AI Startup Funding Rounds 2021-2026

Published: 20 October 2026| Version 1 | DOI: 10.17632/p53tv2j78f.1
Contributor:
Fliar BI

Description

Deal-level funding rounds of AI companies worldwide 2021-2026, stage, amount, lead investor. Sample release of 5,000 of 8,122 funding rounds. Published by Fliar BI (https://fliarbi.com), licence CC BY-NC 4.0, contact b2b@fliarbi.com. Key findings (computed on the full data): 1. 5,412 companies raised 8,122 rounds between 2021 and mid-2026. 6,974 rounds disclosed an amount, totalling $301.5B. 2. 2024 was the busiest year, with 1,768 rounds. Disclosed funding in 2025 was $83.2B, 1.8× the 2021 level ($47.3B). 3. Mega-rounds dominate the dollars. Rounds of US$100M or more are 11% of rounds with a disclosed amount, but 67% of disclosed dollars. 4. Early rounds got bigger. The median Seed round went from $3.0M in 2021 to $4.8M in 2025; the median Series A from $13.0M to $16.5M. 5. Funding is concentrated. The top three countries (United States, United Kingdom, Canada) hold 79% of disclosed dollars. 6. The fastest-rising sectors (share of rounds, 2021–2022 vs 2024–2025): Business Services (3.8% → 6.0%), Enterprise Infrastructure (6.3% → 8.5%), Consumer (2.6% → 3.9%). Also on: kaggle https://www.kaggle.com/datasets/fliarbi/ai-startup-funding-rounds-2021-2026; hf https://huggingface.co/datasets/fliarbi/ai-startup-funding; zenodo https://zenodo.org/record/23431429; github https://github.com/Busineessfree/ai-startup-funding How this dataset was acquired. Every row comes from the Radar API (https://github.com/fliarbi/radar-api-mcp), Fliar's company-data API and MCP server. Radar reads business news continuously and extracts each funding announcement into structured fields: company, stage, amount, currency, investors and date. It then links the company to its record in a database of five million companies. This release covers AI companies worldwide, January 2021 to June 2026. It runs that output through a reproducible cleaning pipeline: repeated reports of one deal are merged, recaps of past rounds are dropped, and every amount is checked against the sentence it was read from. Each row names the article it was read from (news_source, news_url) and quotes the sentence (news_excerpt). The full, continuously updated data is available through POST /v1/companies/funding (reference (https://github.com/fliarbi/radar-api-mcp/blob/main/docs/reference/get_company_funding.md)) or the get_company_funding MCP tool. This file is a 62% sample: 5,000 of 8,122 funding rounds. It keeps every record for a fixed subset of companies. The subset is chosen by a hash, so it stays the same from one release to the next. The findings and tables below were computed on the full data. For the rest, the withheld fields and monthly updates, get a Radar API key (https://fliarbi.com/radar-api/developer) (docs (https://github.com/fliarbi/radar-api-mcp)). Each row's fliar_profile_url opens the company's full profile on Fliar.

Files

Steps to reproduce

1. Extraction. Radar's pipeline reads each funding article and extracts structured fields from it, including the source sentence. 2. Back-fill days removed. Two days on which archived articles were loaded with the load date as their publication date (2024-04-22 and 2025-02-13) are excluded, because those rounds carry no reliable date. 3. Scope. Venture-style instruments only (pre-seed to Series H, growth/PE, bridge, convertible notes, venture debt, grants, strategic). IPOs, QIPs, private placements, rights issues, bank loans, bonds and acquisitions are dropped. Companies founded in 2000 or later; rounds after a company's IPO and group subsidiaries of large conglomerates are excluded. 4. Deduplication. Reports about the same company and stage within 45 days are one round. A round repeating an earlier amount within 180 days is dropped, as are lettered stages below one already announced (recaps). One round filed under two domains of the same company is merged. 5. Amount check. Each amount is compared with the money figures parsed from its source sentence (tolerance 0.6–1.7×). verified matched; unchecked had no parseable figure; failed_check is blank because it did not match, exceeded a stage ceiling, or was a $500M+ figure reported by one article only; not_disclosed had no amount. 6. Currency. US dollars, converted at the time of extraction and rounded to three significant figures. 7. News columns. The excerpt is the earliest reporting sentence that does not name a data provider. Source and URL come from that article when it can be retrieved. 8. Sample. Each company key is ranked by sha1(key). Keys are taken in rank order until the file holds max(25%, 5,000) rows (capped at 100,000). 9. Profile links. Links are built from Fliar's company index and checked against it. A row whose company has no public profile has an empty link.

Categories

Finance, Entrepreneurship, Venture Capital

Licence