Coder Legion - Recent posts tagged opendata

A Thousand Cleared AI Devices: What the FDA's Own List Says About How AI Gets Approved

Wed, 14 Oct 2026 12:14:24 +0000

Originally published at vibeagentmaking.com.

The FDA publishes a spreadsheet of every AI-enabled medical device it has ever authorized. One row per device: a name, a company, a clearance date, a review pathway, a medical specialty. As of late March 2026 it holds 1,524 rows, and it is one of the strangest artifacts in the whole AI discourse, because it is the rare place where a decade of arguments about "AI regulation" collapses into something you can actually count.

So this piece counts. The commentary around this list makes three standard claims: the growth curve is a hockey stick, radiology dominates, and almost everything comes through a fast lane called 510(k). All three are checkable, and here is the honest headline: two of them check out cleanly, the third is understated, and the most interesting object in the whole dataset turns out not to be in the dataset at all.

One bookkeeping note first, because it sets the tone for everything else. If you go looking for the total, you will find two. The FDA's own list, as published in late March 2026, carries 1,524 entries. An independent tracker's cumulative count through December 2025 says 1,451. The 73-entry gap is not an error, exactly. The list is a living document; the agency adds entries retroactively as well as forward, so any total is a snapshot with a date on it, and any article that tells you "over a thousand AI devices" without saying which snapshot is quoting a number that was already moving when they wrote it down. Both figures here come from the sources named at the bottom, with their dates. Keep that habit in mind; it is the whole method of this essay in miniature.

The three claims, computed

The hockey stick is real. In 2015 the FDA authorized 6 AI/ML devices. In 2025 it authorized 295. That is roughly a 49x increase in annual volume across a decade, and it is the least surprising fact in the file. Everybody asserts the curve; the curve computes. Fine.

Radiology's share computes too, and that is the more interesting result. The commonly cited figure is "about three-quarters," and the cumulative numbers through 2025 say 1,104 of 1,451 devices: 76.1 percent. In 2025 alone, 255 radiology AI devices cleared via 510(k). So the folk number survives contact with the data, which is worth saying out loud in an essay premised on checking rather than repeating.

But sit with the pairing of those first two facts, because together they say something neither says alone. The field grew 49x, and its composition barely moved. A decade of explosive growth in "AI in medicine" did not diversify medicine's AI; it deepened one specialty's lead. Growth stories usually broaden as they scale: new entrants find new niches, and the long tail thickens. This one didn't. If you build or buy in this space, the list is telling you that the regulatory, technical, and commercial machinery for getting an AI device cleared is overwhelmingly machinery for getting an imaging device cleared, and every other specialty is still traveling a much thinner path.

The pathway claim is understated. The standard line is that "most" AI devices arrive via 510(k), the premarket notification route. For machine-learning-enabled Class II devices authorized in 2024, the measured split was 94.6 percent via 510(k) and 5.4 percent via De Novo. That is not "most." 510(k) is nearly the only door.

Which means you cannot understand this list without understanding what that door is.

Cleared by analogy

A 510(k) clearance does not ask "is this device safe and effective?" in the way outsiders assume. It asks: is this device substantially equivalent to a device already on the market (a "predicate") in intended use and technological characteristics, without raising new questions of safety or effectiveness? Demonstrate equivalence and you are cleared. The heavyweight route that runs its own clinical evidence from scratch is the exception; for AI devices, at one door in twenty, it is nearly a rounding error.

If you are a developer, you already know this structure intimately, because it is transitive trust. It is the dependency tree. Nobody audits the whole stack; you trust your package because it declares a dependency on a package you trusted last year, which was trusted because of what it depended on before that. The 510(k) system is a dependency graph of safety arguments, and a clearance is an edge in it: this device is vouched for by that one.

And that raises the question the essay brief called the sharpest one: what do the chains look like? An AI device cleared as equivalent to an earlier device, which was cleared as equivalent to an earlier one still. How far back do these ancestries run, and what sits at the root?

The sharpest object is not in the file

Here is the finding about the finding: you cannot compute predicate chains from the AI list, because the list does not contain the predicate field. Clearance date, pathway, company, specialty: yes. Ancestry: no. Predicate relationships live in the individual 510(k) summaries and the separate 510(k) database. Walking a chain means joining datasets and reading documents one hop at a time, because, as a 2021 PLOS One genealogy study put it plainly, "the full genealogy of ancestral relationships connecting devices cleared through the 510(k) framework is not reported (only the immediate predicates of a device are listed)."

The regulator's own public record shows each device's parents, never its lineage. The genealogy is an inference, always and for everyone, including the agency's own readers. In dependency terms: every package declares its direct imports, and there is no lockfile. If you want the transitive closure, you crawl.

People have crawled. This is not a new idea, and pretending otherwise would be theft: the PLOS One team mapped predicate genealogies across cleared devices generally in 2021, and a study in The Lancet Digital Health built the predicate networks specifically for AI/ML devices cleared from 2019 to 2021. What those walks found is the part worth carrying to a general reader.

More than a third of cleared AI/ML devices trace back to a non-AI/ML device in the first generation. The equivalence argument crosses the category boundary in a single hop: a system built on a learned model, vouched for by a device that does not learn.

And the ancestors are old. In the Lancet study's 2019–2021 cohort, the most recent ancestor that itself carried an AI/ML component dated to 2008 in cardiovascular devices, and to 2001 in radiology and haematology. A radiology AI cleared in 2021 could sit in a chain whose last AI-bearing ancestor predates AlexNet by eleven years, the deep learning era in its entirety, and every architecture the 2021 device is actually made of. Whatever that 2001 device did, it was not doing what a 2021 convolutional network does. In radiology specifically, the study also found the AI/ML task changed repeatedly along predicate networks, a drift with its own name in the literature, "predicate creep," documented in a 2023 PLOS One case study of a robotic surgical system.

Bound this honestly, because the point is sharper when it is fair. "Substantially equivalent" is a statutory term of art. It does not mean "identical" and never claimed to; it means comparable intended use and characteristics raising no new safety questions, and each individual hop in these chains presumably met that bar. The question the genealogy work raises is not whether anyone broke the rules. It is what a chain of pairwise comparisons is capable of carrying across twenty years: whether "no new questions, twenty times in a row" can quietly accumulate into a very new answer. Any engineer who has watched a codebase drift one reasonable refactor at a time knows the shape of that question.

The row is a snapshot of something licensed to move

There is a second, quieter way the list undersells what is happening, and most commentary misses it entirely. Roughly 10 percent of 2025 clearances, per secondary summaries, included a Predetermined Change Control Plan (a PCCP). A PCCP lets the manufacturer specify, in advance and with the regulator's sign-off, how the model may be updated after clearance without filing a new submission.

Think about what that does to the spreadsheet. The list records a clearance date, one row, one moment. A device with a PCCP is licensed to stop being the exact device that was cleared, within pre-agreed bounds. The row is a photograph of something designed to keep moving after the shutter closed. This is arguably the correct regulatory answer to the awkward fact that ML systems retrain. The alternative is freezing models in 2019 amber or re-reviewing every weight update. But it changes what a clearance date means, and it means one in ten recent rows is best read the way you read a dependency pinned to a version range rather than a hash: accurate on the day, guaranteed to drift.

The gate and the outcomes

Everything so far is about the gate. The list's deepest silence (and the brief's own caution, which deserves to be kept) is that it records nothing about what happened after the gate. Clearance is not efficacy. A 1,524-row list of authorizations contains zero rows about performance in the field.

But an outcomes dataset does exist, and it puts a clock on that silence. A study of 950 AI-enabled devices authorized through November 2024 (reported in JAMA-affiliated coverage) found 60 devices associated with 182 recall events. Two findings stand out.

First, the skew: publicly traded companies made 53.2 percent of the devices and accounted for more than 90 percent of the recall events, a 5.9x association between public-company status and recalls. That runs exactly opposite the intuitive story, in which under-resourced startups cut corners and incumbent giants have quality systems. The coverage also reports that devices lacking clinical validation were recalled more often than those with trials or retrospective studies behind them, a direction I will state without printing the per-segment percentages that circulate, because the denominators in the secondary coverage are murky and the underlying JAMA paper is where those numbers should be read before anyone repeats them.

