B3IQ Blog · Insights · 4 min read
The GPU lifecycle is way longer than anyone is pricing in
The fight over GPU depreciation misses how GPUs actually age: they move down the job ladder, from training to inference to batch work, and keep earning.
A GPU does not stop being useful when a newer one ships. It moves from frontier training to serving to embeddings and batch jobs, and serving software keeps making old cards faster. Renters pay the going rate for every tier; owners keep the whole career, including the long tail.
Written by Sean Geng, cofounder and CTO.
Fig. 1: Every generation, from the H200 to the RTX 5090, ages down the same job ladder.
The funniest fight in AI right now is an accounting fight.
On one side you've got Michael Burry, saying the hyperscalers are understating depreciation by stretching GPU useful lives - by his math, $176 billion of it between 2026 and 2028.
On the other side, Nvidia's CFO clapped back that the A100s they shipped six years ago are still running at full utilization, and Jensen posted that the A100 fleet is "mission-capable from 2020 through 2029."
Here's my take: both sides are talking past how GPUs actually age.
The market prices a GPU like an iPhone - hot for 2 years, in a drawer by year 4. But anyone who actually runs hardware knows a GPU ages like a work truck. It doesn't stop producing when the new model drops. It moves down the job ladder & keeps earning.
No lab runs on one machine type
The depreciation spreadsheets treat a GPU fleet like one homogeneous blob doing one job. That's not how any real AI operation works.
Imagine you're running one:
- Frontier training wants the newest silicon, period. But training is the minority of compute.
- Serving your flagship model is the demanding end of inference - still wants strong hardware, doesn't need to be bleeding edge.
- Then there's the long tail that quietly eats most of the cycles: fine-tuning, embeddings, batch jobs, draft models for speculative decoding, classification, routing... and the small models that agents call 100 times for every 1 call to a big one.
A GPU's career is a walk down that ladder. The card that trains today becomes a great inference box in 2 years, then a perfectly good embeddings & batch machine 2 years after that. Each step down it earns less per hour than it did at the top - but it keeps earning.
And this isn't theory. The buyers we talk to at B3IQ already operate this way - they don't run one machine type, they run multiple rigs for different workloads, and when a new rig comes in the old one doesn't get sold. It gets re-roled: pushed down to inference, routing, and the lighter jobs that don't deserve new silicon.
Don't take Jensen's word for it
The receipts are public:
- Azure retired its V100 instances in September 2025 - roughly 7.5 years after that chip launched. The K80/P100 generation before it? 7-9 years of service.
- The T4 is a 2018 card and it's still earning rental income on GPU marketplaces today.
- CoreWeave has had H100 capacity from 2022 rebooked at 95% of original pricing.
- The companies with the best data on this (the ones operating millions of these things) run 5-6 year depreciation schedules... and keep finding the hardware outlives them.
Old GPUs literally get faster
This is the part the iPhone comparison gets exactly backwards.
Your phone gets slower over its life as the software bloats. A GPU gets faster, because the software underneath it keeps improving. 4-bit quantization roughly doubled what a given card can serve. Better attention kernels, continuous batching, speculative decoding - all compounding on top.
A 2021 card running a 2026 serving stack pushes meaningfully more tokens per second than it did the day it was new. Wild, but true.
That's the actual substance behind Jensen's CUDA line. Yes it's marketing, but it's also real: the install base keeps inheriting the improvements, so the floor under old hardware keeps rising.
Okay, but doesn't Burry have a point?
Honestly - yeah, partially.
Useful-life assumptions are a choice, and stretching them flatters earnings, so some skepticism is deserved. And two real forces do push against long lifecycles:
- Perf per watt improves every generation. Electricity is a growing share of your cost per token, so an old card can lose on operating cost even while it still works fine.
- The frontier premium decays fast. H100 rental rates went from $7-10/hour at launch to a few bucks by late 2025.
So the honest curve looks like this: the top falls fast, and the tail is long & fat.
The mistake (and it's the market's default mistake) is pricing the entire asset off the top of the curve and calling the tail zero.
So what do you do with this?
Well - it depends on whether you rent or own.
If you rent compute, none of this upside is yours. You pay the going rate for whatever tier you're using, forever, and the option value of aging hardware goes to your landlord. The cloud providers understand truck economics perfectly. Why do you think they keep the trucks?
If you own the machine, the whole career is yours. The box that serves your flagship model in year 1 serves your embeddings, agents & batch jobs in year 4, and can take on outside work in the gaps. When we say B3IQ machines are yours, this is a big part of what we mean - you keep the shiny first act AND the unglamorous, profitable tail that everyone else writes down to zero.
Six years in, the A100 fleet is booked solid.
Compute good. Lifecycle long.