Horowitz Warns of AI Computing Power Arbitrage Risks Amid Loudcloud's Bandwidth Crash
a16z shared Ben Horowitz's reflection on his founding of Loudcloud, which faced bandwidth prices of $1500 per megabit in 1999 and difficulty in obtaining it, forcing the signing of a three-year long-term contract, with prices dropping to $50 by 2001.
He pointed out that locking in high-priced bandwidth through borrowing could lead to company distress after a price crash, a risk beyond the control of entrepreneurs.
Horowitz compared this to the current AI neocloud, believing the core dynamic is a supply shortage rather than an excess, with token prices declining but GPU prices not falling in sync.
Data center space prices are rising due to labor shortages in construction and policy restrictions increasing costs across the supply chain, with demand not yet waning.
He stated these factors create a tailwind for companies like CoreWeave, although the debt market is independently under pressure, leading a16z to refrain from investing in any neocloud.
From a market perspective, this view suggests that intermediaries locking in high-priced computing power and taking on debt face the risk of price reversals, with capital flowing to infrastructure with validated demand, putting pressure on arbitrage computing power lessors.
Supplementary information indicates this discussion emphasizes that input price fluctuations beyond entrepreneurs' control can destroy businesses reliant on long-term contracts and leverage.
Source: Public Information
ABAB AI Insight
Ben Horowitz co-founded Loudcloud in 1999, experiencing the process of bandwidth costs dropping from high long-term contracts to market price crashes during the dot-com bubble. The company later sold its hosting business and transformed into Opsware, which was acquired by HP, highlighting the historical path of learning about price and debt risks from infrastructure operations.
Its capital path at a16z reflects an avoidance of investing in neoclouds reliant on long-term high-price contracts and borrowing, prioritizing demand-side or non-arbitrage businesses, shifting resources from potential price reversal exposure to more controllable software and application layers.
Similar cases include the underutilization of fiber laid during the 2000 telecom bubble leading to asset idleness, and the pressure of long-term capacity contracts during price declines in early cloud computing construction. The current AI infrastructure is in a phase of high capital expenditure and supply constraints.
Essentially, this is a transfer of pricing power: market price fluctuations of input costs (bandwidth or GPUs) can quickly erode fixed contracts and debt structures, mechanically amplifying exposure through long-term locking and leverage, causing intermediaries to bear the brunt during price declines, while end-demand parties benefit from cost reductions.
ABAB News · Cognitive Laws
Long-term locking of high-priced inputs can destroy leveraged businesses upon price reversals.
Supply shortages support premiums, but technological iterations can change cost curves.
External price fluctuations beyond entrepreneurs' control are the hardest risks to hedge.