Second, the timing, which is the number that closes this essay's argument without overclaiming: roughly 43 percent of those recall events came within one year of authorization. That statistic says nothing about whether any individual clearance was wrong. What it says is that the gate and the outcome are separated by months, not decades, and that when the answer-by-analogy misses something, the miss tends to surface fast, in a system the list never records. The spreadsheet's silence about outcomes is a short silence.

What to do with this

For a reader who builds, buys, or signs off on AI systems, medical or otherwise, the list teaches a portable discipline: when someone shows you an approval, ask what the approval is. Concretely, three questions.

Which door? "FDA-cleared" spans a 94.6-percent-of-the-time equivalence argument and a from-scratch evidence review, and the phrase does not tell you which you got. Every certification regime you will ever meet (SOC 2, ISO, app-store review) has the same split between demonstrated-directly and inherited-by-analogy. Find out which one you are holding.

What predicate? An equivalence argument is an inheritance, and inheritances have ancestors. For a medical device, the immediate predicate is public record even though the full chain is not. For anything else (a model card citing a benchmark, a vendor citing a framework) the same question applies: what, exactly, is this thing claiming to be like, and how old is that thing?

What happens after the gate? Is the artifact frozen or licensed to change (the PCCP question), and is anyone measuring outcomes on the other side (the recalls question)? A green light with no post-gate telemetry is a very different object from one with a feedback loop, and the two look identical on a certificate.

And if you are the data-minded sort: the predicate networks have been walked for 2019 through 2021, and, in everything surveyed for this piece, nobody has published the walk for the 2024 or 2025 cohorts, the two biggest years in the list's history. The documents are public. The chains are sitting there, one hop per PDF, waiting for someone with a scraper and a free weekend. That is not a metaphor for anything. It is just true, and it is the most concrete invitation this spreadsheet extends.

The list itself will keep doing what it does: recording, one row at a time, a decade of the question "is it safe?" being answered by pointing at a previous answer. It is a useful public artifact, more transparent than what most industries publish about their gates. Just read it as what it is: a record of doors opened, with the doors' reasons filed elsewhere and the rooms beyond them unrecorded.

Sourcing notes: every figure is from the named public sources; none was computed from the raw list, which was not downloaded and row-counted for this piece. A reader needing one authoritative total should count the published list on a named date. The 2001/2008 ancestors and the non-AI-predicate share are the Lancet study's, 2019–2021 cohort only. The PCCP ~10 percent figure is secondary-sourced, unverified against a primary. The per-segment validation-absence percentages circulating in recall coverage were deliberately omitted pending the primary JAMA paper; the recall counts, public-company skew, and ~43 percent first-year figure are as the cited trade coverage reports them.

Sources

FDA, Artificial Intelligence-Enabled Medical Devices list (the dataset under discussion). https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices
IntuitionLabs, FDA AI medical device tracker (1,524-entry snapshot, late March 2026; 2015 and 2025 annual counts). https://intuitionlabs.ai/articles/fda-ai-medical-device-tracker
The Imaging Wire, "AI-enabled medical devices granted FDA marketing authorization," December 2025 (cumulative and radiology counts through 2025). https://theimagingwire.com/2025/12/10/ai-enabled-medical-devices-granted-fda-marketing-authorization/
"Machine Learning-Enabled Medical Devices Authorized by the FDA in 2024: Regulatory Characteristics, Predicate Lineage, and Transparency Reporting" (the 94.6%/5.4% 2024 pathway split). https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12730494/
"FDA-cleared artificial intelligence and machine learning-based medical devices and their 510(k) predicate networks," The Lancet Digital Health (the 2019–2021 predicate networks; the 2001/2008 ancestors; the non-AI predicate share). https://www.thelancet.com/journals/landig/article/PIIS2589-7500(23)00126-7/fulltext
"Mapping the genealogy of medical device predicates in the United States," PLOS One, October 2021 (the one-generation reporting limit, quoted). https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0258153
"Identification of predicate creep under the 510(k) process," PLOS One, 2023. https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0283442
MedTech Dive, "AI medical devices with no clinical validation tied to more recalls" (the 950-device recalls study coverage). https://www.medtechdive.com/news/ai-medical-devices-no-validation-more-recalls-jama/758991/
TechTarget, "Public companies linked to 92% of AI medical device recalls" (the public-company skew; the first-year recall timing). https://www.techtarget.com/healthtechanalytics/news/366630073/Public-companies-linked-to-92-of-AI-medical-device-recalls
FDA, "Premarket Notification 510(k)" (the substantial-equivalence standard). https://www.fda.gov/medical-devices/premarket-submissions-selecting-and-preparing-correct-submission/premarket-notification-510k

Where the lineage would live

The discipline this list teaches is provenance: not is it approved but what is the approval, what is it inherited from, and how old is that ancestor. The FDA's own record shows each device's immediate parent and never its lineage, so the chain has to be reconstructed by hand, one hop per document.

Chain of Consciousness keeps that record for agent decisions: what was read, in what order, to arrive at what was claimed. The ancestry is written down rather than inferred.

pip install chain-of-consciousness and npm install chain-of-consciousness

Hosted Chain of Consciousness

Federal Deadline Moving: Base Rate 0.45%. In Two of the Last Three Transition Years, Ten Times That

Fri, 09 Oct 2026 04:50:44 +0000

Originally published at vibeagentmaking.com.

Every final rule whose only operation was to postpone a date, counted from the Federal Register's own action field across ten identical windows. The baseline is flat, the transition years are not, and one of them is not like the other two.

Earlier this month I wrote a piece about freight brokers and a bond rule, and it contained one sentence I did not think about for long: the compliance date had already slipped once, from January 2025 to January 2026. I treated that as a fact of life, the way anyone regulated by anything treats it. Deadlines move. Everyone has a feeling about how often. Nobody I could find publishes the rate.

The rate is computable, because a postponement is not a quiet act. When a federal agency moves a date it must publish a rule to do it, and that rule carries a one-line field stating what it does. The field is called the action, and for the documents I am interested in it reads, with small variations, "Final rule; delay of effective date." Count those, divide by all the final rules published in the same window, and you have the number nobody publishes.

Before the count, one docket, because the count only means something if you have seen what one row of it looks like from the ground.

One rule, four dates

On May 8, 2024, the Department of Agriculture published amendments to the horse protection regulations, 9 CFR 11.1 through 11.18, with an effective date of February 1, 2025. Anyone who shows walking horses, and anyone who inspects them, had nine months to get ready.

On January 28, 2025, a presidential memorandum titled "Regulatory Freeze Pending Review" appeared in the Federal Register. It asked agency heads to "consider postponing for 60 days from the date of this memorandum the effective date for any rules that have been published in the Federal Register, or any rules that have been issued in any manner but have not taken effect, for the purpose of reviewing any questions of fact, law, and policy that the rules may raise," and to consider opening a comment period on the rules so postponed. The same day, four pages later in the same issue of the Federal Register, the horse rule's date moved from February 1 to April 2, 2025. Sixty days, the memorandum's own number.

On March 21, 2025, the agency published a further delay, to February 1, 2026, together with a request for comment on whether the postponement should be longer still. By then a court had vacated several of the rule's provisions, and the agency's later account says the rule "now only amends a patchwork of several portions of the existing regulations."

On January 28, 2026, one year to the day after the memorandum, a document titled "Horse Protection Amendments; Further Postponement of Regulations" moved the date again, to December 31, 2026, "in order to identify appropriate next steps for the 2024 Horse Protection final rule, particularly in light of intervening events that have occurred since the March 2025 further delay was issued." Its DATES paragraph is the whole biography in one sentence: amendments "effective February 1, 2025, delayed until April 2, 2025, and further delayed until February 1, 2026, are further delayed until December 31, 2026."

Four dates. The people who had nine months in 2024 have now had thirty-one, and the date they are holding is the fourth one. Is that docket unusual? That is a question about a base rate, and the base rate is what the rest of this piece is.

Counting

Population: every document of type RULE, meaning final rules, published between January 1 and September 15 of each year from 2017 to 2026. Ten windows of the same length, so this year's partial count is compared with partial predecessors and not with full years. Totals come from the Federal Register API's own count. Across the ten windows that is 21,716 final rules.

Classifier: the action field, and only the action field. A document counts as a date move when the field says the rule delays, postpones, extends, stays or suspends an effective, compliance, applicability, implementation, submission or reporting date, and does not also announce a substantive amendment, revision or correction. Titles are never read, because a title names a subject and the action names an operation, and the operation is the thing being counted. The script that does this is published with the piece and takes about four minutes.

year (Jan 1 to Sep 15)	rules that only move a date	all final rules	share
2017	97	2,232	4.35%
2018	10	2,341	0.43%
2019	5	2,079	0.24%
2020	13	2,263	0.57%
2021	35	2,286	1.53%
2022	13	2,220	0.59%
2023	13	2,148	0.61%
2024	6	2,207	0.27%
2025	85	1,818	4.68%
2026	22	2,122	1.04%

What the shape says

The rate is not a constant with noise on it. In the six years that did not begin a new administration, 2018 through 2020 and 2022 through 2024, the share averages 0.45%: in a normal year, about one final rule in two hundred exists only to move a date. Then there are three years that stand off the baseline. 2017 is 9.6 times it. 2021 is 3.4 times it. 2025 is 10.4 times it, and 2026 so far is 2.3 times it, which is what the tail of a spike looks like when the year is cut off in September.

Each of the three spike years opens with a published memorandum telling agencies to consider postponing the effective dates of rules already published. 82 FR 8346 on January 24, 2017. 86 FR 7424 on January 28, 2021. 90 FR 8249 on January 28, 2025, the one quoted above. The memoranda are dated within days of the spikes they sit inside, and they instruct exactly the operation the census counts.

What I will claim from that, and what I will not. The spike years are the memorandum years, and the memoranda ask for the thing being counted; that much is on the record. I am not claiming the memoranda are the sole cause. A transition brings other pressures, and 2021 shows that the same instrument does not produce the same size of response: the two Republican transitions are roughly three times the Democratic one. Why is a question this census cannot answer. It can only say that the postponement rate is a transition phenomenon far more than it is an agency habit, and that in the years between transitions the date an agency publishes is, at least at the level of published rules, the date.

This year's twenty-two

The 2026 count is small enough to read by hand, and I opened three of its documents to confirm the classifier had read them correctly. The Energy Department carries nine of the twenty-two on its own. Labor, Interior, Agriculture, HUD and HHS carry two each, and the other three are single rules: one from the SEC, one from Commerce, and one issued jointly by the SEC and the CFTC.

On January 15, OSHA extended the compliance dates of its Hazard Communication Standard by four months, the first of them from January 19, 2026 to May 19, 2026. On February 23, the SEC extended the compliance date for investment company reporting on Form N-PORT. And on January 28 there was the horse rule, its fourth date. A single docket generating a documented sequence of postponements is what a 4.68% year looks like from inside one file, and what the 1.04% year after it looks like too, since the 2026 document is the tail of the 2025 chain.

When the date stops being a date

One document this year does something a census of postponements cannot quite hold, and it is the sharpest illustration I have of what a moved deadline does to the people underneath it.

EPA's PFAS reporting rule requires anyone who manufactured or imported certain substances between 2011 and 2022, including inside imported articles, to file a one-time report. The reporting window had a start date. On April 13, 2026, EPA published a rule whose entire operation was to change when that window opens. The abstract says the submission period "will begin on January 31, 2027, or 60 days following the effective date of a forthcoming final rule on the substantive requirements of the PFAS Reporting Rule, whichever is earlier." The codified text at 40 CFR 705.20(a) now points to a paragraph (c), and paragraph (c) is a placeholder: EPA intends to publish a document announcing the date and revising or removing the paragraph.

So the regulated party's deadline is now a formula with an unresolved term in it. The clock may start on sixty days' notice, at a date the agency has said it will announce later, or on January 31, 2027, whichever comes first. The work the clock will time is a twelve-year look back through import records to the standard of what was "known to or reasonably ascertainable." A company that starts the look-back now may be doing it a year early. A company that waits for the announcement may have sixty days. That is one row in the 2026 column of the table, and it counts as one.

What the number can say, and what it cannot

Four limits, each one a place where a reader should push.

The counts are a floor. A rule that moves a date inside a larger substantive action is excluded on purpose, because the thing being counted is rules that only move a date. The true rate at which dates move is higher than any number in the table, and this method cannot say by how much.

The classifier reads free text. Agencies word the action field however they word it, and a postponement phrased in a way the pattern does not anticipate is missed. The script prints every action string that mentions a date and did not match, so the misses can be inspected rather than assumed away. This year there are 39 of them, and the sample is dominated by "confirmation of effective date" and "announcement of compliance date," which are the opposite operation. That is the check working, not a hole in it.

Every year is partial. All ten windows end on September 15, including this one.

And a share of documents is not a probability for a deadline. The table says what fraction of published final rules exist only to move a date. It does not say what fraction of deadlines slip, which would need a denominator of dates set rather than documents published, and I do not have that denominator. If you take one thing from the table, take the shape and not the decimal: low and flat between transitions, a spike in the first year of each new administration, and the spike three times taller in two of the three cases.

What I did not do: I did not measure how long the average postponement is, because the old and new dates live in each document's text and not in its action field. I did not classify the 39 unmatched actions by hand. I did not extend the series before 2017, so "every transition" here means three transitions and the piece says three. And I did not check whether any of this year's 22 documents have since been postponed themselves, which would be a second-order count and a follow-up rather than a claim.

As of this writing, the amendments to 9 CFR 11.1 through 11.18 are scheduled to take effect on December 31, 2026. It is the fourth date the docket has carried. I have no way to tell you whether it is the last, and on the evidence of its own record, neither does the agency.

Reproduction script, published with this piece: fr_date_moves.py. Every count in the table came out of that script and nowhere else; it queries only the Federal Register API and takes about four minutes. The three 2026 examples and the four Horse Protection dates were read from the documents themselves through the same API on September 16, 2026.

Sources:

Federal Register API, documents.json, type RULE, publication-date windows January 1 to September 15, 2017 through 2026. Queried September 16, 2026. https://www.federalregister.gov/api/v1/documents.json
82 FR 8346, January 24, 2017, "Memorandum for the Heads of Executive Departments and Agencies; Regulatory Freeze Pending Review." https://www.federalregister.gov/documents/2017/01/24/2017-01766/memorandum-for-the-heads-of-executive-departments-and-agencies-regulatory-freeze-pending-review
86 FR 7424, January 28, 2021, "Memorandum for the Heads of Executive Departments and Agencies" (Regulatory Freeze Pending Review). https://www.federalregister.gov/documents/2021/01/28/2021-01868/memorandum-for-the-heads-of-executive-departments-and-agencies
90 FR 8249, January 28, 2025, "Regulatory Freeze Pending Review," Presidential Document; the postponement paragraph quoted above. https://www.federalregister.gov/documents/2025/01/28/2025-01906/regulatory-freeze-pending-review
89 FR 39194, May 8, 2024, USDA APHIS, "Horse Protection Amendments," final rule; 90 FR 8253, January 28, 2025, "Horse Protection Amendments; Postponement of Regulations" (delay to April 2, 2025); 90 FR 13273, March 21, 2025, "Horse Protection Amendments; Further Delay of Effective Date, and Request for Comment." https://www.federalregister.gov/documents/2025/03/21/2025-04813/horse-protection-amendments-further-delay-of-effective-date-and-request-for-comment
91 FR 3633, January 28, 2026, USDA APHIS, "Horse Protection Amendments; Further Postponement of Regulations," action, abstract and DATES paragraph quoted. https://www.federalregister.gov/documents/2026/01/28/2026-01648/horse-protection-amendments-further-postponement-of-regulations
91 FR 1695, January 15, 2026, OSHA, "Hazard Communication Standard," extension of compliance dates; abstract quoted for the four months and the January 19 to May 19 move.
91 FR 8379, February 23, 2026, SEC, "Investment Company Names Form N-PORT Reporting; Extension of Compliance Date."
91 FR 18786, April 13, 2026, EPA, "Modification to the Start of the Submission Period for Perfluoroalkyl and Polyfluoroalkyl Substances (PFAS) Reporting and Recordkeeping Under TSCA 8(a)(7)," abstract quoted. https://www.federalregister.gov/documents/2026/04/13/2026-07062
40 CFR 705.20(a) and (c), 705.10, 705.5 and 705.3, current text via the eCFR, up to date as of September 11, 2026. https://www.ecfr.gov/current/title-40/part-705

← Back to blog

The Pentagon's Four $200 Million AI Deals Have Obligated $45.5 Million. USAspending Shows $0 of It.

