01
Accumulated normalization
Thirty jurisdictions, thirty schemas, one output. The thirty-first is nearly free. A new entrant redoes all thirty — their cost is linear where ours is asymptotically zero.
SocioClimate The data moat company.
We find markets whose prices nobody publishes, assemble the record, and own it.
The claim
Compute is a commodity. The same weights, the same inference cost, and the same public benchmarks are available to anyone with a credit card. There is no defensible position in the computation layer, and there has not been one for some time.
Observation is not a commodity. A record of what things actually sold for, assembled across thirty jurisdictions that each keep it differently, gets more valuable the longer it runs, and there is only one of it. It is a network good with increasing returns, and the part of it that has already happened cannot be bought at any price.
So the rent has moved to the observation layer. Automation diffuses in proportion to how observable an asset is — how often its price is recorded, how publicly, and in how standardized a form — and is largely indifferent to available computing power. Where a dense public price record exists, decisions get ceded to models. Where it does not, valuation stays with a human whose incentives are set by whoever pays for the number.
The fight in these markets is over who is permitted to see, not who is permitted to compute.
The proof
One model. One housing stock. One day. Four times the error.
Median error of the Zestimate was 1.8% on homes listed for sale and 7.2% on off-market homes in 2025 — the same model, the same housing stock, the same day. The only difference is whether a price was published. Redfin's independently built model reports 1.85% and 7.27% on the same two populations: two competing firms converging on the same error floor, which suggests the floor is a property of the available observations rather than of either model.
Zillow Group FY2025 Form 10-K, filed 2026-02-11; Redfin Estimate published accuracy, 2025.
What a data moat is
Three properties. A position that has all three is hard to take.
02
A time series that cannot be back-bought
History is the only component that cannot be reconstructed at any price once the window has passed. Start the clock early. Never sell the history.
03
Work that is tedious rather than hard
No single scraping target and no one large engineering problem — thirty small interpretation problems. The shape large vendors are structurally worst at, and a small operator is best at.
Where we are doing it
Two, with honest stage labels.
Deed-Restricted DB
Deed-restricted affordable for-sale housing — an unobserved asset class sitting inside the most instrumented housing market in the world.
PorchLight OS
Operating software for a residential brokerage.
Two doors
Door one
Partner with us
You already generate the raw material. Every day your business produces records that exist nowhere else and that nobody has normalized — the by-product of doing the work. We turn that stream into an asset you own a share of, and we do the part you have no reason to be good at: schema design, normalization across jurisdictions, productization, and finding the buyer.
You keep operating. Nothing about the day job changes.
Door two
Sell us your company
We buy small data and services businesses and hold them. To be plain about what you would be dealing with: SocioClimate is a small operator, not a fund. There is no blind pool behind this, no auction process, no forty-eight-hour letter of intent used as a negotiating device.
You would talk to the person who will still be running the business afterwards.