Unpacking Big Tech's quasi-acquisitions of GenAI companies
The rapid takeover of top AI talent
Hi readers, I will be moderating a panel of leading AI investors from Lightspeed (Mistral AI, InWorld AI), Khosla Ventures (OpenAI, Cognition), and IVP (Perplexity, Glean) at the AI Conference in San Francisco next month. In case you want to attend, use my promo code “VC” to save a few hundred dollars with a 30% discount. Tickets are available here: Eventbrite.
In the past few months, Microsoft, Amazon, and Google have engaged in quasi-acquisitions of foundation model companies Inflection, Adept, and Character.ai, respectively. Instead of a traditional acquisition involving stock purchases, these Big Tech firms opted to hire key AI talent from these startups and pay the investors license fees. In this article, I analyze these deals through a corporate development lens, offering details and insights beyond what has already been covered in the media:
Why Microsoft, Amazon and Google pursued those deals
Why they resorted to an unusual deal structure
Why they paid so much money
Who else could acquire and be acquired
Why quasi-acquisitions won’t become a trend
Overview of the deals
Why acquire?
Speed is the primary reason why Microsoft, Google, and Amazon spent hundreds of millions (perhaps billions) on these deals. They are rushing to become market leaders because the rewards are high, competition is intense, and there is a data flywheel effect that accumulates over time.
Reward for winning is high. The revenue potential of the generative AI market is enormous, and the pace of growth is unprecedented. For example, in June 2024, Sam Altman shared in a company-wide meeting that OpenAI’s annualized revenue is $3.4 billion—121 times more than the $28 million revenue the company generated in 2022, making it the fastest-growing company in history. It was estimated that 55% of this revenue comes from individuals (ChatGPT), and the other 45% from businesses, suggesting that both B2B and B2C markets are lucrative.
Competition is intense. With such a large prize at stake, competition is fierce. All the big tech companies are spending billions to build data centers to support AI models. Microsoft’s exclusive relationship with OpenAI drives both Azure spend and access to OpenAI’s latest tech, which is integrated into Microsoft’s other products. Amazon has a deep relationship with Anthropic through a $4 billion investment. If AI cloud spend goes to Microsoft, it might also attract other cloud spending. Google’s search dominance is at risk if it doesn’t build an AI layer on top of its search results, and its cloud prospects might also suffer. Google committed to investing $2 billion in Anthropic.
In addition to these investments, the Big Tech companies have been investing approximately $15 billion each in cloud infrastructure every quarter, and they expect this to increase over the next several quarters.
Alphabet/Google (7/23/2024): “Our reported Capex in the first quarter was $12 billion. Looking ahead, we continue to expect quarterly Capex throughout the year to be roughly at or above the Q1 Capex of $12 billion. We continue to invest in designing and building robust and efficient infrastructure to support our efforts in AI given the many opportunities we see ahead.”
Microsoft (7/30/2024): “Capex, including financing leases, were $19 billion. Cloud and AI-related spend represents nearly all of our total Capex. We expect Capex to increase on a sequential basis, given our cloud and AI demand as well as existing AI capacity constraints.”
Meta (7/31/2024): “We anticipate our full year 2024 Capex will be in the range of $37 billion to $40 billion. We currently expect significant Capex growth in 2025 as we invest to support our AI research and our product development efforts. Llama 3 is already competitive with the most advanced models and we’re already starting to work on Llama 4. The amount of compute needed to train Llama 4 will likely be almost 10 times more than what we used to train Llama 3.
Amazon (8/1/2024): “For the first half of the year, Capex was $30.5 billion. Looking ahead to the rest of 2024, we expect capex to be higher in the second half of the year. The majority of the spend will be to support the growing need for AWS infrastructure as we continue to see strong demand in both generative AI and our non-generative AI workflows.”
Data flywheel. As discussed in my past essay on data moats, there is a data flywheel in AI products. User data helps build better models and products. Researchers estimate that LLMs will have been trained on all public internet data by as early as 2026. Proprietary user data becomes even more important in building better models.
Now that we know why the big companies made these acquisitions, there are two questions: why the quasi-acquisition structure, and why such large amounts of money?
Why not a standard acquisition structure?
The structure of the quasi-acquisitions consists of two parts:
Talent Acquisition: The big tech companies hire the founders and key employees of these startups, selectively acquiring the talent they want. Microsoft hired over 80% of Inflection AI’s team, Amazon hired 66% of Adept’s team, and Google hired a smaller fraction of Character AI’s team.
Non-exclusive technology licensing: Instead of acquiring the entire company, the Big Tech firms obtain non-exclusive licenses to the startups' models. The license fees are then used to pay investors a negotiated return over time.
