Screening the AI Picks-and-Shovels Trade: Six Layers, Two Scores, and the Metrics That Lie

The thesis, and why layering it matters

Nobody knows which AI model or application company ultimately wins, but they all buy from the same short list of suppliers. That observation is common. What makes it actionable is decomposing the supply chain into layers and noticing that every layer is structurally concentrated — a monopoly, a duopoly, or a three-player oligopoly with high switching costs.

1. Compute — merchant GPUs and custom accelerator ASICs.
2. Foundry — leading-edge fabrication plus advanced packaging capacity.
3. Memory — high-bandwidth memory, the tightest physical bottleneck in the chain.
4. Equipment — the wafer-fab tools that make everything above possible.
5. Networking and optics — moving data between accelerators at rack and cluster scale.
6. EDA and power — the software every chip is designed on, and the electrical and cooling infrastructure that turns capex into working data centers.

Layering is not presentational. It forces the screen to compare within a layer rather than across the whole set, which is the only way a 45x-forward equipment maker and a 7x-forward memory maker can be evaluated without the comparison being meaningless.

Two scores, deliberately coarse

Each name gets a quality score and a cheapness score, both 0–5. Coarse scoring is a choice: the inputs do not support decimal precision, and a rubric that outputs 3.7 invites false confidence in a judgment that is really 'good, not great'.

• Quality weighs moat durability (monopoly and oligopoly versus contestable), gross and operating margin, ROIC, revenue growth, balance-sheet strength with net cash beating net debt, and customer-concentration risk — the last one mattering enormously here, since much of this chain sells to a handful of hyperscalers.
• Cheapness weighs forward P/E, growth-adjusted P/E, forward multiple against the stock's own five-year history, and free cash flow yield where available.
• Growth adjustment is weighted heavily on purpose: a fast compounder at 25x can be genuinely cheaper than a slow one at 15x, and a screen on headline multiples alone systematically buys the slower business.

Where the metrics lie, part 1: derived PEG

The verification pass produced the most useful finding in the whole exercise, and it was a data-quality finding. PEG ratios for a substantial share of the names — including several in equipment and the entire EDA and power layer — turned out to be provider-derived or otherwise uncorroborated. Different sources disagreed, and the disagreements were not small.

The failure mode is specific: PEG depends on a forward growth estimate, and where trailing earnings are at a trough or distorted by an acquisition, the denominator is an assumption dressed as a measurement. Two names in this set had exactly that profile — one post-acquisition, one at a cyclical earnings trough — and their PEGs were the most attractive-looking in their layers.

• Treat PEG as directional, never decimal-precise, and say so in the output rather than carrying two significant figures into a ranking.
• Mark every derived figure that could not be independently corroborated. The published table flags them inline rather than footnoting them once.
• Where a figure is simply unavailable, write 'n/a'. Inventing a plausible number is the single easiest way to corrupt a comparison table, and it is undetectable later.

Where the metrics lie, part 2: peak-cycle memory

Memory is the most cyclical corner of semiconductors, and the snapshot caught it at or near a cyclical peak: roughly 85% gross margin at one manufacturer, 72% operating margin at another. Those are not sustainable levels; they are what the top of a memory cycle looks like.

The consequence for screening is counterintuitive. Trailing P/Es looked reasonable at 23–25x and forward P/Es looked extraordinary at 5–7x, with growth-adjusted ratios below 0.1. A mechanical cheapness score ranks these first in the entire universe. But those forward multiples run on peak-cycle EPS, and the classic memory trap is that the multiple inverts when pricing rolls over: earnings collapse faster than price, so a 6x stock becomes a 40x stock without moving.

The handling was to judge memory on mid-cycle earnings rather than the current print, cap its cheapness score with an explicit cyclicality flag, and state plainly that position sizing around cycle timing matters more than any point multiple. The bull case — that HBM wafer intensity, secular demand, and oligopoly discipline stretch the cycle — is stated as a thesis rather than folded into the score as if it were a fact.

What the screen produced

Roughly 30 names across the six layers, scored and sorted by combined quality plus cheapness. All figures are a July 1–2, 2026 snapshot; several of these stocks moved 7–12% in a single session during the research window, which is itself the best argument for treating any table like this as perishable.

The structural finding was that quality and cheapness are strongly anticorrelated in this chain, and the exceptions are where the analysis is worth doing. The highest-quality names in equipment and EDA carried the worst growth-adjusted valuations; the cheapest names sat in memory, where cheapness is a cycle artifact. Only a handful scored well on both, and each of those had a specific, nameable reason the market was discounting it.

The disclosure pattern worth copying

Independent of the subject, this write-up used a disclosure discipline I would apply to any analysis built on imperfect inputs.

• Snapshot date stated at the top and repeated at every table, with an explicit warning that the figures will be stale almost immediately.
• Every corrected or refuted figure kept in the output alongside the correction, rather than silently replaced.
• Uncertainty marked inline at the specific cell, not aggregated into a general caveat nobody reads.
• Missing data written as 'n/a' rather than imputed.
• The single largest risk to the thesis — cycle timing in memory — stated in the methodology section rather than buried in a per-name footnote.