If you allocate to the AI upstream — chips, HBM, data centers, power — here’s a sorting knife. “Upstream scarcity” is a phrase worn smooth by overuse; cut it open and you find three things with completely different shelf lives, deserving completely different valuations.

The upstream isn’t one scarcity — it’s a stack of them. Sort by cause:

Engineering scarcity (chip design): scarcity built from ecosystem, interconnect technology, and accumulated engineering depth. Near-total monopoly (roughly 90% of the training market by widely cited industry estimates), enormous pricing power (~70% gross margins) — but it depends on staying ahead every generation, while its biggest customers spend tens of billions a year building replacements. The signature of engineering scarcity: impregnable until caught, no cushion after.

Cyclical scarcity (HBM and memory): sold out today, record margins — but the memory industry’s century-old law holds: excess profit invites excess capacity, and excess capacity digs the next trough. This scarcity has a shelf life, and the shelf life is printed in the capacity-expansion announcements. The current cycle began in 2023–24; by historical cadence, do the arithmetic yourself.

Physical scarcity (advanced process nodes, power): one company on Earth supplies EUV lithography; new generation and grid interconnection take five to ten years, with permitting queues measured in years. No “open-source” or “build-it-yourself” mechanism can route around it — you can design your own chip to replace a supplier; you cannot design your own physics. Power is also the only layer in the whole chain whose scarcity strengthens over time: demand keeps outrunning supply’s response speed, and the gap is widening.

Different erosion mechanisms deserve different valuations — hence the title: physics gets duration, cycles get trades, contracts get discounts.

Wait — what’s the fourth one?

Contract-Manufactured Scarcity: The Upstream’s False Bulk

Part of today’s upstream “shortage” isn’t physical — it’s manufactured by contract structure: labs sign long-term capacity deals beyond their ability to pay in order to lock in compute; neoclouds need more contracts as loan collateral; clouds stockpile capacity commitments for order-book visibility.

Which produces an absurd accounting reality: the same batch of GPUs can appear simultaneously in three companies’ “contracted capacity” — the lab’s commitment, the neocloud’s contract, the cloud’s expansion justification. Apparent demand is inflated by double and triple counting.

Physical scarcity is immune — its scarcity lives on the supply side. But capacity-type scarcity (GPUs themselves, data-center space) carries a contract bubble: once the financing loop winds down, apparent shortage can flip to apparent glut within quarters, with price swings far beyond what organic demand justifies. That is exactly how fiber went from “never enough” in 2000 to “under 5% lit” after the bust.

So interrogate every upstream story with two questions: who are the counterparties behind your “contracted capacity”? And where does their money come from?

Three sentences to close. Physical scarcity (process nodes, power, interconnection assets) earns the longest duration and the highest valuation tolerance. Cyclical scarcity (HBM) is a trade — exit on the expansion announcements. Engineering scarcity (chip design) pays today but charges a permanent attention tax: customer-built silicon progress, second-supplier share, and the inference-to-training spend ratio — three gauges you never stop watching.

The “AI upstream scarcity” you own — which kind is it? And how much shelf life is left? Comments open.

Data sources (verified, June 2026): chip share and margins (widely cited industry estimates, pending primary anchoring); AMD–OpenAI 6GW second-supplier signal (official press release + SEC, Oct 2025); fiber lit-rate and bubble history (public record).

— From Chapter 5 of a book in progress, working title The Deflation Sandwich

#AI #AIInfrastructure #DeflationSandwich