AI isn’t a bubble. This one runs on contracts.
AI isn’t a bubble. Bubbles run on stories. This one runs on contracts.
THE AI STACK, Q4 edition. Scarcity migrates up the stack: GPUs in 23, memory now, power next.
ENERGY:
$BE
$EOSE
$TE
$VST
The layer you can’t fab. Substations take years.
COMPUTE:
$IREN
$NBIS
$CIFR
$WULF
Compute is revenue, per Jensen. Cheapest per buildout dollar: IREN.
SILICON:
$NVDA
$AVGO
$AMD
$108B guided next quarter. Zero China in the number.
MEMORY:
$MU
$SNDK
$SKHY
The layer with pricing power right now. HBM sold out.
PHOTONICS:
$CRDO
$LITE
$COHR
$AAOI
Where the bottleneck went after GPUs. Every rack talks at 1.6T.
PHYSICAL AI:
$OUST
$AMBA
$VPG
$TSLA
The next demand clock. The most 2027 in the whole map.
SPACE:
$ASTS
$RKLB
$PL
Different clock from AI capex. That’s the point.
SOFTWARE:
$PLTR
$NOW
$TEM
Who keeps the profit. Re-rates last and hardest.
Q4 setup: tape unpinned after OpEx, midterm seasonality, and contracted ARR converting to revenue on a commissioning schedule. Customers are prepaying for capacity that doesn’t exist yet. The receipts are in the filings.
Which layer runs short next?
