Grants

Who funds what, and where the gaps are

Most AI funding flows through venture capital and hyperscaler partnerships, not grants. The grants tier is small by comparison, but it is where independent and sovereignty-positioned projects get their start. This page tracks who funds those projects, what they actually fund, and where the gaps are.

37 funders · 79 grants · 10 named gaps · $419M disclosed

The shape of the data

Data $230M
Weights $223M
Evaluation $157M
Compute $152M
Safety and Guardrails $144M
Governance $98M
Runtime $26M
Training $11M
Agents $7M
Protocols $2M
Identity and Trust $2M
Sovereignty and Decentralization Primitives $1M
Retrieval and Memory $200K

Disclosed dollar amounts attributed to each layer. A grant tagged with multiple layers counts toward each.

Underfunded areas

10 specific gaps where credible project shapes exist and no existing funder is sustainably covering the work. Sorted by priority. Each entry: the gap, why it matters, and concrete project shapes a new grant could fund.

high Runtime

Maintenance contracts for foundational OSS AI runtimes

Critical inference runtimes (vLLM, SGLang, llama.cpp, transformers) and evaluation infrastructure (lm-evaluation-harness) have no dedicated, sustained maintainer funding. Most maintainers are part-time, employer-sponsored, or volunteer. A single corporate exit or burnout incident threatens the open ecosystem.
A new grant could fund:
1. 01 Fixed-term funded maintainer role for lm-evaluation-harness (EleutherAI lineage)
2. 02 Fixed-term funded maintainer role for vLLM (UC Berkeley origin, LF AI hosted)
3. 03 Fixed-term funded maintainer role for llama.cpp (Georgi Gerganov, post-HuggingFace acquisition)
4. 04 Fixed-term funded maintainer role for HuggingFace transformers (Apache 2.0, broad ecosystem dependency)
high Compute

Decentralized / volunteer inference for current-gen MoE models

Petals-class decentralized inference works at small / medium model sizes but no public-goods funder is sustaining research on MoE-class inference at frontier scale. Prime Intellect and Templar have shipped decentralized training; decentralized inference of the same model sizes is mostly stalled.
A new grant could fund:
1. 01 Re-implementation of Petals for current MoE architectures (DeepSeek V3, Qwen Max class)
2. 02 Verifiable computation layer for decentralized inference (proof-of-correct-inference, ZKML-adjacent)
3. 03 Latency-tolerant inference protocols suitable for residential-internet pipelines
high Data

License-clean pretraining data corpora

EleutherAI's Common Pile v0.1 is essentially the only entity doing this work at scale. Dolma (AI2) covers some of it but uses non-Common-Pile licensing assumptions. No other lab is funding the multi-year work of curating, licensing, and shipping fully-redistributable pretraining corpora.
A new grant could fund:
1. 01 License-clean corpus for code (alternative to The Stack)
2. 02 License-clean multilingual corpus (alternative to OSCAR / mC4)
3. 03 License-clean reasoning / math corpus (alternative to proprietary AI-Math, AIME training sets)
high Evaluation

Benchmark-contamination detection tooling

Models trained on the eval data score artificially well. No public-goods detector exists that operators can run to scan a published model checkpoint against published benchmark sets and report contamination probability. The contamination-aware benchmarks (LiveBench) are partial workarounds, not solutions.
A new grant could fund:
1. 01 Open scanner that ingests a HuggingFace checkpoint + benchmark and outputs contamination score
2. 02 Continuous-monitoring service that flags new releases against contamination signatures
3. 03 Benchmark refresh framework that auto-rotates held-out items
high Data

Auditable training-data provenance

C2PA-style attestation exists for images but not for training datasets. There is no public-goods system for proving 'this dataset contains exactly these sources, licensed under exactly these terms, scraped on exactly these dates.' Without provenance, the OSAID v2.0 data-disclosure clause cannot be operationalized.
A new grant could fund:
1. 01 Open data-card generator with cryptographic provenance attestation
2. 02 Verifiable inclusion-list service for training datasets ('is my work in your data?')
3. 03 Continuous-monitoring service for web crawls that signals license changes upstream
medium Evaluation

