Federal Grants for AI Startups: Which Agencies Actually Fund Machine Learning

{/* Schema recommendation: BlogPosting + FAQPage + Dataset.

• BlogPosting: author Nalin Vahil, datePublished and dateModified 2026-07-31. Emitted by the post template from frontmatter.
• FAQPage: emitted from the frontmatter faq block. Do not duplicate the FAQ in the body; the template renders it as an accordion below the article.
• Dataset is a strong secondary candidate and unusual for this site: the piece publishes an original derived dataset (agency and NIH institute AI/ML award counts, FY2023-FY2025) with a stated source, method, date, and license-free provenance. Mark up the agency table and the NIH institute table with name, description, temporalCoverage 2022-10-01/2025-09-30, and isBasedOn pointing at USAspending.gov.
• ItemList candidates: the five-step shortlisting method in "How to turn this map into a three-agency shortlist" and the four limitations in "What this map does not settle". Both are the strongest AI-extraction targets after the tables. */}

Cada has written 100+ grant proposals across 30+ agencies, including NIH, NSF, NASA, and DoD.

Between FY2023 and FY2025, federal agencies made between 1,058 and 1,465 AI and machine learning awards to small businesses, worth $775M to $1.06B. The three largest sources of federal grants for AI startups were NIH (490 awards, $429M), NSF (359 awards, $208M), and the Department of Defense (111 to 518 awards, $81M to $369M).

If that NIH number surprises you, you are not alone. Most founders building AI software assume NIH funds biology and NSF funds software. The award records say otherwise: NIH made more AI and machine learning small-business awards than NSF over that window, spread across 21 institutes and centers.

Everything below comes from a scan of USAspending.gov award records run on July 31, 2026, not from agency marketing pages. The exact filters are in the methodology section so you can rerun it yourself.

Why every guide to federal grants for AI startups is organized backwards

Search "federal grants for AI startups" and you get agency-first content.

One page for NSF SBIR. One page for the Air Force. One page per program, often generated at scale.

That structure only helps a founder who already knows which agency to read about.

The question you actually have runs the other direction. You built a model that does something specific, and you want to know who pays for that. Nobody publishes the map in that direction, so this piece does.

One warning before the data: this map tells you where the money has gone. It does not tell you whether you are eligible, whether a relevant topic is open right now, or whether you would win. Those are different questions, and the last section is honest about them.

How we counted (read this before you trust the numbers)

Every number here comes from the USAspending.gov spending_by_award API, queried on July 31, 2026.

The filters: fiscal years 2023 through 2025 (October 1, 2022 through September 30, 2025), recipient type restricted to small_business, and seven AI and machine learning keywords run separately then deduplicated by award ID. Those keywords were artificial intelligence, machine learning, deep learning, neural network, computer vision, natural language processing, and large language model.

Restricting to small_business is the important move. It is the cleanest available proxy for the SBIR and STTR population, because it strips out the university and national-lab awards that otherwise dominate any AI keyword search at NIH, NSF, and DOE.

Identification differs by agency, because the agencies record things differently:

• NIH: award ID prefixes R43, R44, R41, and R42 identify SBIR and STTR exactly. The institute comes from the CFDA number.
• NSF: assistance listing 47.084, filtered to awards whose description names SBIR or STTR.
• DoD: contracts under NAICS 541715 (R&D in physical, engineering, and life sciences), $50K to $2M. DoD does not flag SBIR on contracts in USAspending, which is why DoD gets a range instead of a number.
• DOE, USDA, EPA: assistance listings 81.049 and 81.135, 10.212, and 66.511.

The keyword scan returned 1,556 deduplicated records, of which 1,055 classified cleanly to an agency or NIH institute.

DoD needed a second pass. Because its contracts carry no SBIR flag, the DoD figures below come from a dedicated contract-level rescan rather than the keyword classification. That rescan finds 111 strictly-identified DoD awards where the keyword pass alone assigned 108, which is why the two totals differ by three.

These counts are floors, not a census. Keyword search only finds awards whose description text mentions AI. An award for a model that its program officer described as "predictive analytics" or "automated triage" does not show up. Read every number below as "at least this many."

Which federal agencies fund AI startups? The full map

Six agencies account for nearly all government funding for AI software companies.

Source: USAspending.gov, scanned July 31, 2026. Method above. DoD medians and recipient counts are computed on the broad population; the low end of each DoD range is the strict count.

Four things worth pulling out of that table before the agency detail.

