The framing that matters

Our research vocabulary today is perspective (thesis), watchlist (tickers), deep dive (per-ticker snapshot), profile (human author). That carves humans-and-public-companies cleanly. It does NOT carve corporate entities that aren't tickers yet (Cerebras, Anthropic, OpenAI pre-IPO) or entities that may never become tickers (Stargate the project, xAI, consortiums like Project Jupiter).

The question this thread is sitting with: what's the right home for AI counterparties that show up across multiple perspectives but don't fit any current type?

The Cerebras IPO on/around 2026-05-15 (9 days out) is the forcing function — not the design problem. We need to decide where the Cerebras S-1 read lives this week, and the answer cascades into Anthropic/OpenAI when their S-1s drop later this year.

Decisions already made in the conversation that opened this thread

(1) CGEH is a sub-thesis inside an existing perspective, not a new perspective

ai-power-bottleneck already covers behind-the-meter data-center power. CGEH (microturbine cogen) sits alongside Bloom (SOFC), GEV (gas turbine), OKLO/SMR/CEG (small nuclear) as competing sub-theses for the same problem. Spinning a new perspective per sub-thesis would shred the macro frame into ticker-shaped pieces. Right home is the perspective's log.md + key_tickers.

This was the easy half of the taxonomy question. Recorded so we don't re-litigate it.

(2) Cerebras / Anthropic / OpenAI are graduating tickers, not permanent entities

These three are all expected to IPO within months (Cerebras ~2026-05-15, Anthropic + OpenAI later this year per user knowledge). Building a permanent entities/ type for things about to collapse into normal coverage is over-engineered. They need a transition path that expires.

Proposed thin type: pre-IPO holding pen at inputs/pre-ipo/<company>.md (or similar) — explicitly designed to expire on listing day. Holds:

• S-1 read (suppliers, customers, risk factors, financial summary)
• Cross-references to existing perspectives, deep dives, watchlists where the entity is named-but-untracked
• Listing-day playbook (watchlist add, perspective log entry, OHLC fetch)
• Known IPO-date or filing-watch expiration

On listing day → ticker activates into normal coverage; holding-pen file becomes archive with a one-line "graduated to TICKER on YYYY-MM-DD" pointer.

Not built yet. Cerebras drives the design — running the Cerebras flow once will tell us what the holding-pen file actually needs to hold.

Deferred questions

Stargate / xAI / consortiums — the genuinely permanent entities question

Stargate is a project, not a company. xAI may IPO eventually but has no filed S-1 today. Project Jupiter is Oracle's 2.8 GW Bloom Energy program, named in our perspectives but not a tradeable thing. These DON'T graduate into tickers on any known timeline — they're permanent watched things.

If the volume of permanent-entity-shaped artifacts grows (>5 of them across our perspectives), they earn their own type. Today we have maybe 3-4 candidates. Defer the decision. Revisit when the population forces it OR when we run the Cerebras pre-IPO flow and have a working pattern to extend.

"Should the holding pen be inputs/pre-ipo/ or a top-level type?"

inputs/ is already overloaded (tweets, podcasts, articles, expert-notes, repos, profiles, triage, etc.). Adding inputs/pre-ipo/ keeps the naming local. But pre-IPO research isn't really a capture of an external signal the way a tweet or podcast is — it's an internal artifact. Might fit better at research/pre-ipo/ or research/graduating-tickers/.

Defer until we write the Cerebras file and see what shape it actually takes.

The Cerebras May 15 forcing function — concrete plan

This is the action that tests the taxonomy. Two-phase:

Phase 1 — Pre-listing (this week, before 2026-05-15)

• Pull the most recent amended S-1 from EDGAR. Cerebras Systems Inc, CIK 0002021728. URL: https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0002021728. Free; no API key. Use a User-Agent string per SEC's request guidelines.
• Extract from the S-1:
  • Supplier list (the open VICR question — "wafer-scale AI engine maker" lead Gen 5 customer — closes if Cerebras names VICR)
  • Customer concentration (top customers % of revenue — G42 was the historical concentration risk that delayed the original IPO)
  • Risk factors section (CFIUS posture, customer concentration, supply chain)
  • Financial summary (revenue, gross margin, cash burn) — most recent fiscal period vs prior
  • Use of proceeds (data-center capex? more wafer-scale fab partnerships?)
  • Underwriters + proposed ticker symbol (probably CBRS or similar on Nasdaq)
• Cross-reference against our existing universe:
  • VICR _beta/deep-dives/2026-05-01-vicr-deep-dive.md — close the open question if Cerebras names VICR as a vendor
  • ai-power-bottleneck perspective — Cerebras's own behind-the-meter / colocation choices
  • optical-supercycle perspective — wafer-scale interconnect requirements drive optical content per rack
  • inputs/profiles/ — any tracked human author who's connected to Cerebras (Andrew Feldman the CEO, Sean Lie the CTO, board members)
• Write the holding-pen file. Path TBD — probably cerebras or inputs/pre-ipo/cerebras.md. Don't over-design; write it once and let the shape teach us what the type wants.
• Optional: build lib/edgar.ts. If the manual SEC fetch is annoying, codify it as a thin EDGAR client (free, no key, by CIK rather than by ticker). Adds an EDGAR fallback to picks/financials.ts for tickers Polygon doesn't cover (foreign, OTC, pre-IPO). Half-day spike. Not blocking — defer if Phase 1 manual fetch is fine for one company.

Phase 2 — Listing day playbook (2026-05-15 or whatever date the IPO prints)

10-minute runbook — pre-position now so it's mechanical on the day:

• Confirm assigned ticker symbol (CBRS or whatever)
• Add ticker to ai-power (and possibly optical-supply-chain.json depending on customer concentration)
• `cd stonks && the desk's own tooling — populate OHLC cache
• _beta/perspectives/ai-power-bottleneck/log.md — entry: "Cerebras IPO graduates from pre-IPO holding pen. Ticker [TICKER] now tracked. Key reads from S-1: [...]. Implications for [...] sub-thesis."
• If Cerebras named VICR as supplier in S-1: VICR deep-dive log entry confirming the lead Gen 5 customer
• Mark holding-pen file graduated: TICKER (2026-05-15), leave the file in place as historical S-1 read

May 15 is also the Q1 2026 13F deadline (separate TASKS line). That's two cluster events same day — pre-position both this week so the Friday isn't reactive.

Anthropic / OpenAI follow-on (date-anchored when S-1s drop)

Same Phase 1 / Phase 2 path. Triggers:

• S-1 filing watch: when Anthropic or OpenAI files an S-1, open the holding-pen file (use the Cerebras file as template by then). Could automate via an EDGAR full-text-search subscription on company-name once lib/edgar.ts exists.
• Underwriter watch: S-1 lead underwriter naming is sometimes the earliest leak (Goldman / Morgan Stanley / JPM filings show up in their filings before the company files). Lower priority unless the EDGAR pipe makes it cheap.

No urgency until there's a filing. Just naming the path so it's not reinvented.

Cross-references — what's already written

The substance of this conversation is partly captured in TASKS.md (concrete actions) and partly captured in _beta/inputs/ (the trigger event). This thread is the architectural-reasoning layer that ties them together so future-Claude/future-user doesn't re-derive it.

What graduates this thread

This thread graduates when either:

1. We've run the Cerebras Phase 1 + Phase 2 once and the pre-IPO holding-pen pattern is durably stable (a path exists, a template file exists, a graduation step exists), OR
2. We decide explicitly that the holding-pen pattern doesn't pay rent and we kill it (everything just goes in TASKS lines and _beta/inputs/).

Anthropic + OpenAI graduations are NOT blocking for this thread — once the pattern exists, those become routine applications, not architectural questions.

Open questions for next session

• Does research/pre-ipo/ or inputs/pre-ipo/ feel right? (decision deferred until first file is written)
• Build lib/edgar.ts proactively, or stay manual until pain shows? (default: stay manual for Cerebras, codify if Anthropic/OpenAI both file the same week)
• Stargate / xAI / consortiums — separate thread when population grows, or fold into this one?

---

Status: opened 2026-05-06 in conversation about @longviewres CGEH report → invezz DGXX article → Cerebras IPO trigger → taxonomy strain. Active until the Cerebras Phase 1+2 runs.