Something to think about this weekend:
Usta Stockman
New: Today, 8:28 AM
Comments (3.23K)
|
|
I appreciate the upgrade logic, but I think you’re looking at the 'AI Factory' through a rearview mirror. You’re citing record Q3 revenues and 73% margins as a reason to buy, but those are trailing indicators. If you look at the forward-looking ROI for Nvidia’s customers, the picture is much darker.
1. The Rental Reality Check:
While you’re calling for a $245 price target, the actual service price for these chips is in a free-fall.
H100s have crashed from $8/hr to as low as $1-$2/hr ([Cybernews]( cybernews.com/...
Even the B200s you mentioned are already sliding toward $2.80/hr (( www.trendforce.com/...
At these rates, the 'Neoclouds' and startups borrowing billions to buy GPUs are effectively underwater. They cannot service their debt at $2/hr.
2. The Nvidia Tax:
You mentioned Google TPUs are a threat and you're right. The market is pivoting to 'Good Enough' inference. Companies are tired of the Nvidia Tax' (the high cost of GPUs plus the 'Micron Tax' for external memory).
Jensen’s $20B Groq deal isn't 'expansion' - it's concession. By licensing Groq’s on-chip SRAM architecture, Jensen is admitting that the 'General Purpose' GPU is too expensive for the inference era. He’s trying to lower the Bill of Materials (BOM) by cutting out the memory suppliers (Micron/Samsung) so he can keep his margins high while the rental market collapses around him.
3. The 2007 Warning:
This mirrors the 2007 housing bubble almost to a fault. We have record 'originations' (GPU sales) fueled by 'circular debt' (VCs funding startups to buy GPUs). But the 'rent' (AI revenue) isn't covering the 'mortgage.'
Jensen is acting like the only banker who sees the crash coming. He’s using his $61B cash pile to buy the 'Inference Moat' (Groq) and colonize the memory stacks (HBM4 with SK Hynix) so he’s the only one left standing when the 2026 'reversion to the mean' hits.
Bottom Line: When a small group of stocks propels the market into a '1-in-75-year' concentration signal, and the lead actor (Jensen) is frantically cannibalizing his own supply chain to survive a rental crash... it’s not time to buy imo
Cyclical Trade
New: Today, 9:01 AM
Investing Group
Comments (4.72K)
|
|
@Usta Stockman Excelling summary.
Usta Stockman
New: Today, 9:08 AM
Comments (3.23K)
|
| @Cyclical Trade the parallel to housing isn’t just rhetorical - the mechanics are nearly identical.
In 2005–2007, consumers borrowed under the assumption that home prices would rise forever.
In 2023–2025, corporations borrowed under the assumption that AI demand would rise forever.
The difference is scale.
Corporate borrowing to chase AI capacity has grown faster than mortgage debt did at the peak of the housing bubble. And now we’re seeing the same early warning signs:
- GPU rental rates collapsing from $8/hr to $2/hr
- B200s already under $3/hr
- Utilization falling
- Token-pricing compressing
- Balance sheets deteriorating for the GPU borrowers
Just like housing, the problem isn’t that prices are falling - it’s that they’re falling while the debt remains fixed. Once the revenue from inference can’t service the loans used to buy the GPUs, the math breaks. That’s exactly what happened in 2007: the cash flows stopped covering the leverage.
The Groq deal only reinforces the point. Jensen is trying to solve the efficiency problem without touching the margin problem. But if the Nvidia Tax stays intact, all he’s really doing is repackaging the same cost structure into a different bill. The underlying debt burden doesn’t go away.
When the economics stop clearing, the correction isn’t gradual. It’s sudden, nonlinear, and confidence-driven. Housing peaked in 2007, and the next four years were the worst economically in decades. The AI infrastructure cycle is showing the same stress fractures.
more comments at seekingalpha.com
|