Agency medians run from $156,492 at NASA to $399,554 at NIH. That is Phase I to early Phase II money. If you are modeling runway, a first federal award is roughly one to two engineer-years, not a Series A.

The spread inside that range tells you something. NIH's median is more than double the Air Force's $179,997, not because NIH writes larger Phase I awards but because its AI money skews toward Phase II. Most NIH institutes sit between $291,000 and $320,000; NIA, NCI, and NIMH pull the NIH-wide figure up.

Recipient counts run close to award counts. NSF made 359 awards to 330 distinct companies, NIH 490 to 381, and DoD 518 to 350. The highest repeat rate at any agency is 1.5 awards per company, at DoD.

No agency in this data is dominated by a handful of incumbents soaking up the AI budget, which is the fear most first-time applicants raise.

The long tail is real but small. DOE, USDA, and NASA together made 98 awards. Worth knowing about, not worth building a strategy around unless your technology genuinely fits.

EPA, AHRQ, and ARPA-H returned zero. More on what that does and does not mean below.

NIH artificial intelligence SBIR: 490 awards, $429M, and the biggest surprise in the data

NIH is the largest single source of AI and machine learning small-business awards in this scan.

That runs against the instinct most software founders have, which is that NIH wants wet-lab work and clinical data. NIH funds an enormous amount of software, and a meaningful share of it has no laboratory component at all: clinical decision support, imaging analysis, remote monitoring, natural language processing over clinical notes, trial-matching systems.

The catch is that NIH is not one funder. It is 21 institutes and centers with separate budgets and separate priorities, and picking the wrong one is the most common expensive mistake in NIH SBIR. The next section breaks that down.

The NIH-wide median in this data is $399,554, which mixes Phase I and Phase II. Read the institute table below rather than the NIH-wide number, because the institute you target moves that figure by a factor of five.

NSF AI startup funding: 359 awards and the widest door for a software-only company

NSF made 359 AI and machine learning SBIR and STTR awards worth $208M, at a median of $294,991.

NSF is the most straightforward entry point for a company whose product is purely software, for two structural reasons.

NSF is not mission-tied. There is no disease and no warfighter requirement your technology has to serve. The bar is technical risk plus commercial potential, which is a bar a software company can clear on its own terms.

NSF also spreads its awards widest: 359 awards across 330 companies, the largest number of distinct small-business AI recipients of any agency in the scan.

The tradeoff is that NSF knows this. It is the best-known door, and the applicant pool reflects that.

DoD AI SBIR topics: the most money, the most confusing front door

DoD is where the dollar figures get large and the counting gets hard.

The strict count, restricted to awards whose description explicitly names SBIR, STTR, or a solicitation topic number, is 111 awards worth $81M. The broad count, covering all small-business R&D contracts in the SBIR size band, is 518 awards worth $369M.

The true number is in between. The gap exists because DoD does not tag SBIR contracts in USAspending the way NIH tags grants.

Within DoD, the distribution is lopsided:

Source: USAspending.gov, scanned July 31, 2026. Rows sum to the 518-award broad count and to $369.3M. The last row is mostly contracts recorded against administering offices rather than the funding component, so read it as bookkeeping, not as a funder.

The Air Force dominates by volume, largely through AFWERX open topics, which accept a company's own proposed application rather than requiring a match to a narrow pre-written requirement. That is the single most accessible DoD door for a software company without defense experience.

Note the inverse relationship between count and median. The Air Force makes many small awards; MDA and DARPA make few large ones. Those are different strategies, and they suit different companies.

DOE, NASA, and USDA: small programs, thin competition

DOE made 50 AI and machine learning small-business awards worth $42M at a median of $318,508, concentrated in scientific computing, grid, materials, and instrumentation applications.

USDA NIFA made 27 awards worth $6.9M at a median of $175,000. Small awards, but the applicant pool for agricultural AI is thin.

NASA made 21 awards worth $7.4M at a median of $156,492, mostly in autonomy, on-orbit operations, and Earth-observation analysis.

None of these should anchor a strategy on their own. All three are worth a look if your application genuinely lands in their mission, precisely because fewer companies apply.

Which NIH institutes actually fund AI?

This is the table no competitor publishes, and it is where the useful surprises live.

Source: USAspending.gov, scanned July 31, 2026. Institute assigned from CFDA number. Institutes with fewer than 10 awards are omitted, as are 31 awards recorded under a generic NIH assistance listing.

Three findings here are worth more than the ranking itself.

The National Institute on Aging leads NIH on AI, not NCI or NIBIB. NIA made 70 AI and machine learning awards worth $70.5M, more than any other institute. Nobody guesses this.

