Methodology and limitations

How the numbers are made, and where they are blind

Source SEC Form DFilings → filings · raises · issuersBuilt

Source

All data comes from Form D notices filed with the SEC: the SEC's quarterly Form D data sets for 2019 Q1 to the latest published quarter, plus every D and D/A filed since then, parsed from EDGAR (each filing's primary_doc.xml, located through EDGAR's form index). When the SEC publishes the quarterly file for a quarter, it replaces the EDGAR-parsed rows. The pipeline stays under the SEC's fair-access limits (5 requests per second, a User-Agent with contact details) and caches every file.

Unit of analysis: a raise, not a filing

A "raise" is one Form D offering: the original filing plus all its amendments. Amendments usually point at the previous amendment rather than the original, so each chain is walked back to its root. Raises whose original filing predates 2019 are marked as such and left out of "new raise" counts. Where the pointer is missing, the amendment is matched to an earlier filing from the same issuer with the same first-sale date, then the same offering amount within 18 months. As of October 2026 every amendment reaches its raise through its own pointers, and a check of 450 filings against EDGAR matched CIK, form, amounts and investor counts on all of them (see the link audit). "Amount raised" is the latest reported total sold, never a sum across filings. Decreases are kept and flagged as corrections.

Startups vs funds

A filing is a fund if the issuer picked "Pooled Investment Fund", ticked pooled-fund interests, claimed an Investment Company Act 3(c) exclusion, or has a vehicle name ("a Series of", "SPV", "Syndicate", "Co-Invest"). A partnership, LLC or trust with the word "Fund" in its name ("Hyla Fund III, LLC") is a fund too, whatever industry it picked, except oil, gas and energy programs ("USEDC 2024 B Drilling Fund LP"), which are counted as other investing; corporations named "Fund" keep the other rules, since fintech startups use the word. Funds are then split by name:

Parallel vehicles ("Fund I" and "Fund I-A") are counted once. An emerging manager is one raising Fund I–III (our cut-off).

A startup raise is an operating company selling equity or a SAFE/convertible. Established companies are excluded: a CIK below 1,400,000 (EDGAR registration before about 2007) combined with incorporation more than five years before filing.

Fund managers and Form ADV. A manager's vehicles are grouped by name stem: the fund name up to its fund number, with platform SPVs grouped under the platform. Managers are then matched by name to the SEC's monthly Form ADV adviser files, which add whether the adviser claims the venture-capital fund exemption, how many private funds it reports, their gross assets, and when it began reporting to the SEC.

AI tag

Definition (written before labelling): an AI company is one whose core product is an AI or ML model, or is built on one. Mentioning AI features does not qualify.

How the hand labels applied it to hard cases:

Form D has no AI field, so evidence comes from four independent sources:

  1. The company name, including former names ("AI", ".ai", "Artificial Intelligence", "LLM"…; "Neural" and "Intelligence" count only as weak signals).
  2. Y Combinator's directory, where a matched company is tagged with AI categories.
  3. The company's homepage, fetched once, politely, with robots.txt respected. AI terms in the title, description or headline count as strong; mentions only in body text count as weak.
  4. Claude's yes/no/unclear judgment, used only for ambiguous homepages (AI mentioned only in body text). The daily build sends these to Claude (Opus 5.5) through Anthropic's Message Batches API. A "yes" counts only when Claude quotes a phrase that appears verbatim in the homepage text and contains an AI term. Local runs can use a local model through Ollama instead, under the same quote check.

The tier follows from the evidence:

Charts count strong and likely. Every company carries its evidence trail.

The sample took equal numbers of tagged and untagged companies, so raw recall overstates how many AI companies the tag finds: untagged companies are about four times as numerous, and the few AI companies among them add up. The weighted figures scale each half by its real size. The tag therefore undercounts AI: treat AI shares in the charts as a floor. Precision and recall were measured on SF Bay Area and Los Angeles companies only, so the tag's accuracy in other metros is unmeasured.

The tag looks backwards from today. Homepage, YC and model evidence describe a company as it is now, so a company that pivoted to AI in 2024 also counts as AI for its 2019 raise, which inflates early years. The Overview therefore offers a strict, time-correct alternative: AI in the company's name as filed on that raise. It misses most AI companies but can't leak. Treat the two as bounds.

Evidence collected in this build: .

Geography

Issuers are placed by the principal business address on their latest filing. The ZIP maps to a county using the Census 2020 ZIP-to-county file (the county with most of the ZIP's land). Mistyped and PO-box ZIPs fall back to the county most issuers in the same city use, then to the 3-digit ZIP prefix. SF Bay Area = San Francisco, San Mateo, Santa Clara and Alameda counties. Los Angeles = Los Angeles and Orange counties. Map coordinates come from the Census geocoder ( by street address, at the ZIP's centre).

What this data misses

Research on about 40,000 VC rounds found most were not accompanied by a Form D, with early-stage tech firms the least likely to file. Many seed rounds (often SAFEs) are filed late or not at all. Amounts go stale when issuers stop amending. Form D never reports valuation, investor names or round labels: any stage shown here is inferred from amount (under $3M seed, $3–15M Series A, $15–50M Series B, $50M+ growth). Of 25 well-known SF AI companies, more than half never filed a Form D under their own name (see Coverage gap). Treat totals as a floor, and trends as more reliable than levels.

Choices that are ours, not standards

Each is one line of SQL or config in the project repository and is easy to change and rerun.

Capital momentum score

For each venture-stage startup, five inputs are scored as percentiles within the peer group (AI or not × metro × size band), then averaged to give 0–100:

Confidence is "low" with one filing, because a missing Form D is not missing momentum.