AI is a cross-cutting federal priority, which means nearly every SBIR agency now funds AI-related R&D. NSF has a dedicated AI topic area. DARPA's biggest office is focused on information and AI innovation. NIH funds clinical AI across 27 institutes. NIST is building the AI safety measurement ecosystem. Even USDA and EPA fund AI applied to their domains.
The result: AI startups have more federal grant options than almost any other sector. The challenge is knowing which agency funds which type of AI, and which path gives you the best shot.
This guide covers every federal program that funds AI at startups -- agency by agency, with amounts, eligibility, and where each type of AI innovation fits best.
The AI grant landscape at a glance
| Agency | Program | Phase I | Phase II | Best For | Application Style |
|---|---|---|---|---|---|
| NSF | SBIR | $305K | $1.25M | Broad AI/ML R&D | 4-page pitch, rolling |
| DARPA | I2O BAA | Varies ($500K-$10M+) | N/A | Defense/intel AI, trustworthy AI | Abstract → full proposal |
| NIH | SBIR | $314K | $2.1M | Clinical AI, health analytics | Standard proposal, 3 cycles/yr |
| NIST | SBIR | $100K | $400K | AI safety, measurement, standards | Annual solicitation |
| IARPA | BAAs | $500K-$5M | N/A | Intelligence-community AI | BAA response |
| DoD | AFWERX SBIR | $75K | $1.25M | Defense-relevant AI | Open Topic, rolling |
| DoD | Navy/Army SBIR | $256K | $1.7M | Domain-specific defense AI | Topic-based, periodic |
| DOE | SBIR | $200K | $1.1M | Scientific computing AI, energy AI | Annual solicitation |
| ARPA-H | SBIR/ISOs | $600K | $3.5M | Breakthrough health AI | PM-directed |
NSF SBIR: the default starting point for most AI startups
If you're building AI technology and have never applied for a grant, start here. NSF has the widest scope, the most accessible process, and the best fit for commercial AI innovation.
What NSF funds in AI: Deep learning, NLP, computer vision, robotics, AI infrastructure, reinforcement learning, generative AI (with genuine technical innovation), AI for scientific discovery, trustworthy AI, neuromorphic computing, edge AI, and AI applied to nearly any domain. NSF's AI topic area is intentionally broad.
Award amounts:
- Phase I: $305,000 (6-12 months)
- Phase II: up to $1,250,000 (24 months)
- Supplemental funding can push Phase II past $1.75M
The process: NSF uses a two-step model. First, submit a 4-page Project Pitch -- a lightweight description of your technology, innovation, and commercial potential. NSF reviews it in about 3 weeks and tells you whether to submit a full proposal. This means you invest minimal effort before knowing if NSF is interested. Submissions are rolling (no fixed deadlines).
Why NSF works for AI startups:
- No preliminary data required (unlike NIH)
- No defense application needed (unlike DoD)
- Rolling submissions -- apply anytime
- PI citizenship not required
- Broadest technology scope of any agency
The catch: NSF funds R&D, not product development. You need to demonstrate genuine technical uncertainty and innovation. "We're building an AI chatbot" isn't enough. "We're developing a novel retrieval-augmented architecture that solves [specific technical problem] where current approaches fail" is the framing that works.
Best for: Commercial AI/ML platforms, AI infrastructure, domain-agnostic AI technology, AI applied to underserved domains, trustworthy AI.
DARPA: the biggest checks for breakthrough AI
DARPA funds AI at a scale and ambition level that no other agency matches. If your AI technology is genuinely novel -- not incremental -- DARPA is where the biggest awards and strongest validation come from.
What DARPA funds in AI: The Information Innovation Office (I2O) covers trustworthy AI, explainable AI, resilient software, human-AI teaming, and cybersecurity. The Defense Sciences Office (DSO) funds fundamental AI research. The Strategic Technology Office (STO) funds AI for large-scale military systems. The Microsystems Technology Office (MTO) funds AI hardware.
Award amounts: No fixed limits. Proposals can range from $500K to $10M+ depending on the technical approach and program scope. Smaller feasibility studies start around $500K. Full program performers can receive $5-10M+ over multiple years.
The process: DARPA publishes office-wide BAAs (Broad Agency Announcements) that accept proposals on a rolling basis. The I2O BAA is currently open through late 2026. You submit a brief abstract first. If DARPA is interested, they invite a full proposal. DARPA program managers have significant discretion -- they're looking for specific capabilities, not checking boxes.
Why DARPA works for AI startups:
- Largest award sizes for AI R&D
- Strong validation signal (DARPA funding impresses investors and customers)
- No size restrictions on the company
- Program managers can move fast when they see something they want
The catch: DARPA wants 10x improvements, not incremental advances. The bar is genuinely high. Most DARPA awards go to established defense contractors and university labs. Pre-seed startups can win, but it's rare without either a prior DARPA relationship or a technology that's clearly ahead of the field. For most AI startups, DARPA should be a second or third grant application, not the first.