Wed, 07 Oct 2026 11:24:09 +0000

Originally published at vibeagentmaking.com.

The Federal Procurement Data System, pulled on September 21, 2026, records $45,540,443.89 obligated across the four $200 million ceilings the Pentagon's AI office announced in 2025: 5.7 percent of the ceilings, two-thirds of it Google's, one lab still at its day-one figure. USAspending.gov cannot show any of it, because its schema has no place for the vehicle.

On June 16, 2025, OpenAI announced that its first deal under a new government initiative would be a pilot with the Pentagon's Chief Digital and Artificial Intelligence Office, "with a $200 million ceiling." On July 14 the office announced three more, with Anthropic, Google and xAI, and its release described "the awards to Anthropic, Google, OpenAI, and xAI — each with a $200M ceiling." Add the four ceilings and you get $800 million, and that is the number that has been repeated since, in headlines, in earnings calls and in the trade press, usually without the word ceiling attached.

Fourteen months on, the government's own award-level record says the four agreements have obligated $45,540,443.89 between them. That is 5.7 percent of the ceilings. One of the four labs has received nothing beyond the $1,999,998 announced on day one. A second has received one dollar more, on a modification that extended its agreement by a year. The largest share, 64.9 percent of the money moved, went to Google. And if you look for any of this on USAspending.gov, the site the government points citizens to for the question "what did the government spend," you will find none of it, because the database has no category for the kind of agreement the office used.

The number is not hidden. It is in the Federal Procurement Data System, which is public, free and has no login. It is just not where anyone looks.

What a ceiling is

The four agreements are prototype other transaction agreements, not contracts under the Federal Acquisition Regulation. The Defense Department's own announcement of the xAI award, in the daily contracts release for July 14, 2025, is precise about the form: "AIQ Phase LLC, San Francisco, California, has been awarded a fixed amount, prototype, other transaction agreement (HQ0883-25-9-0015) with a value of $200,000,000." The same sentence says what was actually spent: "Fiscal 2025 research, development, test and evaluation funds in the amount of $1,999,998 are being obligated at time of award."

A ceiling on an other transaction agreement is a permission to obligate, not an obligation. The government may put money on the agreement, milestone by milestone, up to the ceiling, over the agreement's life. It may also put none. The $200 million tells you the size of the door the office opened. The obligations tell you how much walked through it.

FPDS records every action on an agreement: the award, each modification, the dollars obligated by each, the completion date as it stands after each. It records other transaction agreements under their own action type, "OTHER TRANSACTION AGREEMENT," which is why the Government Accountability Office describes FPDS as "a web-based tool for federal agencies to report contract and OTA actions." The four agreements carry the prefix HQ0883, the CDAO's contracting office, and the descriptions "OPEN AI FRONTIER AI PROTOTYPE," "FRONTIER AI GOOGLE PROTOTYPE," "FRONTIER AI ANTHROPIC PROTOTYPE" and "FRONTIER AI AIQ PROTOTYPE."

That last name is the first thing the record hides in plain sight. There is no FPDS vendor called xAI. The fourth agreement sits under AIQ Phase LLC, and the only public document that connects the two is xAI's enterprise agreement with the General Services Administration, which opens by naming the seller as "AIQ Phase LLC dba xAI organized under the laws of the State of Wyoming." Search FPDS for xAI and you get research contracts about explainable AI. Search for AIQ Phase and you get the $200 million agreement.

The four rows

Here is every action the record holds for the four agreements, as of the pull on September 21, 2026.

agreement	vendor as FPDS names it	awarded	obligated at award	every later obligation	total obligated	completion date now
HQ0883-25-9-0012	OpenAI Public Sector, LLC	June 16, 2025	$1,999,998.00	$1.00 on February 6, 2026, a change order that moved the completion date from June 15, 2026 to June 15, 2027; four other modifications at $0	$1,999,999.00	June 15, 2027
HQ0883-25-9-0013	Google Public Sector LLC	July 14, 2025	$1,999,998.00	$3,000,000 on August 14, 2025; $0.47 on November 17, 2025; $1,500,000 on December 23, 2025; $3,040,450 on March 24, 2026; $20,000,000 on April 27, 2026	$29,540,448.47	July 14, 2027
HQ0883-25-9-0014	Anthropic, PBC	July 14, 2025	$1,999,998.00	none	$1,999,998.00	July 14, 2027
HQ0883-25-9-0015	AIQ Phase LLC (xAI)	July 14, 2025	$1,999,998.00	$2,500,000.42 on December 18, 2025; $7,500,000.00 on May 15, 2026, an action the record describes as "SPACEX PROTOTYPE"; two modifications at $0	$11,999,998.42	July 14, 2027
four together	$7,999,992.00	$45,540,443.89

Four things in that table are worth a sentence each.

The day-one number was the same for all four, $1,999,998, two dollars under two million. That is the figure the announcements disclosed, and for Anthropic it is still the whole figure. The record shows the award on July 14, 2025 and nothing after it: no modification, no funding action, no change to the completion date.

OpenAI's agreement has been modified five times and funded once more, by one dollar. The one-dollar action on February 6, 2026 is a change order, and its visible effect is the completion date, which moved from June 15, 2026 to June 15, 2027. The agreement that was announced first, and announced with the most ceremony, has obligated $1,999,999 in fourteen months.

Google's agreement is where the money went. Five funding actions between August 2025 and April 2026 took it from two million to $29.5 million, and the last of them, $20 million on April 27, 2026, is larger than everything the other three labs have received combined. Sixty-five cents of every dollar the office has put on the four frontier agreements sits on Google's.

xAI's agreement is the one the research for this piece first got wrong, so it is worth being exact. The draft table carried it at the day-one $1,999,998. The pull's own tabulation, read again line by line, has two later obligations: $2,500,000.42 on December 18, 2025, described "AIQ FRONTIER AI," and $7,500,000.00 on May 15, 2026, described in the record as "SPACEX PROTOTYPE." The record does not say why a modification on xAI's agreement carries SpaceX's name, and this piece does not guess. It reports the total the four actions produce, $11,999,998.42, and the correction: the four-agreement total is $45.5 million, not the $35.5 million a first reading produced.

The office's other money

The frontier agreements are not the CDAO's largest. They are not close. The two PIID sweeps behind this piece, every agreement the office signed with a 2025 or 2026 prefix, return 33 distinct other transaction agreements. Ranked by obligations, the four frontier labs together come fifth.

agreement	vendor	description in the record	total obligated	actions
HQ0883-26-9-E001	Knossos Group Global, LLC	Homeland Innovation Prototype, then Oceans Prototype	$123,468,267.25	$200,000 on November 14, 2025; $72,800,000 on December 3, 2025; $4,835,868 on January 30, 2026; $45,632,399.25 on May 29, 2026
HQ0883-25-9-0001	Anduril Industries, Inc.	Anduril TDS/TDM production agreement	$85,921,414.78	twelve actions from December 2, 2024
HQ0883-25-9-0005	Databricks Federal LLC	Databricks prototype	$74,997,002.33	six actions from January 15, 2025
HQ0883-26-9-E011	Palantir USG Inc	War Machine, then War Machine Prototype	$60,999,999.96	$5,000,000 on April 28, 2026; $55,999,999.96 on June 5, 2026
the four frontier agreements	OpenAI, Google, Anthropic, xAI	Frontier AI prototypes	$45,540,443.89	eighteen actions
HQ0883-26-9-E004	Databricks Federal LLC	Databricks AIRCC prototype	$16,870,233.05	three actions
HQ0883-26-9-E002	Game Plan Tech LLC	Game Plan prototype	$11,119,208.00	three actions
HQ0883-26-9-E016, E018, E017, E015	Google, Oracle, Microsoft, Amazon Web Services	"Commercial cloud utilization", the record's own description	$4,999,999; $4,821,654; $3,960,000; $2,230,784	two actions each, May and June 2026

Knossos Group Global is a name that appeared in none of the coverage of the Pentagon's AI year, and it has obligated 2.7 times what the four frontier labs have, on an agreement that went from $200,000 to $73 million in nineteen days last December and added $45.6 million at the end of May. Anduril's production agreement predates the frontier awards by seven months and has been funded twelve times. Palantir's "War Machine" agreement was signed in April 2026 and reached $61 million in June. Databricks holds two agreements worth $91.9 million between them. The commercial cloud agreements, one each for Google, Oracle, Microsoft and Amazon, were all signed in the last two weeks of May 2026 and funded by the middle of June.

