Major technology companies have shut down three promising artificial intelligence startups in the past six months and developed a new M&A playbook that threatens to push venture capitalists to the brink of the AI ​​boom.

Chatbot makers Inflection and Character.AI and AI agent developer Adept had collectively raised more than $2 billion in funding before their top talent was poached by Microsoft, Google and Amazon, respectively.

After these three deals, the big technology companies acquired the founders, researchers and engineers of the start-ups as well as licenses for their products. The venture capitalists, on the other hand, more or less ended up back where they started.

Their early exit is a grim sign for other AI startups trying to develop their own large-scale language models – the systems that underpin OpenAI’s ChatGPT and Google’s Gemini.

The deals will reinforce venture capitalists’ concerns that the winners of the AI ​​boom will be the largest technology companies that can bear the multi-billion dollar costs of developing cutting-edge AI systems.

Earlier this month, Google agreed to sign Character.AI co-founders Noam Shazeer and Daniel De Freitas and license the startup’s celebrity-impersonating chatbots. Character’s founders and other shareholders, the largest of which is Silicon Valley venture capital firm Andreessen Horowitz, will reportedly receive $2.5 billion as part of the deal.

This values ​​Character at 2.5 times its March 2023 price of $1 billion – a respectable but unspectacular return for investors who have invested nearly $200 million in the startup since its founding in 2022.

The merger of Microsoft and Inflection and the Adept deal with Amazon were even less lucrative for the venture capitalists of the two start-ups. They barely got back more than their original investment, say people familiar with the transactions.

The deals have provided “pretty decent” salaries for the founders who work for the big technology companies, says Mike Volpi, a partner at venture capital firm Index Ventures. But for the venture capitalists, they are “not good results.”

Since the majority of all start-ups fail, venture capitalists rely on a handful to be so successful that their miscalculations are covered many times over.

“VCs, especially those with larger funds, need above-average results (and) 2.5x your investment … is really not very useful for a single company,” said Volpi, whose firm has invested in a number of AI companies, including Cohere and Mistral.

Less than two years after the launch of OpenAI’s powerful chatbot ChatGPT sparked a flood of AI investment, many founders who left their corporate jobs to launch startups have returned to the arms of big tech companies.

Shazeer and De Freitas criticized Google’s slow pace when they left the company to start Character in 2022, but ultimately returned. Several Adept executives and Inflection founder Mustafa Suleyman were all Google researchers before founding their companies. They now work at Amazon and Microsoft, respectively.

The acquisition of the three companies by giant technology conglomerates underscores how difficult it is to scale an AI start-up. The resources required to train and run cutting-edge AI models are enormous, and start-ups without established sales channels struggle.

According to David Cahn, partner at venture capital firm Sequoia Capital, these challenges are likely to become more acute in the next phase of AI development.

Over the past year, startups have tried to get a head start by using novel research techniques or better training data. “The next phase in the AI ​​race will look different: It will be driven more by physical engineering than scientific discovery,” Cahn wrote in a recent blog, as AI companies race to build massive data centers costing billions of dollars each year to develop more powerful models.

This would likely benefit major technology companies, which increased their annual capital spending from $138 billion to $229 billion last year, he added.

Even startups that have remained independent rely heavily on partnerships with large tech companies. Microsoft has committed $13 billion to OpenAI; Amazon and Google have invested a combined $6 billion in Anthropic. Smaller players like Cohere and Mistral have also entered into partnerships with large tech companies.

Access to the computing resources and customers of the big tech companies gives startups developing large language models “an inherent advantage,” says Raviraj Jain, partner at Lightspeed Venture Partners. While only a handful of companies can afford to compete at that level, there is still room for venture-backed startups to develop smaller AI models, applications and infrastructure.

Regulatory intervention could shift the balance again. Antitrust authorities in the US and Europe are currently reviewing deals with Amazon, Google and Microsoft – despite efforts to structure the agreements so that the start-ups remain nominally independent. Last month, the UK Competition and Markets Authority announced an investigation into Microsoft’s deal with Inflection.

Venture capitalists point out that consolidation and crashes are typical in the early stages of a technology boom. Even the dot-com bust of the late 1990s couldn’t stop the Internet from becoming ubiquitous, and the most popular consumer applications didn’t appear until years after the introduction of the smartphone.

“We are undoubtedly going through a difficult time for venture capital, with all the excitement concentrated in one sector,” Volpi said. “But there are hundreds of companies doing a lot of interesting projects in AI. There are going to be some needles in that big haystack.”