Best for: Novel AI architectures, AI safety and alignment, adversarial AI, human-AI teaming, AI for cybersecurity, autonomous systems, AI hardware.
NIH: clinical AI, health analytics, and biomedical ML
If your AI is applied to healthcare, NIH is the richest single funder. Each of NIH's 27 Institutes and Centers runs its own SBIR topics, and many now prioritize AI/ML.
What NIH funds in AI:
- Clinical decision support systems
- Medical image analysis (radiology, pathology, dermatology)
- Drug discovery and molecular design ML
- Genomics and precision medicine algorithms
- Mental health screening tools
- Population health analytics
- Electronic health record mining
- Clinical trial optimization
Award amounts:
- Phase I: up to $314,363 (6-24 months)
- Phase II: up to $2,095,748 (1-3 years)
- Commercialization Readiness Pilot: up to $4.19M
The process: NIH runs three review cycles per year with fixed deadlines. Standard SBIR proposals (no pitch step like NSF). Reviewed by study sections -- panels of domain experts who score on a 1-9 scale (1 is best). Match your technology to the right institute: NCI for cancer, NIMH for mental health, NHLBI for cardiovascular, NIAID for infectious disease, NIA for aging, etc.
Key differences from NSF:
- Preliminary data is critical -- NIH reviewers expect evidence that your approach works. First-time applicants without any data face steep odds.
- PI citizenship not required (same as NSF)
- Individual institutes may set lower budget caps than the SBA maximum
- Academic credentials carry more weight -- PhDs are common (not required) among successful NIH PIs
Best for: AI/ML applied to clinical medicine, diagnostics, drug discovery, health data analytics, genomics, public health.
NIST: AI safety, measurement, and standards
NIST is a smaller player but increasingly important as AI safety regulation takes shape. If you're building AI safety tooling, evaluation frameworks, or measurement infrastructure, NIST is a natural fit.
What NIST funds in AI: AI safety measurement, bias detection and mitigation, AI system evaluation, adversarial robustness testing, AI standards development, trustworthy AI infrastructure, and AI applied to manufacturing and measurement science.
Award amounts:
- Phase I: approximately $100,000 (6 months)
- Phase II: up to $400,000 (24 months)
Why NIST matters for AI safety startups: NIST published the AI Risk Management Framework and is building the evaluation ecosystem around it. Companies funded by NIST SBIR to develop AI safety tools are positioned at the center of the emerging regulatory landscape. The award amounts are small, but the strategic value -- building measurement tools that become industry standards -- can be significant.
The catch: NIST has a very specific scope. Your AI needs to connect to measurement, standards, or safety evaluation. General-purpose AI doesn't fit.
Best for: AI safety tooling, bias auditing, model evaluation platforms, AI testing infrastructure, trustworthy AI standards.
DoD SBIR: AI for defense applications
If your AI has defense relevance -- autonomous systems, cybersecurity, intelligence analysis, logistics, situational awareness -- the Department of Defense is the largest single SBIR funder at roughly $2.3 billion annually.
AFWERX (Air Force) is the fastest entry point. Their Open Topic SBIR accepts any defense-relevant AI. Phase I is small ($75K, 3 months) but Phase II jumps to $1.25M. STRATFI/TACFI post-Phase II programs can extend to $3-15M. Start here if you've never done defense. See our AFWERX SBIR guide.
Navy and Army SBIR publish specific AI topics through defensesbirsttr.mil. Phase I up to $256K, Phase II up to $1.7M. Topics include C4ISR, autonomous navigation, predictive maintenance, cyber operations, and sensor fusion.
DARPA SBIR operates alongside DARPA's BAA process. Standard DoD SBIR amounts ($250K Phase I, $1.7M Phase II) but DARPA's scope and review process.
Best for: Autonomous systems, cybersecurity AI, computer vision for defense, predictive maintenance, logistics optimization, human-AI teaming.
IARPA: intelligence-community AI
IARPA funds AI for the intelligence community through rolling BAAs. Awards range from $500K to $5M per performer depending on the program. Topics include facial recognition, foreign language NLP, geospatial intelligence, forecasting, and deception detection.
The catch: May require security clearances for some programs. Export control restrictions apply. Not a first-time-applicant-friendly agency.
Best for: AI for intelligence analysis, OSINT, advanced NLP, geospatial AI, predictive analytics for national security.
DOE: AI for science and energy
DOE SBIR covers AI applied to scientific computing, energy systems, materials discovery, and climate modeling. Phase I up to $200K, Phase II up to $1.1M.
ARPA-E funds AI for transformational energy technology -- grid optimization, battery design ML, carbon capture modeling. Awards average $2-3M but competition is fierce (2-5% acceptance rate).