The reason is that aging research is unusually rich in the problems machine learning is good at: early detection of cognitive decline, passive monitoring in the home, fall prediction, caregiver support tools, and analysis of long longitudinal datasets.

If you build anything that touches older adults, NIA should be on your list before NIBIB.

NIGMS ranks second with 57 awards and has no disease area. NIGMS funds basic biomedical science and research infrastructure. For an AI company, that means tooling: lab automation software, computational methods, data infrastructure, model development for biological problems.

If your product is a research tool rather than a clinical product, NIGMS is often the right institute. It is almost never the one founders name first.

NIMH has a median award of $1,658,412, roughly four times the NIH-wide median. NIMH made 34 awards, but its median is Phase II scale. That means NIMH is concentrating money in fewer, larger, later-stage projects rather than spreading Phase I bets.

If you are pre-Phase-I in mental health, that is a harder first door than the raw award count suggests. NINDS shows a milder version of the same pattern at $666,954.

The generalizable lesson: rank institutes by the problem your model solves, not by the organ or disease your training data came from.

What award data cannot tell you: the Navy problem

Here is the finding that changed how we read all of this, and the reason to distrust any agency ranking that does not disclose its method.

The Navy appears in this scan with 4 AI awards. That is wrong, and it is wrong in an instructive way.

We measured the description-field quality of small-business R&D contracts for each DoD component. The Navy writes generic descriptions: 75.4% of its contract descriptions are stub text such as "RESEARCH AND DEVELOPMENT," with a median description length of 24 characters. The Air Force writes real project titles: 4.8% generic, median length 83 characters.

Source: USAspending.gov, scanned July 31, 2026. Sample is up to 500 small-business R&D contracts per component (NAICS 541715, $50K to $2M, FY2023-FY2025), taken in descending award-amount order: 500 each for Navy, Air Force, and Army, 431 for MDA, and 122 for DARPA, which is the full available population for those last two. "Generic" counts descriptions that are boilerplate or under 25 characters.

A keyword search cannot find an AI award that was described as "RESEARCH AND DEVELOPMENT." So any AI ranking built on description text finds 386 Air Force awards and 4 Navy awards, and a founder reading that would conclude the Navy does not fund AI.

The Navy runs one of the largest SBIR programs in the federal government. The 4 is a measurement artifact.

The general rule: award databases measure what agencies write down, not what they fund. Before you rule an agency out on low counts, check whether that agency writes descriptions at all.

The same caution applies to the three agencies that returned zero in this scan. EPA, AHRQ, and ARPA-H produced no matching AI or machine learning small-business awards.

For EPA and AHRQ that is mostly a size effect, since both run small programs. For ARPA-H it reflects that the agency has no dedicated SBIR assistance listing and a short award history.

Zero results here is not evidence that any of the three funds no AI. It is evidence that this method could not see it.

We are treating this as a real limitation rather than smoothing over it. In Cada's own scoring model, the Navy is explicitly flagged as insufficient-data and left at a neutral score rather than being penalized, precisely so a data-collection gap does not get encoded as a funding judgment.

How to turn this map into a three-agency shortlist

The map is only useful if it narrows your list. Here is the method, which is the same one we run internally.

It applies whether you are looking at SBIR for machine learning companies, STTR, or agency-specific contract vehicles. The mechanism changes, the shortlisting logic does not. If you have not yet decided between SBIR and STTR, settle that first, because it changes who has to employ your principal investigator.

1. Classify your technology by the problem it solves, not the technique.

"We use transformers" is not a classification. "We reduce the time a radiologist spends on triage" is.

Agencies fund problems. Every agency in this scan has AI awards, so "we do AI" narrows nothing.

2. Pull your own counts.

Do not take this article's word for it. Query USAspending directly with your specific application terms, not generic AI vocabulary.

Use recipient_type_names: ["small_business"], a three-fiscal-year window, and an amount range of $50K to $3M for grants or $50K to $2M for DoD contracts.

Search your problem terms ("sepsis prediction," "fall detection," "spectrum sensing"), not "machine learning." Our guide to competitive research on USAspending walks through the query mechanics.

This is the most valuable hour you can spend on the whole process, because it tells you whether anyone has ever funded work like yours.

3. Check award size against your burn.

A $175,000 USDA Phase I and a $1.66M NIMH award are different products. Match the award size to what you actually need.

Chasing a large median at an institute that funds few Phase I projects is a common way to waste a cycle.

