familiar with the discussions told Axios. Its contents will instead be shared with companies participating in the voluntary process.

三位知情人士告诉 Axios，白宫并不打算公布已完成的框架。其内容将转而与参与该自愿流程的公司共享。

That decision turns a national AI security policy into a private operating agreement between the government and a selected group of technology companies. The classified material includes the cyber-capability benchmark and the threshold that determines which systems qualify as "covered frontier models." The White House is also withholding the surrounding framework, which the executive order itself does not designate as classified.

这一决定把国家层面的人工智能安全政策转变为政府与一小部分被选定科技公司之间的私人运作协议。列为机密的材料包括网络能力基准和决定哪些系统符合“covered frontier models”门槛的标准。白宫同时也在扣留围绕这些内容的框架，尽管行政命令本身并未将这些框架指定为机密。

The result is a two-tier policy process. Participating developers can learn how to approach the government, protect model intellectual property during testing and help select "trusted partners" for early access. Developers outside the room cannot assess how the same process may apply to their next model.

其结果是形成了双层政策流程。参与的开发者可以了解如何与政府接触、在测试期间保护模型知识产权，并协助甄选可获得提前访问的“trusted partners”。未在会议室内的开发者则无法评估同一流程将如何适用于他们的下一个模型。

A voluntary review with consequential access

自愿审查与具有重要后果的访问

President Donald Trump signed Executive Order 14409 on June 2nd. It directed federal agencies to build the framework within 60 days, putting the deadline around August 1st. The White House said on August 3rd that it had finished the work on time, according to earlier Axios reporting.

总统 Donald Trump 在6月2日签署了 Executive Order 14409。该命令指示联邦机构在60天内制定该框架，截止日期约为8月1日。白宫在8月3日表示已按时完成该工作，依据为 早先的 Axios 报道。

The order allows developers to ask whether a model under development falls within the government's classified frontier threshold. A developer can then provide federal officials with access for as long as 30 days before releasing that model to other trusted partners.

该命令允许开发者询问其正在开发的模型是否落入政府划定的机密前沿门槛。开发者随后可以在将该模型发布给其他 trusted partners 之前，向联邦官员提供最长30天的访问权限。

The government must protect confidentiality, cybersecurity, intellectual property and model access during that window. Developers also collaborate with federal officials on deciding which trusted partners may receive the model early.

在此期间，政府必须保护保密性、网络安全、知识产权和模型访问权限。开发者还将与联邦官员合作决定哪些 trusted partners 可提前获得模型。

The order expressly bars the framework from becoming a mandatory licensing, preclearance or permitting system for AI development and release. Participation remains voluntary under the text Trump signed.

该命令明确禁止将该框架变成针对 AI 开发与发布的强制许可、预先审查或准入许可制度。根据 Trump 签署的文本，参与仍为自愿。

Voluntary status does not make the process commercially trivial. Early access to a capable model can give cybersecurity agencies and critical-infrastructure operators time to test defenses before broader distribution. It can also give selected companies, researchers or foreign governments an earlier view of capabilities that rivals and excluded partners have not seen.

自愿性质并不意味着这一流程在商业上无足轻重。对具备能力的模型的提前访问可以让网络安全机构和关键基础设施运营者在更广泛分发前有时间测试防御。它也能让被选中的公司、研究人员或外国政府比竞争对手和被排除的合作伙伴更早了解模型能力。

The unanswered operational question is who qualifies for that access. Axios reported that the framework does not publicly identify the trusted partners or establish whether foreign governments can join the group. The European Union and United Kingdom therefore lack a public basis for determining whether their agencies or researchers could participate.

尚未得到回答的操作层面问题是：谁有资格获得这种访问权限。Axios 报道称，该框架并未公开识别 trusted partners，也未明确外国政府是否可以加入该群体。因此，European Union 和 United Kingdom 无从在公开层面判断其机构或研究人员是否可以参与。

Nvidia brings the open-model argument into the room

Nvidia 将开放模型的论点带入会场

Axios reported that open-source models were discussed during the August 4th meetings and that Nvidia staff participated. The reporting does not establish what Nvidia proposed or whether the White House adopted any of its positions.

Axios 报道称，在8月4日的会议上讨论了开源模型，并且有 Nvidia 员工参与。该报道并未说明 Nvidia 提出了何种建议，也未表明白宫是否采纳了其任何立场。

Nvidia has a direct interest in how Washington treats open models. At its March conference, Nvidia argued that open and proprietary systems would develop together, quoting Huang as saying, "It's proprietary and open." Open models widen the pool of developers that can train, modify and deploy AI software on Nvidia's hardware and software stack.

