AI funding is splitting in two: megacaps and the startups losing the race

The funding gap inside AI is getting wider, and two data points published months apart show how it works. In December, sources told CNBC that Swedish vibe coding startup Lovable was raising at a $6.6 billion valuation. By April, Rest of World was reporting that global edtech investment had fallen to less than $3 billion in 2025, down from a $16.7 billion peak in 2021.

Both stories are about AI. Only one of them is about money arriving.

Lovable: three rounds in one year

Lovable's round values the company at $6.6 billion, according to two unnamed sources who spoke to CNBC. That is more than triple the $1.8 billion valuation it reached after closing its previous round in July. It is the Stockholm company's third round of 2025, and it caps a year in which Lovable went from a $1 million annual recurring revenue milestone to $200 million in ARR, reported in November.

Accel is participating, both sources said, a detail CNBC notes had not been reported before. Accel also backed the July round, which raised $200 million and included Creandum, Klarna founder Sebastian Siemiatkowski, ElevenLabs founder Mati Staniszewski and Synthesia founder Victor Riparbelli. Khosla Ventures is in the new round too, according to one of the sources. Lovable, Accel and Khosla Ventures did not respond to CNBC's requests for comment before publication.

Forbes had reported in November that the round would value Lovable at around $6 billion. The company was founded in 2023, uses models from providers including OpenAI and Anthropic, and said in November that 100,000 projects were being built on its platform every day. It is opening offices in Boston and San Francisco.

Lovable is not an isolated case. Anysphere, the maker of Cursor, raised $2.3 billion at a $29.3 billion valuation in November, CNBC reported. Replit reached a $3 billion valuation in September after raising $250 million, and Vercel closed a $300 million round at $9.3 billion.

The other side of the ledger

Edtech tells a different story, and it is not a story about AI failing. It is a story about what happens when a category loses its thesis.

Global edtech investment peaked at $16.7 billion in 2021, when lockdowns kept millions of children out of classrooms, according to Tracxn, a Bengaluru platform that tracks startup funding. By 2025, venture capital had dropped below $3 billion. Most of that money comes from US firms, Tracxn's data shows. The founder pool shrank with it: 645 companies launched in 2025, against almost 10,500 in 2020.

HolonIQ, a research firm that advises governments and investors on education, wrote in a February 6 post that capital had shifted "from volume to intention." The same post said investors concentrated money in AI-enabled products, workforce-aligned platforms and K-12 operations tools that address cost pressures, staffing problems and learning support at scale. Rest of World cites that analysis in its 23 April piece.

The failure modes were not mysterious. High customer acquisition costs, long institutional sales cycles and low retention because learning outcomes stayed murky, according to an analysis by Loot Drop, a database of more than 1,700 startup closures. Byju's, once the world's most valuable education startup at $22 billion, collapsed under a financial crisis and aggressive sales tactics for expensive courses. Nigerian startup Edukoya shut down in 2025, citing weak profitability and waning investor support.

Where online schooling is genuinely necessary, in Afghanistan where girls cannot attend school, or in war zones where buildings are destroyed, nonprofits like Khan Academy and local innovators filled the gap rather than venture-backed companies.

China shows the political version of the same collapse. The government's "double reduction" policy in July 2021 crushed the K-12 online education sector overnight. Yuanfudao, once valued at $15.5 billion and one of only two global edtech decacorns alongside Byju's, ended its core tutoring services and pivoted to AI hardware, launching "learning machines." It is now among the top six players in China's AI learning hardware market, alongside rival Zuoyebang. Other products in that space include AI-powered desk lamps and smart desks.

Loot Drop's conclusion, quoted by Rest of World, is blunt: corporate workforce development, professional certification prep and specialized skill acquisition for high-income careers look more promising than K-12 general education. The winners, it argues, will be vertical-specific tools that plug into existing workflows rather than platforms trying to replace entire educational institutions.

What the split actually measures

Read together, the two stories suggest that "AI startup funding" is not one market. Lovable sells a tool that developers and non-developers can try, share and pay for without a procurement committee. Edtech sold to school districts and ministries, where a sale could take a year and a renewal could be lost to a change in curriculum.

That difference shows up in the numbers. Lovable went from $1 million to $200 million in ARR in under a year, by its own account. Edtech went from $16.7 billion in annual investment to under $3 billion in four years, by Tracxn's count.

It also shows up in who is writing the cheques. Accel backed Lovable twice in 2025 and has put money into Cursor and into Thinking Machines, the AI company founded by former OpenAI executive Mira Murati, according to CNBC. The same firm's interest is not in education.

The risk that does not show up in a valuation

There is a third data point that belongs in this picture, and it is not about money at all.

In July, The Register reported that MeetingTV, a video conferencing and webinar startup, had sued Palo Alto Networks over a threat intelligence blog published by Koi Security, a business Palo Alto acquired. The blog, published on December 30, linked MeetingTV's Zoomcorder product to a Chinese corporate espionage operation it called DarkSpectre, and said the service acted as a "public-facing front" for the operation while lending it "credibility" and serving as a "monetization channel." The complaint alleges Koi used a large language model to generate the report and that the system hallucinated findings that were then published as fact.

MeetingTV founder and CEO Michael Robertson told The Register that Koi never contacted his company before publishing. He said he learned about the report only when service providers started blocking his domains, labelling them malware and command-and-control infrastructure. He said providers including Verizon and Palo Alto Networks were still blocking the company at the time of the article.

A Palo Alto Networks spokesperson told The Register the company was aware of the lawsuit regarding a report published by Koi before the acquisition, and declined to answer specific questions about the allegations. The spokesperson said the company believes Koi's research reflects its commitment to identifying and exposing threats, and that it expects the dispute to be resolved through the legal process.

The lawsuit claims Koi's single-actor theory rested on a fabricated technical pivot: a browser extension identified as "Twitter X Video Downloader" that, according to Robertson and the complaint, does not exist. The blog has since been silently edited to remove references to Zoomcorder, according to The Register.

Robertson's argument, as quoted by The Register, is about due process rather than about AI. He said the case matters because AI systems hallucinate and their findings should not be accepted without human review, and he listed the decisions such systems may soon make: tax status, credit ratings, university admissions, home loans, no-fly lists.

None of that is a funding story. But it is the kind of risk that does not appear in an ARR figure or a valuation, and it sits on the same technology stack that investors are currently paying a premium for.

The through line across all three stories is not that AI is working or failing. It is that capital is moving toward products with short sales cycles and measurable output, away from categories that need institutional buyers to change their behaviour, and into a technology layer whose errors can still be published as fact.