Gavin Baker Says There Are No Idle GPUs Compared to 2000's Dark Fiber
a16z shared Gavin Baker's views, pointing out that during the peak of the 2000 bubble, 97% of the fiber optic cables laid in the U.S. were unlit, whereas currently there are no idle GPUs.
Baker stated that he was an investor during the 2000 tech bubble, which was essentially a telecom bubble, where a large amount of fiber was laid but could not be used due to a lack of supporting optical and routing equipment.
He compared that all current GPUs are in use, noting that one of the main issues in technical papers regarding training operations is that GPUs overheat due to high-intensity usage.
This comparison aims to distinguish current AI infrastructure investment from the overbuilding of 2000, emphasizing that GPU utilization is close to saturation rather than idle capacity.
Baker also mentioned that the return on capital expenditure for large tech companies has improved, contrasting with the debt-driven fiber expansion of 2000.
On the market side, this narrative reinforces the sustainability of AI computing power demand, with funds continuing to flow towards GPU and data center construction, benefiting chip suppliers like NVIDIA, while investors questioning the bubble face pressure to validate demand.
Additional information indicates that this viewpoint comes from a16z discussions, with Baker being a managing partner at Atreides Management, who has publicly compared valuation multiples with infrastructure utilization rates.
Source: Public Information
ABAB AI Insight
As an investor who experienced the 2000 telecom bubble, Gavin Baker was early in investing in technology and communications, witnessing companies like Level 3 and Global Crossing lay fiber on a large scale only to see assets idle due to low utilization. The historical path emphasizes the verification from overbuilding to actual usage.
His current capital path focuses on AI infrastructure and computing power investment, promoting funds to continue being allocated to data centers and chips by comparing dark fiber and GPU utilization, believing that current spending is supported by cash flow rather than solely by debt, and that utilization rates are a core indicator of non-bubble conditions.
Similar cases include the process of excess fiber from 2000 being gradually absorbed by subsequent demand, and the transition of server utilization rates from low to high in early cloud computing construction, with current AI computing power moving from model training expansion to inference and application scaling.
Essentially, this represents capital concentration: computing resources are concentrated in a few entities capable of procurement and deployment, with mechanisms proving demand authenticity through high utilization rates, thereby supporting sustained capital expenditure and allowing chip and infrastructure providers to gain pricing advantages, avoiding the pitfalls of idle assets.
ABAB News · Law of Cognition
Idle capacity is a bubble signal; full utilization is proof of demand.
Laying does not equal lighting; utilization determines asset value.
Historical comparisons are effective, but cycle differences determine outcomes.