The Market Is Not Selling AI Demand. It Is Selling AI Multiples.
How Meta Cracked the Scarcity Story
The chip rally through June was built on a single premise: there aren’t enough GPUs, HBM modules, or packaging capacity to meet AI demand. Memory makers tripled in value. Equipment suppliers doubled. The entire semiconductor complex traded as if supply would remain permanently tight.
Then Meta announced Meta Compute on July 1 — a plan to sell surplus AI training and inference capacity to enterprise customers. In a single session, Meta surged 8.8% while Micron fell 10.6%, AMD dropped 6.9%, and Nvidia slipped 1.3%. If hyperscalers have excess capacity to sell, the scarcity story cracks.
$22 Billion in Cash Deposits Say Otherwise
While the equity market repriced, the corporate spending pipeline did not pause. Amazon committed $200 billion in 2026 capex. Alphabet raised to $180–190 billion. Meta pushed to $125–145 billion. Microsoft tracks above $120 billion. Goldman projects $5.3 trillion cumulative through 2030.
These are not intent letters. Micron alone has collected $22 billion in cash deposits and letters of credit from five-year take-or-pay contracts. The infrastructure buildout is prepaid and binding. Your brokerage app says sell. The corporate treasury ledger says deploy.
TSMC Settled It Overnight
No single company sees the AI order book the way TSMC does. Every Nvidia GPU, every Amazon Graviton, every Google TPU is physically manufactured in a TSMC fab. When TSMC raises its capex by $8 billion and its revenue growth target by 10 percentage points, that is not one company being optimistic. It is every AI customer’s purchase order landing on the same loading dock.
The stock fell 1.55% after hours. The foundry that builds every advanced AI chip on earth confirmed demand is accelerating — and the market priced the cost, not the signal.
What a 4.6% Yield Does to a Growth Multiple
Our view: the selloff makes structural sense without requiring a demand collapse. The 10-year is at 4.63%. The Fed is discussing hiking. When the discount rate rises, even companies growing revenue 34% see their forward multiples compress. The market is not selling AI demand. It is selling AI multiples.
That distinction matters for allocation. In a rate-driven multiple compression, the stocks that recover first are the ones collecting actual corporate dollars — the foundries, the power-grid builders, the packaging specialists — not the ones trading on the dream of future demand.
Where the Recovery Starts
Worth watching: the names that recover first from this selloff will be the names where the purchase order is visible, not projected. The infrastructure layer collects real revenue from real contracts with real prepayments. The application layer still trades on hope. In a 4.6% yield world, the market is learning to tell the difference.
| 