One of Cassandra Unchained’s subscribers received this missive in his email and asked me about it. I am not sure of the source.
Rentability is not my claim, never has been. Depreciation is not a bet on when a chip stops working. It is the recovery of capitalized sunk cost over a defined window of earnings at the frontier, set against the terminal value. A chip can rent and still depreciate very fast economically. I never said the A100s would stop functioning.
In early 2025, Amazon (AMZN) shortened a subset of server lives to 5 years from 6 years and explicitly cited the pace of development. This caused additional expense across 9 months of $677 million. During the same quarter, Meta (META) extended its useful server lives to 5.5 years, reducing the depreciation hit to net income by $2.9 billion. Same hardware but opposite conclusions by two of the biggest.
I covered this in Part 2 of the Heretic's Guide to AI's Stars - the Depreciation Problem.
Depreciation cannot move two directions at once if it were a measurement of something. It is an economic lever, not a physical measurement.
Satya Nadella at Microsoft (MSFT) said in an interview on the Dwarkesh Patel podcast, “I didn’t want to get stuck with four or five years of depreciation on one generation.”
NVIDIA’s Jensen Huang at GTC said when Blackwell ships in volume, “you couldn’t give Hoppers away.”
I discussed this in The Supply-Side Gluttony Recurrence.
The biggest buyer and the biggest seller of these chips both admitted, pointedly, that the frontier moves faster than the depreciation schedules used by the big hyperscalers and neo-clouds.
The missive forwarded by the subscriber treads alongside arguments Coreweave CEO and Nvidia Neo Pal Michael Intrator has sent my way over the last 6 months or so.
Across multiple CNBC and Jim Cramer appearances, Intrator told the same story - his contracts run five years, his A100s stay fully booked, that a batch of H100s coming off contract were leased right back at 95% of the original price. He told Jim Cramer his customers are willing to rent GPUs for six to seven years.
Intrator’s consistent message is that Nvidia’s chips get rented out on successive contracts so the idea depreciation schedules should be less than 5 or 6 years is ignorant.
Intrator also unknowingly argued against this point of his at the Fortune Brainstorm AI conference when he attempted to describe circular financing as really just good logistics securing a volatile supply chain given compute and power take years to build.
Circular is “the incorrect way of looking at it.
It’s a lot of companies working to address an imbalance that is distorting the globe
.”
“The primary constraint is a physical bottleneck associated with getting the most performant compute into the hands of the most cutting edge players.”
“
The reasons that you have challenges in delivering that compute is because of policy…because of infrastructure…because of energy.
You do that by working together.”
The message here is that there is a physical bottleneck due to a “distorted" global supply chain as well as many policy, energy, and other factors.
I agree. I discussed that global supply chain in Nvidia Ratchets Up the Risk.
What is happening now is
not
temporary. It is no export shock. It is not even external. This is coming from within the business plan.
The headline here is the Cash Conversion Cycle is extending permanently along with Days Inventory Outstanding.
This new reality reflects a deliberate decision to lock up supply chain capacity further than Nvidia has ever done before.
While NVIDIA sorts its supply chain, the forward constraint is, as Intrator says, the deployment of these frontier chips. There is not much of a bottleneck in redeploying the old ones. In fact, the old ones are getting work that they would otherwise not have received precisely because of the bottleneck at the frontier.
Two days ago, Analytics India Magazine reported on an Air Street Capital study that made fewer headlines than I thought it would.
More than 95% [
94% by my calculation
]
of the world’s announced
NVIDIA Grace Blackwell GPU
capacity has
yet to be deployed
, according to Air Street Capital’s latest
State of AI Report Compute Index
, which tracks major AI compute deployments worldwide.
The index estimates that 100,128 Grace Blackwell GPUs (GB200/GB300) had been deployed as of July 1, compared with an announced pipeline of around 1.66 million GPUs, meaning less than 5% of the announced capacity is operational. Another 308,640 GPUs are in the installation phase.
Included in the report was this graphic regarding Grace Blackwell GPUs deployment as of July 1, 2026.
The full article is available on Michael Burry's Substack.
Or, Please click on this direct link.
