The DOJ has spent almost a year investigating Andreessen Horowitz over two board seats. One partner sits on the board of Databricks, a $190 billion company. Another sits on Fivetran's. A 112-year-old law, the Clayton Act, says you can't help steer two competitors at once, and the government wants to know if that's what happened.
Nothing has been alleged and nothing has been found. What I can't stop thinking about is the spreadsheet that sits somewhere at every firm with a column called "board seat," and how old the information in that column is.
How the two companies became competitors
An investor put it plainly to TechCrunch: the two companies weren't rivals when the checks were written. They drifted into each other's markets over years, one product launch at a time, and then Fivetran merged with dbt Labs in June. Somewhere in that timeline two portfolio companies became competitors, and I'd bet nobody's systems flagged the day it happened.
VCs told TechCrunch the probe surprised them and that conflicts like this are unavoidable at scale. I believe them. Unavoidable is also what a regulator looks for, and the question an LP asks on the next call is the one the government is asking now: did you see the overlap forming, or did you find out from a reporter?
Where the board seat data lives
In our first nine years at my last company we put over $6 billion of financing through more than 20,000 portfolio companies, so I know where this data lives. The truth in private capital is in the documents, not in the ledger. Board seats sit in voting agreements and side letters. Competitive scope sits in charters and merger papers. Nobody re-reads any of it after closing, so a firm's map of its own board seats is whatever somebody typed into a spreadsheet three financings ago.
The spreadsheet was accurate the day it was typed, and the documents have changed several times since. A priced round closes and the amended charter changes who sits on the board, a side letter grants an observer seat, a portfolio company buys another one and its charter now describes a business that competes with something else you own, and each of those events produced a piece of paper that went into a folder and stayed there.
What the documents already say
That map should build itself. 1) Every voting agreement names who holds each board seat; roll that up and you have every seat the firm holds today. 2) Every priced round files an amended charter; each one updates the picture the day it lands. 3) The Fivetran and dbt merger agreement redefined what the combined company does; that paper existed in June, not the day it made news.
None of that requires clairvoyance. It requires reading the documents you already have, at the speed they arrive, and keeping what they say somewhere a partner can look at it on a Tuesday afternoon without asking a lawyer to re-read the data room first.
How we read the documents
Handing every closing document to one big general-purpose model and asking "who's on the board?" never worked. The model fills in gaps it shouldn't fill, one field's answer bleeds into the next, and once somebody catches a wrong director's name, nobody trusts the rest of the table.
So GoodStream doesn't do it that way. We run purpose-built extraction agents, one per field, each grounded in the shape of a specific document, so the agent that pulls the board designee off a voting agreement is a different agent from the one that pulls the authorized share count off the amended charter. Every field gets read twice, by two different models, and where the two readings differ someone on our side opens the document to that page and settles it. Click any board seat in the roll-up and you land on the paragraph of the executed voting agreement that names the director.
The a16z probe reads to me as a story about how private capital keeps its own records, and how far behind the paperwork those records run. A firm that can pull up its board seat map the day a portfolio company changes what it does has already answered the question before the LP, or the DOJ, gets around to asking it, and that is the version of transparency I spend my time on, a firm knowing its own book before a reporter does.
If your board seat map lives in a spreadsheet, you're one product launch away from learning about an overlap the same way a16z did. The voting agreements and the charters that would tell you are already in your folders, and reading them at the speed they arrive is now a job for software rather than for an associate. In the a16z case, the merger agreement that mattered had been sitting in a folder since June.
