Musk and Huang: AI’s Power Needs and the Case for Agent Monitoring

Musk predicts a link between power and growth; Huang describes isolating and independently monitoring AI agents.

EMElon MuskゲストJHJensen Huangゲスト · NVIDIA
チャンネル
Right Side Broadcasting Network
公開日
2026年9月29日
動画
26:37
読了時間
5 分
原文
英語
YouTube で開く
スクロールすると動画は右下に小さくなって再生を続けます。字幕は閲覧言語に合わせて切り替わります。
閲覧言語2 言語で読めます
この記事はまだ日本語に翻訳されていないため、英語の原文を表示しています。
要約

30 秒でわかる

Elon Musk argues that the United States needs much more electricity and chipmaking capacity to expand AI, and predicts a link between increased power use and GDP. Jensen Huang describes a security approach that restricts AI agents’ access and watches them from a separate system. Anthropic co-founder Tom Brown says newer models can handle longer delegated tasks, while the speakers also discuss local data-center jobs and a safety declaration.

  1. 1Energy and chipmaking are long-term constraintsMusk says the United States needs enough electricity and secure logic and memory production to compete.
  2. 2Musk links one percent more electricity use to one percent more GDPHe offers it as a prediction based on improving computing efficiency, not a measured outcome.
  3. 3Huang expects infrastructure construction to create jobsHis estimate of about a million jobs accompanies an assumed buildout of 10–20 gigawatts a year.
  4. 4An agent needs both restricted access and independent monitoringHuang describes OpenShell isolation and monitoring outside the agent’s environment using BlueField hardware.
  5. 5Brown says newer models can take on longer tasksThe Anthropic co-founder expects more AI-assisted engineering to improve software over time.
  6. 6The speakers describe a joint safety declarationMusk and Huang say it includes monitoring, internal controls, audits and sharing practices.

ポイント

  1. Musk predicts that a 1% rise in power use could correspond to roughly 1% more GDP.
  2. Huang recommends restricting agent permissions and monitoring from outside the agent’s environment.
  3. Huang estimates a 10–20 gigawatt annual buildout could create about a million construction-related jobs.
  4. Brown says newer models can handle longer delegated tasks than previous ones.
全文

チャプターごとに整えた全文を読む

文字起こしのみ
01

Power and chips set the pace

Huang and Musk say AI infrastructure depends on expanding electricity and chip production.

The moderator asks how the United States can lead in what the speakers call superintelligence and make the benefits reach ordinary Americans. Jensen Huang says the shift is reinventing the computing stack, from chip design and software training to applications.

Huang says energy is the foundation: power runs computing, which runs algorithms and applications. He argues that the United States needs power generation and delivery as well as strong chips and algorithms, and says data-center builders must do more to partner with communities and share prosperity.

Elon Musk agrees that energy and chip production need to scale. He identifies China as the main competitor and says the long-term challenge is securing enough electricity and capacity to fabricate logic and memory chips in the United States or a secure region.

02

The economic case for more power

Musk proposes a relationship between electricity use and economic output, and discusses solar-powered computing in orbit.

Asked about SpaceX's publicly discussed power plans, Musk estimates average U.S. power consumption at 500 gigawatts and says another five gigawatts of steady power would represent roughly 1%. He predicts that a 1% increase in power use could correspond to about 1% more GDP as the useful computing per watt improves.

Huang and Musk work through an illustrative estimate of economic output per gigawatt. Using Musk’s assumption, they arrive at roughly $60 billion per gigawatt a year.

Musk says SpaceX is looking to orbital computing to access more energy than is available on the ground. He describes a goal for SpaceX and Tesla to produce 200 gigawatts of solar capacity per year, arguing that solar equipment in space can receive substantially more sunlight than installations on Earth.

03

Data centers and local communities

Huang forecasts construction jobs; Musk describes local spending and infrastructure around xAI facilities.

Huang says strong potential returns from computing could make launch costs affordable relative to the benefit. The moderator asks what data centers and related energy projects offer ordinary Americans and the towns where they are built.

Huang calls the facilities 'superintelligence factories' because they produce something of economic value, rather than merely store data. He lists power, computing, networking and cooling as parts of the infrastructure and imagines it making services available from anywhere.

Huang estimates that building 10–20 gigawatts of infrastructure annually could create around a million construction-related jobs, including work on power plants, cooling and pipes. He presents this as an opportunity to bring more skilled trades and industrial work into the U.S. economy.

Asked about xAI's facilities, Musk says the economic effect on the surrounding area has been positive and that employers have had trouble finding enough workers. He also mentions discounted Starlink access and a planned $250 million water-recycling plant as ways to contribute locally.

Huang argues that demand for AI infrastructure is encouraging private investment in different energy technologies, including nuclear power and storage. In his view, this creates a chance to finance future energy systems through market demand.

04

Contain and monitor AI agents

Huang describes isolating agent permissions and using independent hardware to monitor behavior.

The moderator asks about Nvidia's newly introduced agent safety framework. Huang says systems need secure containment during evaluation and continuous monitoring even when they are thought to be contained. He recommends monitoring from outside the agent's own environment.

Huang describes OpenShell as a way to isolate an agent and restrict its permissions, and says a separate BlueField chip can watch its behavior and escalate actions that violate policy. He compares granting access and monitoring an agent with assigning permissions to an employee.

Huang argues that safety and technical progress need not conflict: in his view, building safer systems can enable wider deployment. The moderator then asks Musk how Starship contributes to a future of solar-powered computing in orbit.

05

Starship and the orbital-computing ambition

Musk discusses launch cadence, space cargo and a proposed long-term expansion of orbital computing.

Musk says Starship's first orbital flight deployed Starlink V3 satellites and predicts the rocket could launch weekly or twice weekly the following year. He presents those launch rates as expectations, while describing much larger ambitions for delivering cargo to the Moon and Mars.

Anthropic co-founder Tom Brown joins after a delayed arrival. Musk estimates that a self-sustaining city on the Moon or Mars would need roughly a million tons delivered to its surface, and discusses hundreds of gigawatts a year of potential orbital computing as a long-term ambition.

06

What AI can do for users today

Brown describes longer delegated tasks; Huang and Musk give examples from research and everyday use.

Brown says Claude Opus 5.5 can handle longer delegated tasks than earlier models. He expects AI-assisted software development to increase the amount of engineering work that can be done, potentially improving products and lowering their cost over time.

Huang says he uses Claude for comparative research: he supplies a list of questions and asks for a report with charts. He says a task that previously took him four to eight hours of weekend research can return a report in around three minutes.

Musk says AI can help people organize their lives and work. He recounts hearing anecdotes of AI identifying medical problems after a doctor made an error.

07

Judgment, safety and a White House declaration

Huang raises questions about AI judging users, and the speakers describe a safety declaration.

Asked about values in AI tools, Huang questions how much judgment a system should impose on a user's intentions. He uses a knife as an analogy and notes that defensive cybersecurity activity can resemble an attack, even though the user's goal differs.

Asked about the preceding White House meeting, Musk says participants signed a joint declaration on superintelligence safety, including monitoring and committees. Huang describes internal controls, documentation, audits and sharing practices across the industry as elements of the declaration.

同じチャンネルの記事