Cogent Communications (NASDAQ:CCOI) founder and CEO Dave Schaeffer said artificial intelligence is “real” and “transformational,” but he cautioned that the economics of AI are still developing and that much of the value may ultimately accrue to end users and businesses built on top of AI rather than the companies funding infrastructure. Speaking in a conference session, Schaeffer also outlined where he expects AI-related demand to benefit connectivity providers, including greater data collection and storage, more fiber buildouts at the edge, and increased demand for “Layer one” transport services to connect remote training facilities.
AI’s impact: transformative, but monetization remains unclear
Schaeffer framed AI’s potential by comparing it to the internet’s impact, arguing the internet proved highly deflationary and shifted value away from service providers toward consumers and new application-layer business models. He noted that roughly $2 trillion was invested during the telecom bubble of the late 1990s and early 2000s, producing broad societal benefits even as telecom providers’ share of the S&P 500 and GDP declined.
- Tokenization, which he said is falling in cost at about 85% per year.
- Large language model building, where compute costs are declining, but model size and complexity are increasing as training datasets expand.
Schaeffer emphasized that AI’s development relies on data generated via the internet, describing stored data as the “raw material” for model training. He said the internet has generated “something like 20,000 zettabytes of data,” with about 1,000 zettabytes stored and usable for model development. He also pointed to proprietary datasets as potential advantages, offering the example that data from Twitter users could be a source of value for xAI’s Grok.
Traffic patterns and power constraints
Schaeffer described three broad “flavors” of AI inference. In the simplest case, he said, a model can be applied to data at an edge site and return results almost instantly. A second scenario involves proprietary data that exists only inside a company’s network or on a device, which he said could “radically change the internet” by shifting traffic patterns from download-heavy to more upload-intensive workflows as data is sent upstream for processing. In a third scenario, he said, an inference request may require access to massive centralized data farms holding “tens of thousands of zettabytes,” resulting in longer response times.
He also highlighted power availability as a major constraint. He said traditional (non-AI) data centers now consume about 2.2% of power produced in the developed world, while AI adds roughly 2.5% of grid load, bringing the combined figure to about 4.7%—nearly eliminating the “slack” that previously existed in a long-standing model of power growth. He said this is pushing data center construction toward remote locations where land, cooling, water, and power are available, and he argued the industry cannot “just keep throwing power at the problems,” implying future progress will require greater efficiency or different processor technologies.
Despite AI’s momentum, Schaeffer said “no one has a good monetization model for AI,” and suggested that firms spending heavily may lack durable moats, leading to commoditization. He added that his expectation is that “the $1 trillion that’s going into training facilities” will likely earn returns “far below the average cost of capital.”
Where Cogent sees connectivity benefits
Asked how connectivity could benefit from AI-driven demand, Schaeffer outlined three areas. First, he said the increasing value of raw data should drive more collection and storage, supporting demand for core internet transit. He described Cogent as “the largest carrier of transit traffic globally,” carrying about “2 EB a day” and “about 25% of global internet transit traffic.”
Second, he said AI-driven use cases will require more fiber investment at the edge because coaxial networks cannot provide the upload speeds likely to be needed, and mobile networks face significantly higher costs per bit-mile than fiber. While he said broader fiber proliferation benefits society, he cautioned it does not guarantee attractive returns for the companies building that fiber.
Third, he said training facilities are likely to be built in locations that are not co-resident with data storage sites, creating demand for long-haul transport services. He argued AI training will require “old-fashioned” Layer one connectivity—such as dark fiber and wavelengths—because GPU-intensive training workloads prioritize low latency and consistent performance. He added that TCP/IP’s buffering and handshake behavior is not well-suited to AI training.
On how new routes may be deployed, Schaeffer cited three mechanisms: using existing fiber inventory laid decades ago, pulling fiber through empty conduits, and new greenfield builds to remote sites. However, he said a handful of hyperscalers have “monopsony power” that can push build economics to “very low IRRs,” often in the low single digits, contributing to industry capital deployment below the cost of capital. He said Cogent has been “extremely disciplined” in capital allocation.
Business update: Sprint integration, leverage, margins, and growth outlook
Schaeffer addressed the company’s share price decline over the prior year and argued the market reaction has been “far more violent than the operations justify.” He said Cogent went public in 2005 at $6 per share and delivered 18 years of organic compounded growth of 10.2% with average margin expansion of 220 basis points per year, along with a policy that returned $2 billion to shareholders.
He characterized Cogent’s acquisition of Sprint assets as “complicated” and “messy,” noting that the company received $700 million in cash over 54 months from T-Mobile as part of the transaction. He said investors have focused on headline total revenue declining, which he said was anticipated. Schaeffer said that over the nine quarters Cogent owned the Sprint assets, Cogent’s core business grew 27% while the acquired Sprint business declined 64%.
He also pointed to the company’s Wavelength business, which he said grew 100% year-over-year and 19% sequentially, though from a small base. He said leverage rose from 3.5x to 6.6x due to work required to repurpose Sprint’s network and buildings, and he attributed some of the stock pressure to that higher leverage and slower consolidated growth.
Looking ahead, he said the company is “back to the point of inflecting to positive top-line growth,” estimating multi-year growth of 6% to 8%, below the 10.2% pace achieved over the prior 18 years. He also described a sharp post-acquisition margin swing, saying EBITDA margins fell from 40.5% before the acquisition to 1% immediately after, and have since recovered to 22%. He said the goal is to return to 40%, primarily through selling more on-net services and reducing exposure to low-quality, off-net business.
Schaeffer provided figures showing a shift toward on-net sales, including that 80% of sales in the most recent quarter were on-net. He also described how, immediately after the acquisition, the combined company was 47% on-net, 48% off-net, and 5% non-core, and that by the end of Q4 2025 it was 61% on-net, 38% off-net, and 1% non-core.
Balance sheet actions: IP address securitization and real estate sales
On deleveraging and monetizing assets, Schaeffer said Cogent pursued three approaches. First, the company leased out a portion of its IP address inventory and securitized the lease cash flows via an asset-backed securitization, which he described as novel. He said the company raised $380 million through a ring-fenced SPV structure.
Second, he discussed the technical buildings acquired with the Sprint transaction: 482 fee simple owned facilities totaling 1.9 million square feet with 230 MW of existing power. He said many were filled with legacy telephone switching equipment—23,500 cabinets—which Cogent removed. He added that the facilities were built around -48V DC power systems and required conversion to support typical computing equipment, leading the company to convert power systems in a subset of sites.
Third, he said some sites were too large for Cogent to fully utilize with its go-to-market model, prompting plans to sell 24 facilities totaling 109 MW and about 1 million square feet. He said 10 of those facilities were under letters of intent to be sold, and that the conversion work began in June 2024 and was completed in June 2025.
Schaeffer also noted that the company does not have conduit to monetize because the Sprint fiber network is direct-buried, but said Cogent does have dark fiber and has sold a limited amount, with a primary focus on using the fiber for its Wavelength network.
About Cogent Communications (NASDAQ:CCOI)
Cogent Communications (NASDAQ:CCOI) is a multinational Internet service provider specializing in high-speed Internet access and data transport services. The company operates one of the largest Tier 1 IP networks in the world, offering wholesale and enterprise customers reliable, low-latency connectivity. Cogent’s core services include dedicated Internet access, Ethernet transport, wavelength services, and MPLS-based IP Virtual Private Networks, all delivered over its privately owned, fiber-optic backbone.
In addition to network connectivity, Cogent provides data center colocation and managed services designed to support businesses with demanding bandwidth and redundancy requirements.