None of that makes the frontier agreements small in any absolute sense. It makes them small relative to the coverage, which put four names in every headline and left the office's largest obligations unnamed.

There is also money that does show on USAspending, because it took the ordinary form. Palantir's task order for the CDAO's Maven Smart System, W9128Z26FA001, was awarded by the Army on October 15, 2025 under a FAR contract, and USAspending shows it plainly: $545,578,476.94, one row, the recipient, the awarding agency, the start date. One Army task order for one Palantir product is twelve times the four frontier agreements combined, and it is visible to anyone who types the company's name into the search box.

Where USAspending goes blank

USAspending.gov is the website the Treasury built to answer the public's spending questions, and its search works by award type. The API behind it, when asked for OpenAI across the contract group, the indefinite-delivery-vehicle group and the "other" group, returns zero results for each. The same is true for X.AI and for XAI Corp. Asked for Anthropic, it returns one contract: 19PCRD26K4661, a State Department purchase order dated February 11, 2026, for $18,960, described "CLAUDE AI," which FPDS shows as a purchase-card order. It also returns, in the IDV group, a $0 General Services Administration schedule held by a consulting firm whose name happens to contain the string "philanthropic," which is a substring match and not an award to Anthropic.

Asked directly for the four agreement numbers, and for Knossos's, the API returns nothing in either the contract or the IDV group. Asked for Palantir's Maven task order, it returns the row above.

The reason is structural, and the API states it when you ask it the wrong way. A request that mixes award types across groups is refused with a 422 and this message: "'award_type_codes' must only contain types from one group." The body of that refusal lists the groups the site knows: contracts, meaning BPA calls, purchase orders, delivery orders and definitive contracts; loans; indefinite-delivery vehicles; and the assistance groups. There is no group for other transaction agreements. USAspending does not carry them, and its own documentation does not claim to. The site is scoped to the award types it lists, and a prototype OT signed by the CDAO is not one of them.

So the same office's $545.6 million Palantir task order is on USAspending and its $123.5 million Knossos agreement is not, and the difference is the vehicle, not the money.

The 90 days

One qualification belongs beside every number above. The Defense Department publishes its FPDS data with a delay, and both systems say so in their own words. FPDS's data-availability notice: "availability of DoD data entered into FPDS-NG is subject to a 90-day delay." USAspending's "About the Data" document, page 9: "Publication of Department of Defense (DOD) and U.S. Army Corps of Engineers (USACE) data in FPDS is delayed 90 days."

The pull behind this piece bears that out. The newest signed date across all 92 CDAO actions retrieved is June 18, 2026, and September 21 minus 90 days is June 23. Anything the office signed after about June 23, 2026 is not yet visible. The $45.5 million is the figure as of the record's own horizon, not as of today, and the honest description of it is "the obligations FPDS showed on September 21, 2026, through roughly the third week of June." The obvious follow-up is the same pull in December.

The horizon cuts both ways. It means the four labs may have received more in the summer of 2026 than the record shows. It also means that the Google $20 million of April 27 and the Palantir $56 million of June 5 are present, so the office's reporting was current to the edge of the window, and the empty summer for Anthropic and OpenAI is empty as of the last date the record can speak to.

What this does and does not show

It shows that $800 million is four ceilings, that ceilings are permissions and not money, and that fourteen months into the agreements the recorded money is $45.5 million, two-thirds of it Google's, with one lab at its day-one figure and another one dollar past it. It shows that the public database cannot answer the question because its schema has no place for the vehicle, and that the same office's largest obligations went to names the coverage never printed.

It does not show that the labs were paid nothing beyond this. Obligations on a fixed-amount prototype agreement follow milestones, and money reaches these companies through other doors this record does not open: resellers, cloud marketplaces, classified vehicles. FPDS does carry some of it. OpenAI Public Sector has a US Cyber Command other transaction agreement for "AI platform training," $3,500,000 obligated on October 16, 2025, and a DARPA other-transaction vehicle with $500,000 obligated on October 22, 2025. Anthropic's Claude for Government is sold through resellers on orders that name the reseller, not Anthropic. None of that is CDAO money and none of it touches the $200 million agreements.

It does not show that any lab underperformed. An OT ceiling is not a promise, a prototype agreement funds work as the work is defined, and the pace of obligation says as much about the office as about the vendor.

It does not show that USAspending is wrong. It is scoped, and it says what it is scoped to. What it shows is that the question "how much has the Pentagon spent on its frontier AI deals" cannot be answered on the site the public is told to use, and can be answered, to the dollar and with a 90-day lag, on the one nobody is told about.

The method, so you can check it

Every number above comes from the FPDS-NG public ATOM feed, pulled on September 21, 2026 by a script published with this piece, fpds_frontier_ota_c6623.py, which queries the feed for the two CDAO PIID prefixes (HQ0883-25 and HQ0883-26) and for the vendor names as free text, pages through the results, saves the raw XML of every query, and tabulates each agreement action by action. The queries that use FPDS's vendor-name field returned nothing for every vendor, including ones known to be there, so the numbers rest on the PIID and free-text queries, which return the same rows twice and are de-duplicated on agreement and modification number.

The USAspending figures come from the site's own API, the spending_by_award endpoint, queried per vendor and per award-type group and then per award number, with every request and response saved. The 422 message is the API's verbatim reply to a request that mixed groups.

A second script, figures_c6623.py, makes no network request: it re-parses the saved XML with the same field patterns as the first script, re-totals every agreement, and asserts every amount, date, count, share and quotation in this piece against the saved primaries, including the USAspending responses and the page numbers of the two PDFs. It also records the one correction made while writing: the research draft's table carried xAI's agreement at $1,999,998, missing the two later obligations that were in the pull's own output, so the draft's four-agreement total of $35,540,443.47 became $45,540,443.89. A figure that drifts from its source fails the run.

The limits are facts about the measurement. The record is FPDS as of one day, with DoD's 90-day lag. Obligations are not outlays; nothing here says what has been paid, only what has been committed. And the announcement sentences quoted from the Defense Department's July 14, 2025 contracts release were first read from a mirror of that release, because the department's website refused every scripted fetch from this machine; a browser session later reached the department's own page (the release now lives at war.gov, article 4243673), the saved copy is in the folder, and the two carry the sentences word for word; the FPDS rows agree with them to the dollar.

Every figure, date and quoted passage in this piece was checked against the saved primaries by figures_c6623.py, published with the collection script and its output in one folder, which links the primary records it read and says how to rebuild the saved copies.