Multilingual evaluation datasets for low-resource languages

Cohere Aya is one of few movers but covers ~100 languages. Thousands of languages with millions of speakers remain unevaluated. Without local-language evals, open-weights models cannot demonstrate non-English competence, which keeps closed-frontier models dominant in non-Anglo markets.
A new grant could fund:
1. 01 African-languages eval suite (Yoruba, Swahili, Hausa, Amharic) at MMLU-Pro depth
2. 02 Indigenous-languages eval suite (Quechua, Cherokee, Maori) at basic-reasoning depth
3. 03 Dialectal Arabic / dialectal Chinese eval suite extensions
medium Silicon

Open RISC-V AI accelerator reference design

Tenstorrent is the only company shipping open-trending AI accelerators, but their work is for-profit. No fabbable, open-licensed reference design exists for a RISC-V-based AI accelerator that academia or sovereign-tech communities can iterate on without commercial entanglement. Sovereign Tech Fund flirts with hardware but lacks AI-specific commitments.
A new grant could fund:
1. 01 Tinygrad-class open accelerator reference design (chiplet-tape-out target)
2. 02 OpenTitan-style open AI accelerator IP for FPGA bring-up
3. 03 Open MLIR / Triton dialect lowering targeting open accelerators
medium Sovereignty and Decentralization Primitives

Civic / public-comment AI infrastructure

Tools for legislatures, public consultations, FOIA, and democratic deliberation are mostly absent from the open-source AI stack. Closed vendors (Polis, some Mozilla-funded efforts) provide partial coverage; no end-to-end open civic-deliberation toolkit exists.
A new grant could fund:
1. 01 Federated, end-to-end-encrypted LLM-assisted comment synthesis for public consultations
2. 02 Open transcription + summarization toolkit for legislative hearings
3. 03 FOIA-request-assist tooling for journalists and researchers
medium Safety and Guardrails

Long-horizon sandboxing and sandbox-escape evaluations

Apollo Research and METR have done first-pass scheming evaluations. No open public-goods benchmark for agent runtime safety beyond the closed work exists. AISI Inspect is a framework, not a benchmark suite; the gap is shared evaluation data.
A new grant could fund:
1. 01 Open sandbox-escape benchmark for agent runtimes (analog to SWE-bench for safety)
2. 02 Long-horizon scheming evaluation suite (multi-step deception detection)
3. 03 Capability-token enforcement evals (does the agent honor scoped permissions)
medium Evaluation

Independent reproducibility infrastructure

Almost no funder pays for re-running released model training to verify reported numbers. Cost reasons (frontier training is $millions) plus credibility reasons (third parties cannot easily access the data and code). Without this, claimed benchmark numbers go unverified.
A new grant could fund:
1. 01 Public-goods reproduction lab for small-and-medium model training (OLMo 1B, Phi-3.5 scale)
2. 02 Continuous reproduction service that re-runs published recipes on standardized hardware
3. 03 Recipe-card standard for end-to-end training runs (data + code + checkpoints + reproduction log)

Funded vs. underfunded, by layer

Per-layer rollup of grants flowing in versus named gaps in the ecosystem. "High-priority gap" counts underfunded areas marked as having credible budgets and near-term impact in data/underfunded.yaml.

Layer	Grants	Funders	USD documented	Gaps	High-priority gaps
Data core	14	9	$230M	2	2
Weights core	12	9	$223M	0	0
Evaluation meta	33	15	$157M	3	1
Compute core	1	1	$152M	1	1
Safety and Guardrails meta	14	10	$144M	1	0
Governance meta	22	12	$98M	0	0
Runtime core	13	6	$26M	1	1
Training core	8	5	$11M	0	0
Agents core	23	9	$7M	0	0
Protocols core	7	6	$2M	0	0
Identity and Trust meta	5	4	$2M	0	0
Sovereignty and Decentralization Primitives meta	10	4	$1M	1	0
Retrieval and Memory core	6	3	$200K	0	0
Silicon core	0	0		1	0

One grant can span multiple layers (e.g., OMAI hits weights + data + compute), so the grant count column sums to more than the absolute number of documented grants.

Who funds what, and where the gaps are

The shape of the data

All grants

Active funders

Underfunded areas

Maintenance contracts for foundational OSS AI runtimes

Decentralized / volunteer inference for current-gen MoE models

License-clean pretraining data corpora

Benchmark-contamination detection tooling

Auditable training-data provenance

Multilingual evaluation datasets for low-resource languages

Open RISC-V AI accelerator reference design

Civic / public-comment AI infrastructure

Long-horizon sandboxing and sandbox-escape evaluations

Independent reproducibility infrastructure

Funded vs. underfunded, by layer