facebook twitter instagram linkedin google youtube vimeo tumblr yelp rss email podcast phone blog search brokercheck brokercheck Play Pause
%POST_TITLE% Thumbnail

The Hidden "Network Tax" on AI Capital

By Jim Brilliant, CFA, CIO, Portfolio Manager

The massive AI boom is being built on a precarious capital bet: that the GPUs filling today’s data centers will remain economically useful for many years. On financial statements, operators typically capitalize these chips as long-lived assets, depreciating them over four to six years—treating them like data center real estate or power infrastructure.

This assumption has sparked a fierce debate. Skeptics argue that rapid innovation cycles will render current chips obsolete long before their accounting life ends. Optimists counter that secondary markets and inference workloads will preserve their value.

Both sides are missing a crucial dimension. The true useful life of a GPU is not defined by how many years it sits in a rack, but by how much of that time it spends doing math versus waiting on the network.


The Hidden "Network Tax"

Every GPU hour effectively splits into two states: computing and waiting.

  • Computing: The GPU is crunching tensors and delivering value.
  • Waiting: The GPU is idle, waiting for gradients during training, parameters in distributed setups, or data fetched across regions.

This "waiting" time is a hidden tax on capital. Two operators with identical GPU fleets will have vastly different unit economics if one utilizes a modern, flat 400G/800G fabric while the other is throttled by legacy links. A GPU that spends 30–40% of its wall-clock time idle behind a slow network is aging prematurely, regardless of when it was manufactured.


The Math: The High Cost of Waiting

The financial impact of this network bottleneck is not theoretical—it is quantifiable.

Research by Lumen Technologies on emerging "Cloud 2.0" architectures highlights the staggering cost of data movement. Their analysis calculates the cost of "waiting" while transferring a single petabyte of training data to a 64-node B200 cluster (burning ~$2,000/hour):

  • At 10 Gbps: The transfer takes ~222 hours. "Waiting Cost": ~$431,000.
  • At 100 Gbps: The transfer takes ~22 hours. "Waiting Cost": ~$43,000.
  • At 400 Gbps: The transfer takes ~6 hours. "Waiting Cost": ~$11,000.

By moving from a legacy 10G link to a modern 400G fabric, an operator saves roughly $420,000 on a single data load. When you scale this to exabytes of data and thousands of nodes, the move to 800G and 1.6T stops being a luxury and becomes a financial necessity.


Why Speed is an Economic Shield

The transition to high-speed optics is the primary lever for extending GPU life across both training and inference.

  • For Training: High-speed optics sharply reduce "serialization time"—the time it takes to put data on the wire. In distributed training, where steps cannot advance until the slowest node reports back, shaving 20% off communication time directly reduces the GPU-hours required to train a model. The chip hasn't changed, but its productivity has skyrocketed.
  • For Inference: Modern AI applications fan out across microservices and vector databases. On legacy networks, congestion creates "tail latency" (p99 delays), forcing operators to over-provision GPUs just to meet SLAs. Faster networks smooth out these tails, allowing each GPU to handle higher query density without breaking latency budgets.


The Future is Purpose-Built

This economic reality is driving the emergence of what Lumen describes as Cloud 2.0: a move away from the "internet swamp" of congested, carrier-neutral overlays toward purpose-built optical fabrics. These architectures extend dark fiber directly to power-rich AI corridors, treating extreme bandwidth and low latency as first-class design goals.

Seen through this lens, the future network is really the other half of the GPU useful-life equation. High-speed optical fabrics that can move exabytes without leaving silicon idle are what make four-, five-, or six-year depreciation schedules defensible. In the Cloud 2.0 economy, the network does not just connect the computer; it saves the computer. By eliminating the bottlenecks of legacy infrastructure, modern fabrics ensure that billions in capital spend their life delivering returns, rather than waiting in traffic.

_____________________________________________________________________________________________________________________________________

Members of the CM Portfolio Management and Research Team

  • Arnold Van Den Berg, Founder, CO-CIO
  • Jim Brilliant, CFA, CO-CIO, Portfolio Manager
  • Scott Van Den Berg, CFP, ChFC, CEPA, AIF, President
  • Aaron Buckholtz, CFA, Portfolio Manager, Director of Trading
  • Brill Brilliant, Fixed Income Portfolio Manager
  • Adrian Kachmar, Senior Equity Analyst
  • Greyson Brilliant, CFA, Senior Equity Analyst


Disclosures: Century Management Financial Advisors ("CM") is an independently registered investment adviser with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. CM is also registered as a Portfolio Manager in the Province of Ontario.  

The information contained herein is for educational and informational purposes only and should not be construed as individualized investment advice or a recommendation to buy or sell any security, strategy, or investment product. The views expressed are those of the author as of the date of publication and may change without notice. Certain statements may be forward-looking in nature and involve known and unknown risks and uncertainties; actual outcomes may differ materially from those expressed or implied.

Any third-party data, research, or market examples referenced in this material are believed to be from reliable sources, but their accuracy or completeness cannot be guaranteed. All economic and market data are illustrative and are not intended to reflect the performance of any specific investment, portfolio, or advisory service. References to technology, companies, or market participants are for context only and should not be considered endorsements.

Past performance is not indicative of future results. Investing involves risk, including the potential loss of principal. Clients should consider their investment objectives and risk tolerance before investing and should consult directly with CM or another qualified professional regarding their specific situation.A full description of our Firm’s business practices, including our Firm’s investment management services, wealth plans and advisory fees, are supplied in our Form ADV Part 2A and/or Form CRS.