A reusable instrument, not a one-shot. When an AI lab or compute company files a public S-1 (or an S-1/A that adds financials), run this to convert the filing into thesis-routed datapoints fast, instead of scrambling. The S-1s are the first audited demand-side disclosure our bottleneck theses have ever had — every prior read was supply-side physics + inferred demand.

Ownership: cross-cutting. This serves the bottleneck perspectives (ai-power-bottleneck, ai-power-delivery, optical-supercycle, memory-supercycle) and the counterparty/IPO threads. The AI mega-IPO Fast Entry lane is one trigger, not the owner. Cross-link out; never relabel a perspective datapoint as lane-owned (and vice-versa). The flow/float/lockup rows belong to the ai-mega-ipo-fast-entry-liquidity-regime lane; the demand/economics rows belong to the bottleneck perspectives.

When to run

Trigger: a qualifying public S-1 / S-1/A lands. Live watches:

• OpenAI — confidential draft ~2026-05-22 (secondary); public S-1 pending.
• Anthropic — confidential draft filed 2026-06-01 (primary company statement); public S-1 pending.
• xAI — bundled inside the SpaceX S-1/A (CIK 1181412); read xAI segment disclosures there.
• Any neocloud / compute-co S-1 (CoreWeave-class) on the same theses.

Do NOT run on a confidential draft (not public, not readable). Wait for the public S-1 or an S-1/A that adds the financial statements.

The crosswalk

Each S-1 section maps to the datapoint to pull, the thesis it informs, and where it routes. Pull every row; "not disclosed" is a valid, useful finding (often the most informative — see Cerebras row 1).

The bull/bear pivot (row 3 is load-bearing)

Rows 1–2 + 8 + 11 say how much compute/power is being funded (the Capital-Formation Accelerant bull case → bottleneck theses harden). Rows 3–4 say whether the marginal compute is profitable (the Burn-Reveal bear case → the AI complex de-rates). A demand spike with deeply negative unit economics is NOT bullish — it's the fuse. Always read rows 1–2 and rows 3–4 together and state which dominates. This is the single most important judgment the crosswalk forces.

Routing protocol

1. Emit a dated read artifact per filing: YYYY-MM-DD--s1-read.md(orinvestigations/` if it's mostly a thesis question), citing the S-1 by EDGAR URL, with the 12-row table populated.
2. Route each row to its owning perspective via a log-entry (the desk's own tooling) — NOT by editing the perspective README. Use --per-ticker for named suppliers/customers.
3. Watchlist gaps: any named supplier/customer not in coverage → file a coverage-decision row in TASKS-FOLLOWS.md (don't auto-add).
4. Flow rows (3-margin, 12-float/lockup) route to the AI-mega-IPO lane, not the perspectives.
5. Run feeds + closeout (research:feeds, then research:closeout --since <date>).
6. Keep this playbook's "Known tells" and "Limitations" current as each filing teaches us something (living-document).

Worked example — Cerebras Systems S-1 (validated 2026-06-02)

Ran the crosswalk against the operative S-1/A (filed 2026-05-11, FY2025 financials; CIK 0002021728) and the original 2024 S-1. This is the validation run; it both proves the instrument and seeds the OpenAI/Anthropic reads.

What it told us (3 theses):

• ai-power-bottleneck — confirmed from the demand side. The catalyst is a

  $20B OpenAI deal denominated in MW (750 / 250), with the filing naming power, not chips, as the binding constraint. Strongest single datapoint.

• optical-supercycle — Cerebras is a COUNTER-case. Wafer-scale keeps compute+memory on-die (44GB, 21 PB/s) to avoid the switches/cables/optics a GPU cluster needs. It's a TSMC-5nm + packaging story, not an optics/HBM story. (Important: a compute-co S-1 can weaken a thesis — route the negative.)
• circular-financing / counterparty — MORE important post-IPO, not less. The canonical G42 loop died in CFIUS, but the pattern migrated to OpenAI (customer + lender + ~$0-strike warrant) on a still-62%-single-customer, UAE-concentrated base under revocable export licenses.

Biggest non-disclosure: no hard compute-footprint numbers — you get a forward 750 MW and GM%, but cannot back out per-system unit economics or installed capacity. Expect the same gap in OpenAI/Anthropic; pre-commit to flagging it.

Known tells (update as filings teach us)

• MW-denominated deals (Cerebras 750 MW) = the demand curve in its rawest form → ai-power-bottleneck primary receipt.
• A customer that is also a lender/warrant-holder = circular-financing red flag → counterparty thread, quantify the triple.
• Gross margin guided lower right after a capex spike = burn-reveal risk → lane bear hypothesis.
• "Not disclosed" on unit counts is itself the finding — it caps how hard you can stress-test the capacity ramp.
• A wafer-scale / custom-silicon filer can WEAKEN optical/HBM — don't assume every AI S-1 confirms every bottleneck; route negatives too.

Limitations / traps

• Confidential drafts are unreadable — wait for the public S-1/A with financials.
• Press valuations/float/raise are NOT filed facts until the offering blanks fill (cf. SpaceX S-1/A offering still blank as of 2026-06-01).
• One filing is one counterparty's view; cross-source against the other side (a lab's cloud contract should also appear in the hyperscaler's disclosures).
• GAAP net income can be distorted by forward-contract / warrant accounting (Cerebras FY25 +$237.8M GAAP vs −$75.7M non-GAAP) — read the non-GAAP/cash burn.
• Keep flow rows (margin pivot, float/lockup) in the IPO lane; keep demand rows in the bottleneck perspectives. Don't let the lane annex disclosure-reads.

Cross-links

• AI mega-IPO lane (trigger + flow rows): 2026-06-02-ai-mega-ipo-fast-entry-liquidity-regime (Hypothesis D2 "Capital-Formation Accelerant" + "Unit-Economics / Burn Reveal").
• Counterparty taxonomy (graduating-ticker home for pre-IPO entities): 2026-05-06-ai-counterparty-taxonomy-cerebras-trigger.
• Anthropic child track: 2026-05-27-anthropic-ipo-compute-financing.
• Trigger task: TASKS-RESEARCH "Anthropic + OpenAI S-1 filing watch".