Original author: Zhao Ying

Original source: Wall Street Journal

NVIDIA is secretly advancing a project across the United States that is unrelated to chips, but may have more strategic defense value than any GPU: acquiring dark fiber that has been laid but never activated nationwide and building its own telecom-grade network infrastructure.

On Tuesday, according to disclosures from Needham and Wolfe Research, the project is expected to cost between $5 billion to $10 billion over the next three years, with a total bandwidth potential of up to 7.6 Petabits per second once completed—this scale surpasses enterprise-level demand and aims directly at control of telecom operator-grade infrastructure.

Unlike directly purchasing bandwidth, acquiring dark fiber means NVIDIA gains complete usage rights to the fiber lines, defining everything from fiber transmission device selection to network topology design independently. This is a path that requires heavier assets and longer cycles but is also more exclusive.

Wolfe Research characterized this layout in their report as NVIDIA's "ultimate insurance strategy" in response to the rise of ASICs. In the context of major clients developing their own chips that continuously erode GPU market share, NVIDIA is attempting to extend competition from "whose chip is faster" to "who can deliver computing power directly to customers"—by controlling the physical transmission layer, opening a second moat outside hardware competition.

For investors, this news reflects a fundamental shift in NVIDIA's strategic positioning: from being an upstream component supplier for cloud service providers to evolving into a computing power infrastructure operator that directly reaches end customers. This means the long-standing subtle symbiotic relationship between NVIDIA and AWS, Azure, and GCP may enter a new phase of competition and cooperation.

7.6 Petabits per second: An underestimated figure

The total bandwidth limit of NVIDIA's dark fiber network—7.6 Petabits per second—needs to be understood within the appropriate context. Currently, the single fiber transmission capacity of the global internet backbone typically ranges from tens to hundreds of Tbps, while NVIDIA's planned scale corresponds to the aggregation capacity of thousands of backbone-level optical fibers.

The dark fiber network can connect multiple data centers, forming a massive cluster. Analysts believe that NVIDIA can lease this network to Neoclouds, thus narrowing the gap with hyperscale data center operators, many of whom possess extensive optical fiber networks.

The activation and commercialization cycle of dark fiber typically takes 18 to 36 months. From fiber acquisition to fiber transmission device deployment to the final mile connections with various data centers and corporate nodes, NVIDIA needs to complete a path that traditional telecom operators take five to seven years to finish within three years. The aggressiveness of this execution pace itself is sending signals to the market.

"Ultimate Insurance": How dark fiber hedges against ASIC risks

To understand the strategic urgency of dark fiber, it is essential to examine the changes in NVIDIA's competitive landscape. Companies like Broadcom and Marvell are customizing AI-specific chips (ASICs) for cloud giants such as Google and Amazon, gradually eroding GPU market share in specific scenarios like inference. When major clients have the option to develop their own chips, NVIDIA's clout in the computing power delivery chain becomes less secure.

The layout of dark fiber is a response to this structural risk—by building their own network to directly reach end customers, NVIDIA is able to transform computing power delivery from "resold through cloud service providers" to "end-to-end self-delivery." Wolfe Research's choice of the term "ultimate insurance" implies that even if future clients choose ASIC over NVIDIA GPUs at the computing layer, as long as computing power is delivered through NVIDIA's network infrastructure, NVIDIA still retains authority and economic benefits over the delivery layer.

This logic is highly consistent with a series of strategic moves NVIDIA has made in recent years: a $20 billion investment in acquiring Groq's technology and team established the GPU+LPU heterogeneous computing paradigm; a $30 billion investment in OpenAI locked in the largest client for model training demand. The dark fiber network completes the last piece of the entire chain from computing power production to delivery—the physical transmission layer, which is the most difficult to replicate.

Dual pressures from cloud giants and telecom operators

For AWS, Azure, and GCP, the direct sales model of GPU as a service poses a tricky prospect: the most important upstream supplier is becoming a direct competitor in specific scenarios. In the past, when enterprise clients rented GPU instances on cloud platforms, cloud service providers extracted considerable platform premiums from that. In the future, if NVIDIA delivers directly through its own network, this portion of profit will be redistributed between supply and demand.

For traditional telecom operators, a new buyer with a market capitalization exceeding $5 trillion making large acquisitions of long-distance dark fiber may drive a reevaluation of related assets. Dominant players in the dark fiber market, such as Zayo and Crown Castle, face not only rising costs from competitive buyers but also accelerated penetration of tech giants into telecom infrastructure—this trend has extended from data centers to transmission networks.

However, it is important to note that the commercial operation of the dark fiber network is not an easy task. Optical transmission equipment, network operation teams, and connections to various data center enterprises in the final mile—NVIDIA has virtually no experience in these areas. Even if the network is built, efficiently operating it and achieving commercialization still pose significant execution risks.

The acquisition of dark fiber is not NVIDIA's isolated adventure but a part of its transformation from a chip company to a full-stack AI infrastructure operator. In just the past year, NVIDIA's layout has covered computing architecture (Groq), customer ecosystem (OpenAI), supply chain (SK Hynix HBM collaboration), and geopolitical layout (collaboration with Japan's robotics and automotive industry)—and the dark fiber network is the part of this puzzle involving physical control.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。