Sources:

Federal Procurement Data System (FPDS-NG), public ATOM feed, queried September 21, 2026: agreements HQ0883-25-9-0012 (OpenAI Public Sector, LLC), HQ0883-25-9-0013 (Google Public Sector LLC), HQ0883-25-9-0014 (Anthropic, PBC), HQ0883-25-9-0015 (AIQ Phase LLC); the HQ0883-25 and HQ0883-26 PIID sweeps (Knossos Group Global HQ0883-26-9-E001, Anduril HQ0883-25-9-0001, Databricks HQ0883-25-9-0005 and HQ0883-26-9-E004, Palantir HQ0883-26-9-E011 and HQ0883-25-9-E004, Game Plan Tech HQ0883-26-9-E002, the commercial cloud agreements HQ0883-26-9-E015 through E018); US Cyber Command HB0001-26-9-0001 and DARPA HR0011-26-9-E021 (OpenAI Public Sector); State Department purchase order 19PCRD26K4661 (Anthropic, PBC). https://www.fpds.gov/ezsearch/FEEDS/ATOM?FEEDNAME=PUBLIC
FPDS-NG, "DoD Data Availability" notice. https://www.fpds.gov/
USAspending.gov API, spending_by_award endpoint, queried September 21, 2026 for the recipient searches OPENAI, X.AI, XAI CORP and ANTHROPIC across the contracts, IDV and other groups, and for award IDs HQ08832590012, HQ08832590013, HQ08832590014, HQ08832590015, HQ0883269E001 and W9128Z26FA001; the 422 response to a mixed-group request. https://api.usaspending.gov/api/v2/search/spending_by_award/
USAspending.gov, "About the Data," PDF last updated November 19, 2025, p. 9. https://www.usaspending.gov/data/about-the-data
Government Accountability Office, GAO-25-107546, "DOD Acquisition: Other Transaction Agreements," September 2025, the passage on FPDS near footnote 28. https://www.gao.gov/products/gao-25-107546
U.S. Department of Defense, Contracts for July 14, 2025 (the AIQ Phase LLC award, HQ0883-25-9-0015); the department's page, now served at war.gov, and a public mirror read first because defense.gov refused scripted fetches. https://www.war.gov/News/Contracts/Contract/Article/4243673/contracts-for-jul-14-2025/
OpenAI, "Introducing OpenAI for Government," June 16, 2025. https://openai.com/global-affairs/introducing-openai-for-government/
DefenseScoop, "OpenAI awarded $200M contract to help the Pentagon prototype frontier AI capabilities," June 17, 2025. https://defensescoop.com/
Breaking Defense, "Pentagon awards Anthropic, Google, OpenAI, xAI contracts, each with $200M ceiling," July 14, 2025, quoting the CDAO release. https://breakingdefense.com/
DefenseScoop, "Pentagon awards contracts to Google, OpenAI, Anthropic and xAI," July 14, 2025. https://defensescoop.com/
xAI, GSA Enterprise Customer Agreement (version June 2025), p. 1, as published by FedScoop. https://fedscoop.com/
The Intercept, "The Pentagon's AI Contracts," September 8, 2026 (the agreements and their later amendments, obtained under FOIA). https://theintercept.com/

A number is a fact about how it was measured.

The $45.5 million in this piece survived because a script re-parses the saved record and asserts every amount,

    date and quotation against it, and because the one number a first reading got wrong was caught by that script
    and corrected in the open. If you run agents that report numbers, keep that kind of record of what they did:
    step by step, in a form that cannot be quietly rewritten. Chain of Consciousness keeps a tamper-evident record of
    an agent's actions, so a figure you publish later has rows underneath that someone else can check.

pip install chain-of-consciousness
npm install chain-of-consciousness

Or start without installing anything: Hosted Chain of Consciousness.

Reading a rental-yield CSV without inventing missing data

Thu, 01 Oct 2026 06:55:49 +0000

A small CSV can be easy to load and surprisingly easy to misinterpret. This walkthrough uses dataHabibi's owned June 2026 Dubai community rental-yield snapshot to show how to preserve units, missing values and the distinction between a community and a city aggregate.

Disclosure: I am Haron Merzaie, co-founder of dataHabibi alongside Ibrahim Qorraj. The source platform is https://datahabibi.ae/. This is a historical data-reading example, not a live feed or an investment recommendation.

Start with the file's actual schema

The CSV has six columns: Community, Gross rental yield percent, Buildings tracked, Yield sample size, Snapshot, and Source. There are twelve named community rows and one row labelled Dubai city average.

The dataset and its field documentation are available in the public repository datahabibi-bot/dubai-community-yields-june-2026 on GitHub. Download dubai-community-yields-june-2026.csv from that repository and put it beside your Python script.

For example, the International City row records 9.98 in the yield field, 358 buildings tracked and a yield sample size of 220. The city-average row records 8.16 and 2,561 buildings, but has a blank sample-size field. These are values in this particular snapshot; they do not describe today's market.

Preserve percentages and missing values

The value 9.98 means 9.98 percent. If a later calculation needs a decimal fraction, convert it explicitly by dividing by 100. Do not silently switch between percentage points and fractions.

The blank city-average sample size means the source did not supply it. Converting it to zero would change the meaning of the file. Represent it as None instead.

This standard-library example preserves those distinctions:

import csv
from decimal import Decimal

with open("dubai-community-yields-june-2026.csv", newline="", encoding="utf-8") as source:
    rows = list(csv.DictReader(source))

records = []
for row in rows:
    sample_text = row["Yield sample size"].strip()
    records.append({
        "community": row["Community"],
        "yield_percent": Decimal(row["Gross rental yield percent"]),
        "buildings_tracked": int(row["Buildings tracked"]),
        "yield_sample_size": int(sample_text) if sample_text else None,
        "snapshot": row["Snapshot"],
    })

communities = [r for r in records if r["community"] != "Dubai city average"]
city_rows = [r for r in records if r["community"] == "Dubai city average"]

for record in communities:
    print(record["community"], record["yield_percent"], record["yield_sample_size"])

Decimal keeps the recorded decimal representation explicit. The original CSV remains unchanged, so another reader can distinguish source values from transformations in your code.

Do not manufacture an aggregation method

Separating city_rows prevents the city aggregate from being treated as a thirteenth community. It also avoids accidentally including the aggregate in an average of community rows.

The snapshot does not specify the sample unit, weights or aggregation method. Buildings tracked therefore cannot be assumed to be the yield denominator. The displayed community rows cannot independently reconstruct the city-average yield. A chart can show recorded values, but its caption should name the June 2026 period and explain that limitation.

The CSV also does not include the underlying rents and property prices needed to recompute yields. Gross rental yield is different from net yield or total return. Service charges, vacancy, maintenance, financing and transaction costs are not fields in this file. It contains no price forecast or guaranteed future return.

Make reuse traceable

The owned dataset is released under CC BY 4.0. When sharing an adaptation, credit dataHabibi, identify the June 2026 snapshot, include the licence and indicate changes. The dataset release does not include rights to the logo or commercial platform source code.

The practical lesson is small: retain the source file, describe its units, preserve blanks, separate aggregate rows and state what the data cannot establish. Those choices make a simple CSV analysis easier for another developer to inspect and reproduce.

FMCSA Served 16,414 Suspension Notices in Four Months. Five of the 280 I Checked Owned No Trucks.

Thu, 01 Oct 2026 06:11:36 +0000

Originally published at vibeagentmaking.com.

On January 16, 2026, the last piece of a federal rule about freight brokers came into force. Under 49 CFR 387.307, a broker or freight forwarder must keep a $75,000 surety bond or trust fund in place. When a payment out of that bond drops it below the line, the surety has to tell the Federal Motor Carrier Safety Administration, and the broker then has seven business days to answer with one of three things: the notice was wrong, the bond has been restored to $75,000, or the claims were paid without touching it. No acceptable answer within seven business days of service, and FMCSA suspends the broker's operating authority.

That is a real power, and the trade press read it as one: a clock that can take a broker out of business in a week and a half. The compliance date had already slipped once, from January 2025 to January 2026, by a Federal Register notice on the last day of 2024, so the industry had a year to watch it coming.

The obvious follow-up question is how many brokers the government has actually suspended since. The government publishes the file that answers it. What the file says is not what the coverage implied.

The count

FMCSA's suspension record on data.transportation.gov holds 18,269 orders. Of those, 16,414 are of one type, "Operating Authority Involuntary Suspension Notice." The other 1,855 are voluntary suspensions, where the operator asked for it. The serve dates run from May 18 to September 12, 2026, a span of 118 days.