Why not just buy the shares like a regular acquisition?
This is to avoid further regulatory scrutiny, which would be triggered by an HSR filing. An HSR filing (Hart-Scott-Rodino Act) is a pre-merger filing to both the FTC (Federal Trade Commission) and DOJ (Department of Justice) and is required when it meets certain thresholds such as a $120 million deal size. After filing, the agencies will have 30 days to do an initial review. During this period, the agencies can approve the transaction or request for more information. Companies must substantially comply with the data requests. That takes weeks to prepare. Agencies can take weeks to months more to review the new information, which could lead them to challenge the deal. Given the strategic importance of speed and certainty to close, Big Tech firms (like all firms) prefer to avoid this scrutiny, especially when the FTC announced in January 2024 that it had already launched inquiries into AI deals done by them. While the HSR Act applies to acquisitions of assets and voting securities, the definition of assets conveniently excludes non-exclusive licenses and people.
Why not just give large bonuses to talent and avoid paying license fees?
Investors would block the acquisition. If key talent is hired, the startup’s value plummets. The license fees give investors a satisfactory return on their investments. Since the founders are also owners, license fees serve as exits for them. For example, in the Microsoft-Inflection deal, Series A investors will get 1.5 times their $225M investment, and Series B investors will get 1.1 times their $1.3B investment. From the date of the funding announcement to the date of the acquisition announcement, Series A investors achieve a 25% IRR, and Series B investors, 14%. Google is more generous with a 2.5x return. Adept investors get their money back. It is likely that these returns correlate with consumer product presence:
Adept is targeting enterprises and has not publicly released a product yet
In March 2024, Inflection reportedly had 1 million daily active users
In July 2024, Character reportly had 6 million daily active users
Why that much money?
If the acquired companies were struggling, as media reports suggest, how did the teams at Microsoft, Amazon, and Google justify the high “license fees” they paid? It’s not because of the models themselves. They obtained non-exclusive licenses to models that will be outdated in six months. With a gutted team and license fees going to investors, the acquired companies have no resources to continue building frontier models. So it’s a mix of other factors:
Talent: The primary driver behind these deals is the acquisition of top AI talent. These startups have assembled teams of the best researchers and engineers, and Big Tech companies are willing to pay a premium to bring them in-house. This talent acquisition accelerates their own AI development efforts and potentially prevents competitors from gaining an edge.
Dataset: While the models might not be the primary focus, the datasets these startups have accumulated can be invaluable. These datasets, often fine-tuned for specific tasks or domains, can provide a significant advantage in training and refining the acquiring company's models. It was reported that Amazon will gain access to Adept's dataset, which likely includes valuable proprietary data related to web browsing, productivity tools, and other user interactions as part of the deal. Unlike public web text, user interaction data is hard to obtain.
GPU access: Acquiring startups that have secured significant GPU capacity can help Big Tech companies overcome potential bottlenecks and accelerate their AI initiatives. Inflection, for example, had secured access to a cluster of 22,000 NVIDIA H100 GPUs, one of the largest in the world, which further incentivized Microsoft's acquisition.
Users: Users provide a testing ground for new AI features and products, and their engagement data can further enhance the acquiring company's models. Additionally, a loyal user base can translate into potential revenue streams and market share. Character AI, with monthly traffic rivaling Gemini, presented an attractive target for Google in this regard.
VC approval: Venture capitalists who have invested in these startups expect a return on their investment. The high price tags associated with these quasi-acquisitions ensure that investors are adequately compensated, preventing potential roadblocks.
Who else could be quasi-acquisition targets?
According to The Information, there are seven AI startups that are struggling and may look for buyers: Reka (raised $56 million, valued $300 million), Imbue (raised $232 million, valued >$1 billion), Cohere (raised $980 million, valued $5.5 billion), Pika Labs (raised $141 million, valued $470 million), Ideogram (raised $97 million, valued $580 million), AI21 Labs (raised $336 million, valuation $1.4 billion), and Essential AI (raised $65 million)
But who could potentially buy these companies? Acquiring model-centric companies only makes sense for cloud service providers. Given the additional scrutiny the recent deals have attracted, it is unlikely they would pursue another high-profile acquisition. IBM and Oracle, the 4th and 5th largest US cloud providers, are unlikely to acquire talent to start building their own frontier models. Two application-layer companies, Salesforce and Adobe, could be potential buyers. Salesforce has cloud ambitions and a broad surface of applications to deploy AI talent. Similarly, Adobe also has a wide range of applications, including its family of Firefly models. But neither Salesforce nor Adobe has antitrust concerns regarding AI. They would likely buy the stocks directly or pursue a typical acquihire.
Interesting reads: