Debunking the AI Infrastructure Overbuild Myth: The Unseen Potential

In the contemporary discussion on AI infrastructure, a recent post by Sequoia's David Cahn posits an intriguing argument that the current AI infrastructure is overbuilt, illustrated by some stark numbers pertaining to NVIDIA's GPU revenue and the required AI revenue to justify it. However, Guido Appenzeller strongly counters this argument, shedding light on the analytical missteps and bringing to fore the colossal potential AI infrastructure holds.

The Overbuild Argument Deconstructed:

David Cahn's argument hinges on the following figures:

- NVIDIA’s annual GPU revenue stands at $50 billion, indicating a requirement of $200 billion in “AI revenue” to justify this.

- The current "AI revenue" only amounts to $75 billion, allegedly leaving a gaping $125 billion hole.

Guido’s Counterargument:

1. Mismatched Calculations:

The analytical framework employed mixes capex (GPU cost), annual opex, cumulative revenues during GPU lifetime & annual revenues from AI applications to arrive at the $200 billion figure. A more refined calculation should focus on the annual return earned on the capital invested by GPU buyers, which presents a more accurate financial picture.

2. Misinterpreted Electricity Costs:

The argument oversimplifies the cost dynamics by suggesting a $1 electricity cost for each $1 of GPU. Realistically, an H100 PCIe, priced around $30k, utilizes approximately 350W of power, escalating to 1kW with server/cooling. With electricity priced at $0.10/kWh, the electricity cost over a 5-year lifecycle is significantly lower, about $0.15 for every $1 spent on hardware.

3. Underestimating the AI Revolution:

AI models, much like CPUs, databases, and networks, are integral infrastructure components. Currently, nearly all software leverages CPU/DB/Network, and a similar trend is anticipated for AI models in the near future. The AI infrastructure will fuel virtually all software, hence, its potential and impact are monumental.

4. Relabeling Infra Spend and Generated Revenue:

Networking infrastructure spending amounts to over $200 billion annually. However, this does not translate to $800 billion in "networking software" revenue. For instance, Google utilizes networking infrastructure for ad selling, categorizing the revenue as ad revenue. Similarly, Microsoft’s revenue from O365 doesn’t explicitly reflect networking infrastructure spend.

Conclusion

The purported $125 billion hole is a misrepresentation stemming from an oversimplified analysis. The expenditure on AI infrastructure, like the $50 billion spent on GPUs, can be easily amortized over a colossal $5 trillion worldwide IT spend. The extensive integration of AI in every software-laden product will ensure a robust return on investment, underscoring the immense potential and indispensability of AI infrastructure. Guido's excitement to invest in AI infrastructure and its myriad applications mirrors the burgeoning confidence and anticipation enveloping the AI domain, signaling a promising and transformative future.