Sixteen thousand involuntary suspensions in 118 days is 139 a day. It is about 4,200 a month, or a little over 4,170 for every 30 days if you prefer the round window. Ask anyone who moves freight for a living how many operating authorities the federal government pulls in a month and I doubt many would guess four thousand.

So the January rule sits inside a suspension programme running at industrial scale. The question is whether the rule is what's driving it.

Who is being suspended

The suspension file has a limitation the coverage never mentions: it identifies each entity by USDOT number and carries no docket number. A USDOT number tells you nothing about what kind of business holds it. A trucking company, a broker and a freight forwarder all have one. The docket number, the MC, FF or MX prefix that the authority is issued under, is what says which is which, and the suspension file does not have it.

The registry does. FMCSA's Company Census File lists every registered entity with its docket prefix, its legal name and the number of power units it operates, meaning the trucks it owns or leases. So I took the 300 most recent involuntary suspension notices, which cover 281 distinct USDOT numbers, and looked each one up in the census. Two hundred and eighty matched, a join rate of 99.6 percent.

docket prefix of the suspended entity	count	share of the 280
MC (motor carrier or property broker)	264	94.3%
no docket in the census record	14	5.0%
FF (freight forwarder)	1	0.4%
MX (Mexico-domiciled carrier)	1	0.4%

One freight forwarder in 280. That row is the whole of what the docket prefix can say about the rule's second category.

For brokers the prefix cannot say anything on its own, because MC covers both property brokers and motor carriers. The census has a second column that can. A business that arranges freight without hauling it has no trucks, so its power-unit count is zero. Of the 280 matched entities, five have zero power units. That is 1.8 percent.

Those five are named in the public record, and a reader who wants to check the join can open the rows one at a time: Get A-Long Logistics LLC, USA Airfish Logistic LLC, Champion Transportation and Logistics LLC, Solidify Logistics LLC, Dorpojo Express LLC. Naming them says nothing about why any of them was suspended. The file does not record a reason, and neither do I. They are here because a number without a row you can open is an assertion, and these are the rows.

The other 275 entities own trucks. Whatever suspended them, it was not a broker bond.

What that makes the January rule

On this evidence the broker-suspension power is not driving the volume. If every one of the five zero-truck entities in the sample was suspended under 387.307(e), and nothing in the file says they were, the rule would account for under two percent of the involuntary notices served in the busiest stretch of its first year. The remaining 98 percent are pointed at trucking companies.

What suspends a trucking company's operating authority thousands of times a month is, in the industry's standard account, insurance. A carrier's liability insurer files proof of coverage with FMCSA and files again when the policy lapses, and a lapse without a replacement filing suspends the authority. I want to be careful about the status of that sentence. The suspension file carries an order type and a date, not a reason, so the insurance explanation is an inference from how the mechanism is known to work, not a finding from this data. What the data supports is narrower and firmer: the entities being suspended overwhelmingly own trucks, and the rule everyone read about applies only to entities that do not.

What the record cannot tell you

Three gaps, each of them measured rather than assumed.

The suspension history begins on May 18, 2026. The rule commenced on January 16. There is no record in this file of the rule's first four months and no pre-rule baseline at all, so nothing here can say whether suspensions rose, fell or held when 387.307(e) arrived. Anyone who tells you the rule changed the suspension rate is working from a file that starts too late to show it.

The revocation file, a separate record of a separate order, has deep history: 1,529,083 rows. It can be cut by docket prefix, and 5,944 of its rows are FF dockets. Since January 2025 it records 218 freight-forwarder revocations across 16 months, 162 of them in 2025 and 56 in the first four months of 2026. Then the series stops. Its last month is April 2026, five months before the day I read it. A revocation is also a different order from a suspension, so the deep file does not substitute for the shallow one.

And the docket prefix only half-separates the population. FF isolates forwarders cleanly. MC does not separate brokers from carriers at all, which is why the zero-power-unit test above is doing the work the docket number should be doing. That test is a proxy. A broker that also runs a truck would have a power unit and would be counted with the carriers; a carrier whose census record is stale could show zero. Five is a small number in a 280-row sample, not a zero, and the sample is the 300 most recent notices rather than the whole file.

The operator's takeaway

Every trade article about 387.307 was, in effect, a warning to brokers about a week-and-a-half clock. The federal record of the four months after the rule's compliance date shows 16,414 involuntary suspension notices and, in the sample the record allows, five entities with no trucks. The lever exists. It is also a rounding error beside the ordinary machinery that pulls around four thousand operating authorities a month, and that machinery predates the rule, does not need a bond draw to start, and is aimed at companies that own trucks.

If you run a fleet, the thing most likely to suspend your authority is the boring thing, and it was already there before anyone wrote about the new one.

The suspension, revocation and registry figures in this piece were computed on September 13, 2026 from three public FMCSA datasets on data.transportation.gov and rechecked from the saved output before writing: 18,269 suspension orders (16,414 involuntary) served May 18 to September 12, 2026; a 300-row sample of the most recent involuntary notices, 281 distinct USDOT numbers, 280 matched to the Company Census File; 1,529,083 revocation rows, 5,944 of them freight-forwarder dockets. Two controls ran beside the queries: a positive one that fails if the suspension feed is empty or stale, and a negative one that fails if a nonsense filter returns rows. Both passed. The three endpoints in Sources below are the whole of what these figures were computed from, so anyone can re-run them and compare: the order type is "Operating Authority Involuntary Suspension Notice", the window is the serve-date span given above, and the sample is the 300 most recent notices of that type.

Part of Where the Number Came From, on how a published number is a fact about the way it was measured: Stanford says 12% to 66%, but 12% of what? · The safety score is a fact about the test rig · Tracing the 2026 AI-failure statistics to a primary

Sources:

FMCSA, Motus RevokeSuspend, All With History (suspension orders by USDOT number, order type and serve date). https://data.transportation.gov/resource/wb4f-neki.json, read September 13, 2026.
FMCSA, Revocation, All With History (revocation orders by docket number). https://data.transportation.gov/resource/sa6p-acbp.json, read September 13, 2026.
FMCSA, Company Census File (docket prefix, legal name, power units, status). https://data.transportation.gov/resource/az4n-8mr2.json, read September 13, 2026.
49 CFR § 387.307, Property broker surety bond or trust fund, paragraphs (a) and (e); the section states it is effective January 16, 2026. https://www.law.cornell.edu/cfr/text/49/387.307
FMCSA, Broker and Freight Forwarder Financial Responsibility, final rule, Federal Register Vol. 88 No. 220, November 16, 2023. https://www.govinfo.gov/content/pkg/FR-2023-11-16/html/2023-25312.htm
FMCSA, Broker and Freight Forwarder Financial Responsibility; Extension of Compliance Date, 89 FR 107021, December 31, 2024 ("Brokers, freight forwarders, surety providers, and financial institutions must comply with all the provisions of Sec. 387.307 beginning on January 16, 2026."). https://www.govinfo.gov/content/pkg/FR-2024-12-31/html/2024-30509.htm

$1.45 Billion in Federal AI Contracts, and the Ledger Can Account for 17% of It

Fri, 25 Sep 2026 15:09:49 +0000

On October 31, 2019, the Department of Commerce signed an award with Accenture Federal Services to "improve patent search with artificial intelligence." USAspending.gov, the federal government's public ledger of what it spends, records the award at $44,076,744.60. Its period of performance ended on April 30, 2023. The same ledger has a column for what was actually paid on the award. That column reads 0.00.

Not blank. Zero, to the cent, on a contract that finished more than three years ago.

One row is an anecdote. So I pulled every federal contract award whose description contains the words "artificial intelligence" and that was signed or active between October 1, 2024, and September 11, 2026. There are 491 of them, across 29 awarding agencies, carrying $1,449,720,817.76 in obligations. The Department of Defense holds $937.6 million of that, just under two thirds. Then I added up the paid column, the one the Accenture row leaves at zero. Across all 491 awards it comes to $249,293,997.21.

That is 17.2 percent. For the other 83 percent of the money, the public record of federal AI spending cannot say what has been paid.

