--- title: "AI Competition Enters the GW Era: Musk and Jensen Huang Discuss Power, Compute, and Security" type: "Topics" locale: "en" url: "https://longbridge.com/en/topics/44125906.md" description: "On September 29, a core group of figures from the US AI industry converged in Washington. In the morning, Trump posted to America.gov; during the same event, Musk, Jensen Huang, and Gavin Baker of Atreides Management held a nearly half-hour discussion. Subsequently, Trump met with executives from tech giants including Google, Meta, OpenAI, Anthropic, NVIDIA, and xAI at the White House..." datetime: "2026-09-30T05:46:08.000Z" locales: - [en](https://longbridge.com/en/topics/44125906.md) - [zh-CN](https://longbridge.com/zh-CN/topics/44125906.md) - [zh-HK](https://longbridge.com/zh-HK/topics/44125906.md) author: "[潘驴邓晓闲缺一](https://longbridge.com/en/profiles/27015735.md)" generator: "portal-rs" --- # AI Competition Enters the GW Era: Musk and Jensen Huang Discuss Power, Compute, and Security On September 29, a core group of players in the U.S. AI industry converged in Washington. In the morning, Trump released **America.gov**; at the same event, Musk, Jensen Huang, and Gavin Baker from Atreides Management held a nearly half-hour discussion. Later, Trump met with leaders from tech giants including Google, Meta, OpenAI, Anthropic, NVIDIA, and xAI at the White House, culminating in the signing of the **White House Accord on Super Intelligence**. Government applications, compute expansion, and safety governance were all placed on the same table in a single day. America.gov addresses the application layer first. Trump signed an executive order defining it as the unified entry point for U.S. federal online services: users can query government information using natural language, with future integrations into specific tasks like passport renewals and Medicare enrollment. The U.S. government is no longer just an AI policy maker but also becoming a large-scale adopter. What Musk and Jensen Huang discussed at length was an even more fundamental issue: **electricity**. The U.S. utility power generation in 2025 was approximately **4.43 trillion kWh**, translating to an average annual generation capacity of about **506 GW**. Adding 5 GW to the entire U.S. power system represents nearly a 1% scale. As top AI companies discuss data center expansion, the units have shifted from how many GPUs to how many MW or GW, and when they can actually be powered up. The EIA explicitly noted that a significant portion of recent new electricity demand in the U.S. comes from the commercial sector, including data centers. SpaceX's changes are very intuitive. In Q2 2025, the company's **Nameplate Compute Draw** was only **0.4 GW**, rising to **1.4 GW** by Q2 2026; during the same period, quarterly capital expenditures reached $18.4 billion, with AI-related capex hitting **$15.8 billion**, accounting for the vast majority. Musk projected in the Q2 earnings call that it would exceed 2 GW by the end of this year, and by the end of 2027 might "be closer to 10 GW than 5 GW." To ensure compute ramp-up, the company even plans to build out power, cooling, and supporting electrical systems at a larger scale in advance. This also explains why Jensen Huang increasingly prefers using **AI Factory** rather than data center to describe this business. At the August earnings call, NVIDIA had already calculated the economics per GW clearly: in the Hopper era, the revenue opportunity corresponding to every GW of AI infrastructure was about **$18 billion**, which rose to **$25 billion** with Blackwell, and further increased to **$40 billion** with Vera Rubin. As SpaceX/xAI continues to expand from 1.4 GW to several GWs, **electricity not only determines when data centers can turn on, but also directly dictates when upstream chips, networking, and servers can recognize revenue.** Musk further pushed the energy issue onto solar and orbit. His long-term goal given that day was for **SpaceX and Tesla to jointly achieve an annual solar manufacturing capacity of 200 GW**. The boundaries of large AI companies are shifting from renting and buying power toward building their own energy capabilities. Orbital compute follows the same logic. Google's **Project Suncatcher** is testing whether TPUs can adapt to radiation, vibration, and vacuum heat dissipation environments, with the subsequent goal of connecting multiple computing satellites via high-speed optical interconnects. Google currently still defines it as a research project, far from GW-level space data centers in the short term, but this route has entered the real hardware validation phase. While AI infrastructure continues to expand outward, another cost is beginning to surface: **safety**. The **White House Accord on Super Intelligence**, signed by leaders from Google, Anthropic, Meta, OpenAI, xAI, and NVIDIA, splits frontier model governance into four layers: internal controls during training and deployment, verification by independent internal teams, third-party external audits, and oversight by an independent committee under the board of directors. The scope covers cyber, biological, and chemical risks, as well as issues like unauthorized model access to technical systems; participating companies also agreed to regularly establish common standards and retained the possibility in the document of writing some requirements into law or regulatory rules in the future. Currently, this remains a voluntary agreement. NVIDIA's release of the **Open Agent Safety Platform** the previous day perfectly mapped these safety requirements down to the product layer. Among them, **OpenShell** establishes independent operational boundaries for AI agents, tracks behavior, and enforces permission policies; **Sentry** runs on BlueField-4 DPUs, continuously monitoring outside the agent, capable of isolating breaches in milliseconds. Safety is moving from the model layer into **CPU, DPU, runtime environment, identity permissions, and audit systems**; what NVIDIA can cover is no longer just GPUs. Putting all the information from this day together, the boundaries of AI capex have widened significantly compared to two years ago. Initially, the market calculated GPUs, then moved on to HBM, optical modules, and switches; now it needs to continue calculating power, cooling, gas turbines, land, and grid connection. Once Agents truly enter production environments, isolation, permissions, and audits will need to be added as well. **Compute is being measured in GWs, with each GW corresponding to tens of billions in hardware value, and safety is once again moving from the software layer into CPUs and DPUs.** Sources: The White House, GSA, SEC, EIA, NVIDIA, Google. ### Related Stocks - [TSLA.US](https://longbridge.com/en/quote/TSLA.US.md) - [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md) - [SPCX.US](https://longbridge.com/en/quote/SPCX.US.md) - [ANTH.NA](https://longbridge.com/en/quote/ANTH.NA.md) - [OpenAI.NA](https://longbridge.com/en/quote/OpenAI.NA.md) - [META.US](https://longbridge.com/en/quote/META.US.md) - [GOOG.US](https://longbridge.com/en/quote/GOOG.US.md) - [GOOGL.US](https://longbridge.com/en/quote/GOOGL.US.md) - [VERA.US](https://longbridge.com/en/quote/VERA.US.md) - [UXSD.SG](https://longbridge.com/en/quote/UXSD.SG.md) --- > **Disclaimer: This article is for reference only and does not constitute any investment advice.**