---
type: idea
title:
created: 2026-07-07
status: seed
origin_sources: []
related_auto: []
used_in: []
---

## Claim

## Why it matters

## Evidence / examples

## Counterpoints

## Related

## Draft hooks
- Possible headline:
- Possible angle:
- What this is really about:


Recently listened to a podcast with Lambda CTO Stephen Balaban, and the conversation touched on the spicy GPU depreciation policy topic: the pre-2025 allegations that CSPs were gaming their GPU depreciation estimates to achieve supernormal margins. Stephen mused that Lambda is among the few neoclouds today still running GPUs that have been fully depreciated, and that the people who expressed skepticism around the depreciation policy were "completely wrong."

Credit where its due - it is great to be able to reap economic value from a fully depreciated asset. I do believe Lambda's ability to still milk its old fleet is more a testament to how wildly agentic use cases picked up after March 2025 (Claude Code launch). Most of the concerns around depreciation policy predated Claude Code -- and there was no reason for the skeptics to believe otherwise back then. Hence, I don't fully agree with Stephen's "completely wrong" posturing. 

Today after Claude Code and the 'agentification' of our worlds, the 5-year average life estimate may look too conservative in hindsight, given how much economic value these GPUs have kept generating. Some companies have extended their GPU life estimates to 6 years to spread the cost of these GPUs over a longer period -- and that's only fair given the expectation of extracting real economic value from the hardware now that token-hungry workflows have been discovered.

Do upcoming GPUs follow the same curve, though? I don't think there is any way to reliably predict that today. Situation when this holds -- the labs and/or application-layer players discover novel use cases that stretch the economic value of compute beyond 6 years (before extremely good newer GPU generations necessitate labs to declare their older fleets obsolete, or inference-stack innovations make models way more compute-efficient and shrink that value). That's a lot of things falling together -- only then do the labs get to see supernormal economic value from the GPUs, like Lambda saw. As things stand today, though, this may or may not happen. And that's why I am so keen to keep watching the CSP earnings and track their margins.