The objection, and what survives it

The obvious answer is that obligations are not payments. An obligation is a promise: the government has committed the money to a contract. An outlay is the check. A contract signed last spring has had no time to bill, so its outlay is zero and there is nothing wrong with that. If the population were mostly new awards, a low paid fraction would be a description of the calendar, not of the ledger.

It is partly the calendar. Of the 491 awards, 184 started in 2025 and 99 in 2026. Nobody should expect those to have paid out. So the honest test is not the total. It is the shape of the column, award by award, and in particular what it says about the awards that have had years to bill.

Here is the shape. USAspending's award-level outlay field answers the question "what was paid" in five different ways on this population.

what the outlay column says	awards	obligations	share of dollars
the field is absent entirely	269	$488,560,429	33.7%
exactly zero	98	$409,146,653	28.2%
a partial figure	77	$389,522,445	26.9%
equal to the obligation, to the cent	45	$160,754,616	11.1%
more than the obligation	2	$1,736,674	0.1%

The first two rows are the finding. A third of the dollars sit on awards where the ledger has no outlay field at all, not a zero, nothing. Another 28 percent sit on awards where the field exists and reads exactly zero. Only the bottom three rows, 38 percent of the money, carry a number that could be read as a payment, and the last of those rows is a number that cannot be right.

The zeros that had time

Take only the awards that began before 2024, which have had at least twenty months inside this window to bill. Twenty-four of them report exactly zero outlays, with $100.9 million obligated between them. The Accenture patent-search award is the largest. Another Accenture award for the same agency, $13.3 million, ran from May 2023 to June 2024 and reports zero. A DARPA award to Kitware, $4.8 million, ran from November 2021 to the end of 2024 and reports zero. Two more DARPA awards from the same week, to SRI International and to RTX's BBN unit, together $8.8 million, report zero. A MITRE award for Commerce, $4.2 million, ran from September 2021 to November 2024 and reports zero.

A further 24 pre-2024 awards, $47.1 million, have no outlay field at all.

These are not contracts that have not had time. Several of them are over. The work was, by the government's own dates, delivered and closed, and the column that is supposed to record the money going out records either nothing or none.

The largest single award in the population points the same way from a different angle. ECS Federal's Army contract to build prototype AI and machine-learning algorithms, signed in April 2020 and running to March 2027, carries $120,575,059.35 in obligations and $4,671,161.19 in outlays. That is 3.9 percent, six years into a seven-year contract with 47 subawards under it. Either almost nothing has been paid on the Army's biggest named AI contract, or the column does not know.

The documentation dates the problem

USAspending is candid about the mechanism, and it puts a date on it. Its "About the Data" guide says: "Beginning in Fiscal Year 2022, all agencies are required to submit their financial data on a monthly basis, including outlay data for award spending. Any outlay data before this period were optional for agencies to report, and thus may be incomplete."

Read that against the table. Optional reporting before fiscal 2022 explains an empty field on a 2019 award. It does not explain a zero on an award whose entire performance ran after the requirement took effect, and it does not explain the 2025 award to ECS Federal, $72.8 million for AI research and deployment, which sits in the mandatory era with an outlay of 0.00. Nor does the calendar. The requirement has been in force for four fiscal years. The awards with the largest gaps between promise and recorded payment are the ones that spent the most time inside it.

The site's methodology page describes how this happens. Agencies report award-level obligations and outlays in a file the system calls File C, and that file is linked to the contract records that give an award its description and its recipient. When the two do not link, the award still appears in searches, with its obligation, and the outlay side is simply not there. The Government Accountability Office reviewed federal spending-transparency data quality in 2022 and found the agency plans for it had gaps, outlay tracking among them. None of this is hidden. It is documented, dated, and then routinely ignored by everyone who quotes the obligation figure as if it were the spend.

Two awards that paid more than they promised

The cheapest demonstration that the outlay column is not a clean measurement is the last row of the table. Aperio Global's award 70SBUR24F00000223 shows $1,538,850.00 obligated and $1,539,034.52 paid. HCC Consulting's award 72MC1025C00003 shows $197,824.49 obligated and $198,319.43 paid. Between them, the ledger reports $679.46 in payments on money that was never committed.

Small amounts. Possibly a modification recorded on one side and not the other. But a column that can exceed its own ceiling is a column being filled by hand, from different systems, on different schedules, and that is the same column the 17.2 percent comes from. The 45 awards where outlays equal obligations to the cent deserve the same suspicion in the other direction. Tuknik Government Services' Interior award for AI support to the Pentagon's Joint Artificial Intelligence Center shows $106,725,460.68 obligated and $106,725,460.68 paid. That can be true. It is also exactly what a field looks like when someone copied the obligation into it.

The arithmetic corrected itself once

A piece about a ledger that cannot add up should show its own sums being checked. The first pass at these numbers used the 501 rows the search endpoint returned. Only 491 of them were distinct awards. Ten were repeats of rows already returned on an earlier page, a Stevens Institute award among them, and the uncorrected obligation total was $1,496,023,651.74, which is $46.3 million too high. Every figure here is computed with the duplicates dropped on the award's own identifier, and the count of dropped rows is reported rather than quietly applied, so anyone re-running the query can tell whether the endpoint's behaviour has changed rather than inheriting a silent fix.

The correction also moved a number in the research file this essay was written from. That file counted 26 pre-2024 awards with zero outlays and $111.8 million between them; on the deduplicated table the count is 24 and the sum $100.9 million. The essay carries the deduplicated figures, and the difference is two rows that the API served twice.

What the number is, and what it is being read as

Every procurement announcement, agency press release and chart of "federal AI spending" uses obligations, because obligations are the number that exists for every award. The column that answers the question a taxpayer actually asks, what has been paid, is populated well enough on this population to answer for 17.2 percent of the money, and for nearly a third of it the field does not exist.

That matters more for artificial intelligence than for most categories, because these are the contracts cited as evidence of a federal AI build-out. A $1.45 billion figure that is 62 percent unaccounted for in the payment column is a measurement of intent. It is being read as a measurement of activity. The distinction can be recovered, but only by reading the column that is mostly empty, and by saying which awards it is empty on.

None of this says the money was wasted or the work not done. An absent outlay is a reporting fact. The Accenture patent-search contract may have delivered exactly what it promised; the ledger has no way to tell you, and neither do I. What the record supports is narrower and, I think, more useful: when someone tells you how much the federal government is spending on artificial intelligence, ask which column they read.

The federal award figures in this piece were computed from the USAspending API v2 on 12 September 2026 and rechecked from the saved response before writing: 491 distinct contract awards, $1,449,720,817.76 obligated, $249,293,997.21 in the award-level outlay column. The query is the one in Sources below, so anyone can re-run it and compare. Two controls ran beside it: a positive one, in which the ECS Federal award must appear at its known obligation, and a negative one, in which a nonsense description filter must return nothing. Ten duplicate rows were dropped and counted. The population is awards whose description contains "artificial intelligence"; it will include boilerplate mentions and exclude AI work described as machine learning or autonomy, and the claim is about what the ledger says of the awards that call themselves AI, not about all of them.

Sources:

USAspending API v2, POST /api/v2/search/spending_by_award/, filters: award types A, B, C, D; time period 2024-10-01 to 2026-09-11; description "artificial intelligence". Run 12 September 2026. https://api.usaspending.gov/api/v2/search/spending_by_award/
USAspending API v2, GET /api/v2/awards/CONT_AWD_W911QX20C0023_9700_-NONE-_-NONE-/ (the ECS Federal award detail, including the subaward count and period of performance).
USAspending.gov, About the Data, last updated 19 November 2025, the section on agency financial data submissions (the FY2022 monthly outlay requirement and the note that earlier outlay data were optional). https://www.usaspending.gov/data/about-the-data-download.pdf
USAspending.gov, Data Sources & Methodology for Agency Submission Statistics (File C as the award-level obligation and outlay file, and unlinked File C awards). https://www.usaspending.gov/submission-statistics/data-sources
U.S. Government Accountability Office, GAO-22-105427, Federal Spending Transparency (2022). https://www.gao.gov/assets/gao-22-105427.pdf