Question: AI-compute-shortage flagship cohort EDGAR financial audit — 13 names across 4 layers (chips/CPU, optical, power, nuclear) Verdict: open

What we're asking

Do the financials confirm the AI-compute-shortage thesis (THESIS.md Pillar 1), or is the cohort priced on narrative? Pulled SEC EDGAR XBRL for 13 picks-and-shovels (memory already done separately) and fanned out a 4-layer parallel audit — the question per name: does the filed financial reality support the re-rate, and what does it show that the price doesn't?

What we found

Verdict: the thesis CONFIRMS in the financials across all four layers — real revenue acceleration (+28% to ~3x YoY), genuine (mostly recent) profitability — but the cohort is largely priced for it (most names CONFIRM-BUT-PRICED after +150–255% 3M runs). The signal is in the exceptions. (Prices / RSI / 3M / 52wk figures throughout are quoted verbatim from research/market-engine/data/summaries/*.json, 2026-06-25 — price truth; all financials from EDGAR XBRL on disk.)

Three places the financials say something the price doesn't:

• BE (Bloom Energy) — the cleanest inflection. Turned GAAP-profitable, operating-driven and cash-backed (operating income ≈ net ≈ OCF, all ~$73M), sustained (FY2025 full year + Q1'26), on +134% YoY revenue, with a growing $2.49B cash war chest. Not a one-time item — a structural cross-over. CONFIRM.
• AVGO — fundamentals LEAD the tape. ~48% operating / 42% net margin (post-VMware drag rolling off), OCF > net — yet RSI 44, weak-down, −23% off highs. The only name where the chart trails the financials. (Coverage note: not in the cpu-shortage key_tickers.)
• LEU (Centrus) — the "burn" is a war chest, not a burn. $1.87B cash is financing-built (DOE HALEU inflows/raises), not operations-drained: profitable net income + positive full-year OCF every recent year; the −$35M Q1 OCF is a working-capital swing. Bullish optionality for the HALEU-monopoly build — but the tape is rolling over (−62% off high), so accumulate-on-confirmation, don't chase. CONFIRM (structure) / pre-confirmation (setup).

The pattern that matters — don't trust GAAP net on the chips. MRVL ($34.5M net on $2.42B rev) and AMD (14% margin) look thin on a screen but are dragged by acquired-intangible amortization + SBC (Inphi/Innovium, Xilinx); OCF runs 2–18× net. Earnings are fine; GAAP optics aren't — same lesson as the quantum warrant-gains: read operating cash flow, not net income. Caveat for MRVL/ALAB/CRDO: the entire profitable era is ~one fiscal year old (all lost money operating through FY24/FY25), so the +180–255% 3M re-rate rests on an unproven margin track record.

The one real risk flag — COHR. Net income $191M but Q3 operating cash flow NEGATIVE (~−$94M); a ~$925M inventory+receivables ramp build (9-mo OCF/NI = 0.018). Earnings real, but cash hasn't followed — and the August Q4 print is the gate that turns that build into operating cash or a reversal. Plus $3.2B II-VI merger debt. By contrast LITE is the higher-quality optical expression: OCF/NI 1.71, $3.2B cash, and the laggard price (−22% off high) — CONFIRM without the priced-in caveat.

Utility caveat (power layer): CEG/VST GAAP earnings are hedge-mark-distorted — CEG printed positive net but negative OCF in FY2023 and FY2024 (collateral postings); read OCF. TLN is the cleanest IPP (GAAP loss FY2025 but +$704M OCF = fresh-start-accounting artifact, not a burn). CEG's −34% de-rate is PJM/FERC valuation unwind, not demand collapse.

Per-name verdicts

Verdict + reasoning

Pillar 1 is financially sound, not narrative — every layer's picks-and-shovels show real, accelerating, mostly cash-backed economics. The actionable edges: (a) BE is the cleanest fundamental inflection; (b) AVGO and LITE are the two names where price lags confirmed fundamentals; (c) LEU is bullish optionality awaiting tape confirmation; (d) COHR carries the one genuine cash-quality flag (Aug Q4 gate). The cohort-wide discipline: read OCF over GAAP net (amortization on the chips, hedge marks on the utilities), and respect that most of these are one fiscal year into profitability after huge 3M runs.

Engine flags surfaced (filed): (1) the EDGAR key-financials extractor returns all-nulls for IFRS/40-F filers (CCJ has the IFRS XBRL on disk; the extractor only maps US-GAAP) — sharpened into the EDGAR v2 backlog (ops/tasks/TASKS-ENGINE.md). (2) LEU summary vs the perspective DATA-TABLE drift (live strong-down $174 vs table weak-up $191, data_as_of 2026-06-22) — a Phase 0b refresh reconciles it.