4. Check the description-quality trap before you rule anything out.

If an agency shows near-zero counts, pull 50 of its awards with no keyword filter and look at the descriptions. If they are stub text, your count means nothing and you need a different source for that agency.

5. Pick one high-volume, one mid-volume, and one adjacent.

One agency that clearly funds your category, one that funds it at lower volume with a thinner applicant pool, and one adjacent agency whose mission your technology touches sideways. Three targets, not one and not nine.

A worked example (fictional)

Say you built a model that flags medication errors from nursing-home records. Entirely made up, but the reasoning transfers.

The instinct is "healthcare AI, so NIH, so NIBIB, because that is the bioengineering one." Working the method instead:

The problem is medication safety in older adults. That is NIA's territory, and NIA is the largest AI funder at NIH with 70 awards and a $499,945 median. That is the high-volume pick.

The population gives you a mid-volume second target: NINDS if the errors are neurological, NIMHD if the angle is disparities in care quality, both in the 12-award range with thinner competition.

The adjacent pick is not NIH at all. It is the Defense Health Agency, which runs military treatment facilities and has the same medication-safety problem. Small in this data at 4 contracts, but almost nobody in digital health thinks to look there.

That is a three-agency shortlist in about 20 minutes, and the first-instinct answer, NIBIB, did not make it.

What this map does not settle

Four things, stated plainly.

Eligibility. These counts say nothing about whether you qualify. SBIR requires a US for-profit small business, more than 50% US-owned and controlled by individuals or eligible entities, with fewer than 500 employees.

The principal investigator must have the legal right to work for the small business in the US, which citizenship, permanent residency, or an appropriate visa can each establish. That standard is the same at NIH and NSF. We cover the ownership and PI rules in detail, including which problems are curable before you apply.

Whether a topic is open. Historical awards tell you an agency has funded your area. They do not tell you a relevant solicitation is open now. DoD in particular runs topic cycles where the specific topic text decides everything.

Your odds. Award counts are not win rates. This data has no denominator, because USAspending records awards, not applications.

Whether the pattern holds. SBIR reauthorization timing has been uncertain for several cycles, and agency AI priorities move faster than most federal budget lines. FY2023-FY2025 data describes the recent past. Re-pull before you commit 80 hours.

Where to go from here

The honest summary: nearly every federal research agency funds AI, the largest sources are not the ones founders guess, and the single highest-value hour you can spend is querying award records for your specific problem rather than reading agency marketing pages.

The method above is complete. Nothing is held back, and a founder with an afternoon can run all five steps without help.

If you would rather not spend that afternoon on USAspending, Cada runs a free agency-fit scan. You describe your technology, we run it against the same award data, and you get back a ranked shortlist of agencies and NIH institutes with the award counts behind each one.

Bring two paragraphs on what your model does and what problem it solves, plus your rough stage and timeline. You get the shortlist back within one business day, and there is no pitch, no obligation, and nothing to pay.

If the data says no agency has funded work like yours, we will tell you that instead of selling you a proposal.

Either way, the sequence matters more than who does it: figure out which agencies fund federal grants for AI startups working on your specific problem before you write anything. The most expensive mistake in this process is 80 hours spent on a well-written application to an agency that was never going to fund your category.

Sources

• USAspending.gov Award Search API -- the spending_by_award endpoint, recipient-type and NAICS filters, and every award record behind the tables above
• SBIR.gov -- SBIR and STTR Policy Directive, the small-business eligibility thresholds, and the list of participating agencies
• NIH SEED -- NIH small business program structure, the R43/R44/R41/R42 activity codes, and institute-level participation
• NIH RePORTER -- searchable index of NIH-funded projects by institute, useful for checking an institute's AI portfolio directly
• NSF SBIR/STTR (America's Seed Fund) -- Phase I solicitation, the technical-risk and commercial-potential review criteria, and PI work-authorization standard
• DoD SBIR/STTR Innovation Portal (DSIP) -- component topic cycles and open-topic mechanisms including AFWERX
• USDA NIFA SBIR and NASA SBIR/STTR -- program scope for the two smallest funders in this scan
• Award classification, institute assignment, description-quality sampling, and all derived medians are Cada's own analysis of the USAspending records described in the methodology section (internal Cada data, 100+ proposals across 30+ agencies)

Data: USAspending.gov spending_by_award API, FY2023-FY2025, small-business recipients, scanned July 31, 2026. Counts are floors, not a census, and the method and filters are documented above so you can reproduce them. Agency programs, topic cycles, and eligibility rules change between solicitations, so verify against the solicitation you are applying under. Nothing here is legal advice. All company examples are fictional and used for illustration only.