We follow the money
to its primary record.
We follow capital into emerging technology from the moment it is announced, through the point a government commits to pay it, to where it is actually paid and converted into facilities and orders. What matters is not the size of a figure but how far along that path it has travelled.
Most capital gets reported at Announced and the story stops there. What we watch is whether it moved further. Today we can follow US federal capital from announcement through payment.
Following money requires discipline about counting. Treating a contract ceiling as inflow, counting the same money at two stages, mixing reporting with primary records — none of it takes bad intent, and all of it inflates the numbers. Below is what we hold to in order to avoid it.
Every capital event is ingested from USAspending transaction records, and each award links straight back to the government's own record.
Every figure says where it came from
We separate figures taken from the issuing record — a government transaction, a filing — from figures based on an announcement. The same “$20M” means different things depending on what it traces back to.
Today every capital event is the former. Nothing enters a total unless it traces back to the issuing record.
A contract ceiling is not inflow
A federal contract has a ceiling, an obligation (money the government is committed to pay) and an outlay (money actually paid). These can differ by an order of magnitude. The figures we publish are obligations. Ceilings are kept separate and never enter a total.
Outlays are not present in the transaction feed, so we do not estimate them. A number we do not have stays empty.
The same money is counted once
One contract accumulates many funding actions — the original award plus later increases and decreases. Summing them naively inflates a single contract by however many times it was touched. We collapse actions onto their contract before counting, and each brief states how many actions became how many awards.
We also separate new awards from modifications. Both are real spending, but they mean different things.
When data is missing, we say so
Department of Defense contract data is published with roughly a 90-day delay. In our own measurements a settled month held about 12,000 transactions while the two most recent months held 14 and 28. Other agencies — NASA, DOE, DHS — show no such lag.
Reporting on “this week” without knowing about the lag confuses “nothing happened” with “nothing has been published yet.” That is worse than a wrong number, so we measure currency on every run and state it before any figure.
| Agency | Type | Current through | Lag | Weekly? |
|---|---|---|---|---|
| NASA | Contracts | 2026-07-01 | 1 mo | |
| DHS | Contracts | 2026-08-01 | 0 mo | |
| DOE | Contracts | 2026-07-01 | 1 mo | |
| DOE | Grants | 2026-07-01 | 1 mo | |
| DoD | Contracts | 2026-05-01 | 3 mo |
This table is re-measured on every daily run. These are the values as of the moment you opened this page.
We do not ask codes to do what they cannot
We classify awards using federal industry codes (NAICS), product and service codes (PSC) and assistance listings (CFDA). None of them has a code for “AI.” The furthest they reach is “software development.”
So we refuse to give a segment a name broader than the codes actually support. That is why you see “Defense Software” rather than “Defense AI.”
One award is not counted in full twice
An award can belong to more than one theme. Counting the full amount under each makes the segment totals exceed the real total. We split by weight instead, and where the basis for splitting is weak we tag the award without assigning any amount to it.
Nothing appears without a source
Every figure in a brief traces to a federal record. Open the link and you can confirm we added nothing to it.
Where this runs out today
- We cover US federal capital. Venture rounds, corporate capex and commercial contracts are not yet integrated.
- Weekly coverage means NASA and DOE. Defense is treated as a retrospective because of the publication lag.
- Awards that match on product code but not industry code are missed.
- Duplicate-award clustering is mostly automated; a human has re-checked only part of it.
- Our indices are early hypotheses, not a basis for investment decisions.
Check it against a real one
Everything on this page is visible in a single brief. No signup required.