Build AI in America

闪电行动

为了让美国在人工智能领域引领全球，我们必须在计算能力和电力方面进行大量投资，使得在美国本土构建人工智能成为可能。随着技术的发展，所需的资源、基础设施和监管环境也需要相应调整。因此，我们发布了一系列政策建议，支持建设必要的国内基础设施，以在美国训练世界上最先进的人工智能模型，并推动人工智能在各个行业的广泛应用。

由美国领先企业推动的人工智能进步正在重塑我们的经济和社会。在Anthropic，我们看到我们的人工智能模型每天被美国企业、科学家、教育者、政府机构等广泛使用，这些应用提升了员工的生产力，促进了经济繁荣，并增强了国家安全。

现在采取果断行动，我们可以确保未来在美国本土构建。

我们的完整建议列表见新发布的报告《在美国构建人工智能》。以下是内容摘要。

全球人工智能领导力的能源需求

前沿人工智能模型训练需要持续稳定的电力供应，实现这一目标需要美国在多种能源技术上具备卓越的能源产能。用于开发最先进模型的数据中心需要稳定的电力才能高效运行，鉴于人工智能开发的能源需求，必须采取“多管齐下”的策略，确保美国在人工智能领域的领导地位。该策略将帮助美国满足即时能源需求，扩大稳定电力来源，并最大化人工智能推动新兴能源技术（如下一代地热和先进核能技术）的机会。

尽管美国在能源发电设施建设方面落后，但中国正在快速建设人工智能能源基础设施，去年新增了超过400吉瓦的电力容量，而美国仅增加了几十吉瓦。考虑到人工智能开发的需求，这种差距尤其令人担忧：我们预计Anthropic在2027年和2028年分别需要2GW和5GW的数据中心来训练单个先进人工智能模型。随着其他美国公司也在开发前沿人工智能模型，预计到2028年，美国前沿人工智能训练的总需求将达到20-25GW。作为对比，这大约是纽约市峰值电力需求的两倍，且日常人工智能模型推理（inference）所需的电力至少相当。换句话说，到2028年，美国人工智能行业至少需要50GW的电力容量，以保持全球人工智能领导地位。

尽管特朗普政府已采取重要措施，如设定雄心勃勃的核电目标和加快《国家环境政策法》（NEPA）审查，但为了与中国竞争，美国必须进一步采取行动，解决可能延误能源项目的监管挑战。

美国可以通过加快许可审批、支持重点输电项目以及与公用事业合作加快电网接入流程，在战略地点建设大规模人工智能训练基础设施。除了训练基础设施，我们还提出更广泛的改革建议，以支持全国范围内的人工智能部署，加强供应链和劳动力发展。这些建议包括加快地热、天然气和核能许可审批，制定美国陆军工程兵团的“全国许可”以加速《清洁水法》审查，建立关键电网组件和燃气轮机的战略储备，扩大技术培训项目。通过我们的提案，政策制定者还有机会通过监管权力和增加整体供应，保持电价对用户的低廉。

美国人工智能基础设施的两大支柱

我们提出两大战略支柱，涵盖从大规模训练设施到支持全国人工智能部署的基础设施的全方位挑战。

支柱一：建设大规模人工智能训练基础设施

• 开放联邦土地作为人工智能基础设施建设选址，避免漫长的州和地方分区审批流程；
• 加快人工智能基础设施的NEPA审批，包括对某些项目进行提前审查；
• 与私营部门合作，加快关键输电线路的建设和升级；
• 支持公用事业进行关键改革，加速接入流程，包括利用人工智能提升效率。在国家安全关键领域，利用现有联邦权限确保人工智能训练基础设施及时接入电网。

支柱二：建设支持全国人工智能创新的广泛基础设施

• 加快地热、天然气和核能许可审批；
• 设立国家利益电力输电走廊，加快人工智能数据中心增长区域的输电许可；
• 通过贷款和贷款担保项目加强关键电网组件和燃气轮机的国内生产，并建立这些产品的战略储备；
• 支持关键能源工人、电工和建筑工人的培训、学徒和创业项目。

美国具备经济实力、技术能力和创新精神，能够应对人工智能基础设施的挑战。我们的建议提供了一个切实可行的蓝图，利用现有联邦权限，打破长期阻碍能源发展的监管壁垒。

通过政府、产业和社区的合作，我们可以为美国未来数十年的人工智能领导地位奠定坚实基础。

阅读全文报告
[https://www-cdn.anthropic.com/13e4cfd61ee168c0372256d7de83d57c9df7c6d1.pdf]

For the United States to lead the world in AI, we must make substantial investments in computing power and electricity that make it possible to build AI in America. As technology evolves, the required resources, infrastructure, and regulatory conditions need to change too. That’s why we are releasing a set of policy recommendations to support building the necessary domestic infrastructure to train the world’s most capable AI models in the U.S. and support the broad deployment of AI across sectors. 