Nvidia 对美国政府如何对待开放模型具有直接利害关系。在其3月的大会上，Nvidia 主张开放与专有系统会共同发展，并引用 Huang 的话称，“It's proprietary and open.” 开放模型扩大了能够在 Nvidia 硬件和软件堆栈上训练、修改和部署 AI 软件的开发者池。

Huang, who co-founded Nvidia in 1993 after working at AMD and LSI Logic, has spent his entire tenure pushing Nvidia into the next large computing market. That strategy turned a graphics-chip maker into a central supplier for model training and inference. Nvidia reported $215.938 billion in fiscal 2026 revenue, including $193.737 billion from data-center products.

Huang 在加入 AMD 和 LSI Logic 工作后，于1993年共同创立了 Nvidia，他整个任期都在推动 Nvidia 进入下一个大型计算市场。这一策略将一家图形芯片制造商转变为模型训练与推理的核心供应商。Nvidia 报告称其 2026 财年收入为 $215.938 billion，其中包括 来自数据中心产品的 $193.737 billion。

Huang met President Trump and Commerce Secretary Howard Lutnick (@howardlutnick) in Washington during the week before the framework meeting, Axios reported. The timing places Nvidia close to the administration as officials confront the tension between distributing American AI technology widely and limiting access to models with advanced cyber capabilities.

Axios 报道称，Huang 在框架会议前的一周于华盛顿会见了总统 Donald Trump 和商务部长 Howard Lutnick (@howardlutnick)。这一时间点使 Nvidia 在官员们在广泛分发美国 AI 技术与限制对具备高级网络能力模型的访问之间权衡时，处于接近政府的地位。

Nvidia's presence also reflects how far Huang's influence extends beyond selling chips. RuntimeWire reported in July that Nvidia invested in Safe Superintelligence and provided access to Vera Rubin compute. Separate reporting indicated Nvidia was considering a financial backstop tied to an OpenAI data-center lease. Policy participation adds another channel through which Nvidia can shape the conditions surrounding frontier AI development, though attendance at the White House meetings alone does not prove policy control.

Nvidia 的出现也反映出 Huang 的影响力远超单纯卖芯片。RuntimeWire 在7月报道称 Nvidia 投资了 Safe Superintelligence 并提供对 Vera Rubin 计算资源的访问。另有报道表明 Nvidia 正在 考虑与 OpenAI 数据中心租赁相关的财政后备措施。参与政策讨论为 Nvidia 提供了另一个影响前沿 AI 发展条件的渠道，尽管仅凭出席白宫会议并不能证明其掌握政策主导权。

Secrecy may weaken the consistency developers requested

保密性可能削弱开发者所要求的一致性

Frontier labs have publicly asked for predictable federal testing. In a July 15th policy post, OpenAI said the cyber-evaluation process needed to be consistent and repeatable across state and federal systems. OpenAI also said it was discussing the framework with the administration, peer companies and business groups.

前沿实验室公开要求联邦测试具有可预测性。在7月15日的一篇政策文章中，OpenAI 表示，网络评估流程需要在州和联邦系统之间保持一致且可重复。OpenAI 还表示正与政府、同行公司及商业团体讨论该框架。

Axios previously identified OpenAI, Anthropic and Google as labs that supplied feedback on a draft. The White House said it was engaging with additional industry partners, without publishing a participant list.

Axios 早前认定 OpenAI、Anthropic 和 Google 是向草案提供反馈的实验室。白宫表示正在与更多行业合作伙伴接触，但没有公布参与者名单。

A private framework can still produce consistent treatment among participating developers. It cannot demonstrate that consistency to companies outside the process. Those developers will not know whether comparable models receive comparable scrutiny, how intellectual-property protections work in practice or whether access decisions favor established labs with existing government relationships.

一个私有框架仍可能在参与的开发者之间产生一致的对待。但它无法向流程之外的公司证明这种一致性。那些开发者将不会知道可比模型是否接受了同等程度的审查、知识产权保护在实践中如何运作，或访问决定是否偏向于与政府已有关系的既有实验室。

Researchers and policymakers face a related problem. Keeping dangerous benchmark details classified can prevent adversaries from using the government's tests as a capability roadmap. Withholding the administrative rules surrounding those tests also blocks outside examination of who receives access, how disputes are handled and whether the voluntary process operates as written.

研究人员和政策制定者也面临相关问题。将危险的基准细节列为机密可以防止对手将政府的测试当作能力路线图来利用。但对这些测试的管理规则进行保密也阻碍了外界对谁获得访问、争议如何处理以及自愿流程是否按文本运行的审查。

The White House has chosen confidentiality as the price of getting model developers and national-security agencies into the same review process. The framework now depends on a smaller circle of companies trusting the government with unreleased models, while everyone outside that circle must trust a system they cannot inspect.

白宫选择以保密性作为代价，使模型开发者与国家安全机构进入同一审查流程。该框架现在依赖于一小圈公司信任政府交付未发布的模型，而圈外的所有人则必须信任一个他们无法检查的制度。