Best for: Scientific ML, materials discovery, energy optimization, climate modeling, computational chemistry.
ARPA-H: breakthrough health AI
ARPA-H is NIH's moonshot sibling. SBIR Phase I goes up to $600,000 (nearly double NIH's cap). Phase II up to $3.5M. Non-SBIR programs can reach $50M+ per performer.
ARPA-H wants 10x improvements in health outcomes -- not incremental advances. If your AI could fundamentally change how a disease is detected, treated, or prevented, ARPA-H is the right funder. If it's an efficiency improvement to an existing workflow, stick with NIH.
Best for: AI diagnostics that detect conditions years earlier, AI-driven drug discovery platforms, precision medicine at population scale, AI for pandemic preparedness.
Which agency fits which type of AI?
| Your AI Innovation | Best Agency | Second Choice | Why |
|---|---|---|---|
| Novel ML architecture or training method | NSF | DARPA | NSF is broadest; DARPA if defense-relevant |
| Clinical decision support | NIH | ARPA-H | NIH for proven approaches; ARPA-H for moonshots |
| Computer vision (general) | NSF | AFWERX | NSF for commercial; AFWERX if defense-dual-use |
| Computer vision (medical imaging) | NIH (NCI or NIBIB) | NSF | NIH if health-specific; NSF if broadly applicable |
| NLP / language models | NSF | IARPA | NSF for commercial; IARPA for intelligence |
| AI safety / alignment | NIST | NSF | NIST for measurement; NSF for research |
| Autonomous systems | AFWERX | DARPA | AFWERX for entry; DARPA for breakthrough |
| Drug discovery ML | NIH (NCATS or disease IC) | ARPA-H | NIH for standard track; ARPA-H for platform plays |
| Cybersecurity AI | NSF | DHS | NSF for research; DHS for homeland security |
| Scientific ML / simulation | DOE | NSF | DOE for energy/materials; NSF for general science |
| AI hardware / neuromorphic | NSF | DARPA (MTO) | NSF for research; DARPA for defense hardware |
| Edge AI / embedded ML | NSF | AFWERX | NSF for commercial; AFWERX for tactical |
What doesn't qualify: the R&D test
Not all AI work is grant-fundable. SBIR funds R&D -- work where the outcome is genuinely uncertain. Here's the line:
Fundable:
- Building a model architecture that doesn't exist yet
- Developing a training approach for a domain where existing methods fail
- Creating an AI system that operates under constraints (latency, privacy, adversarial conditions) that current solutions can't handle
- Advancing AI safety evaluation or alignment methods
Not fundable:
- Fine-tuning GPT-4 on your industry's data
- Building a wrapper around existing APIs
- Applying standard ML libraries to a new vertical without algorithmic innovation
- Building a better UX on top of existing AI infrastructure
The gray area is real. If you're unsure whether your AI qualifies, NSF's Project Pitch is the lowest-risk test -- 4 pages, 3-week review, zero cost beyond your time.
Applying to multiple agencies: the stacking strategy
AI startups are uniquely positioned to apply across agencies because AI is relevant to nearly every federal mission. The rules allow simultaneous applications as long as the proposed work doesn't overlap. Here's how to do it:
Same platform, different proposals:
An AI company building a medical image analysis platform could submit:
- NIH (NCI): Proposal focused on cancer detection accuracy, clinical validation, and oncologist workflow integration
- NSF: Proposal focused on the novel computer vision architecture, training methodology, and generalizability of the approach
- AFWERX: Proposal focused on battlefield medical triage application of the same imaging technology
Each proposal emphasizes a different aspect of the platform for a different agency mission. This is legitimate and common among sophisticated applicants.
Sequencing for maximum impact:
- Start with NSF (lowest barrier, fastest feedback via Project Pitch)
- Apply to the mission agency matching your domain (NIH, DOE, DoD)
- Use Phase I results from one agency to strengthen proposals at others
- Phase II at multiple agencies can yield $2-4M+ in non-dilutive funding
The bottom line
AI startups have the broadest federal grant landscape of any sector. The combination of NSF's open scope, DARPA's large checks, NIH's deep health funding, and NIST's emerging AI safety focus means there's a federal path for nearly every type of AI innovation.
The founders who capture the most funding don't just pick the agency with the biggest check. They match their specific AI innovation to the agency whose mission it serves, start with NSF to learn the process, and build a multi-agency pipeline that compounds over time.
Want to know which AI programs fit your specific technology?
The matching table above gives you the starting point. But which specific topics, institutes, or BAAs are the strongest fit -- and how to frame your AI innovation for each agency's reviewers -- is where expert guidance pays for itself. Our Strategy Review identifies the 3-5 programs where your AI startup has the best shot, ranked by fit, award size, and timeline.