Advances in AI, driven by cutting-edge American companies, are reshaping our economy and society. At Anthropic, we see our AI models used every day by American businesses, scientists, educators, governments, and others—in ways that make workers more productive, our economy more prosperous, and our nation more secure.

By acting decisively now we can ensure that the future is built here in America.

Our full list of recommendations can be found in our new report, “Build AI in America.” Below, we summarize its contents.

Frontier AI model training requires continuous access to firm, reliable power sources, and meeting this goal will require extraordinary U.S. energy capacity across a range of energy technologies. Data centers used to develop state-of-the-art models need stable power access to operate effectively, and given the energy demand for AI development, an “all of the above” approach is required to ensure American AI leadership. This approach will help the U.S. meet immediate energy needs, expand sources of firm power, and maximize opportunities for AI to catalyze emerging energy technologies, like next-generation geothermal and advanced nuclear technologies.

While the U.S. is lagging in bringing energy generation online, China is rapidly building energy infrastructure for AI, having added over 400 gigawatts of power capacity last year compared to the just several dozen gigawatts added in the U.S. This disparity is especially concerning in the context of AI development: we project that 2GW and 5GW data centers will be needed to develop single advanced AI models for Anthropic in 2027 and 2028, respectively. With several other U.S. companies developing frontier AI models, we expect total frontier AI training demands in America to reach 20-25GW in total by 2028. For context, this is roughly twice New York City’s peak electricity demand, and at least as much will likely be needed for serving everyday AI model demand (called “inference”). Put another way, the U.S. AI sector needs at least 50GW of electric capacity by 2028 to maintain global AI leadership.

While the Trump Administration has already taken important steps toward removing barriers by setting ambitious nuclear power targets and accelerating National Environmental Policy Act (NEPA) reviews, to compete with China, it is critical that the U.S. take further action to address regulatory challenges that can delay energy projects.

The U.S. can enable the buildout of large-scale infrastructure for AI training in strategic locations, including by accelerating permitting, supporting targeted transmission projects, and collaborating with utilities to speed up grid interconnection processes. Beyond infrastructure for training AI, we propose broader reforms to support AI deployment nationwide and strengthen supply chains and workforce development. These include accelerating geothermal, natural gas, and nuclear permitting, developing a U.S. Army Corps of Engineers “nationwide permit” for AI data centers to speed Clean Water Act reviews, creating strategic reserves of critical grid components and gas turbines, and expanding technical training programs. Policymakers also have the opportunity through our proposals to keep prices low for ratepayers both through regulatory authority as well as by increasing overall supply.

We propose two strategic pillars that will address the full spectrum of infrastructure challenges—from large-scale training facilities to infrastructure supporting nationwide AI deployment.

Pillar 1: Building large-scale AI training infrastructure

• Make federal lands available as options for AI infrastructure construction, avoiding years-long state and local zoning processes;
• Accelerate NEPA processes for AI infrastructure, including by conducting advance reviews of certain projects;
• Partner with the private sector to expedite critical power line buildouts and upgrades; and
• Support utilities in making key reforms to accelerate interconnection processes, including by deploying AI to boost efficiency. Where critical for national security, leverage existing federal authorities to ensure that AI training infrastructure receives timely interconnections.

Pillar 2: Building broad-based infrastructure for AI innovation nationwide

• Accelerate geothermal, natural gas, and nuclear permitting;
• Establish National Interest Electric Transmission Corridors to speed transmission permitting in areas of significant AI data center growth;
• Strengthen domestic production of critical grid components and turbines through loan and loan guarantee programs and building strategic reserves of these products; and
• Support training, apprenticeship, and entrepreneurship programs for critical energy workers, electricians, and construction workers.

America has the economic strength, technical prowess, and innovative spirit to meet AI's infrastructure challenges. Our recommendations offer a practical blueprint that uses existing federal authorities while cutting through regulatory barriers that have held back energy development for too long.

By partnering across government, industry, and communities, we can build the foundation for decades of American leadership in AI.

Read the full report.

Generated by RSStT. The copyright belongs to the original author.

Source