Big money investments, not acquisitions, fuel GenAI startups

2023 AI partnerships

It has been a year of big AI partnerships.

AWS revealed plans to invest up to $4 billion in AI safety and research startup Anthropic. In return, the big tech vendor will take a small minority stake Anthropic.

Google has also invested in Anthropic. The cloud giant, which has its own large AI R&D team, has pumped about $300 million into Anthropic. It's also reportedly in talks to invest hundreds of millions of dollars in AI chatbot startup Character.ai. The smaller vendor provides services that let users chat with a virtual celebrity or anime characters.

In the biggest move that exemplifies the new tech giant AI investment strategy, Microsoft has committed $13 billion in services and cash to ChatGPT and Dall-E creator OpenAI, which was founded as a nonprofit in 2015 and changed to a for-profit company in 2019. Microsoft is the exclusive provider of the computing power the startup needs for research, products and programming interfaces.

Meanwhile, both Oracle and AI hardware/software giant Nvidia have invested in foundation model provider Cohere in deals that placed the value of the AI startup at about $2.2 billion. Oracle's investment will loosely mirror the relationship between Microsoft and OpenAI, with Oracle integrating Cohere's LLMs into its products. In June, both Oracle and Nvidia participated in a Series C funding round in which Cohere raised $270 million.

Microsoft, AWS, Oracle, OpenAI, Anthropic and Cohere did not respond to requests for comment about the partnership deals.

With all these investments, acquisitions have been sparse.

Other than Databricks' $1.3 billion acquisition of MosaicML in June and Thompson Reuters' purchase of Casetext for $650 million in August, the pace of AI acquisitions has slowed dramatically from its peak in the mid-2010s.

For example, the second quarter of 2023 saw no AI acquisitions by Amazon, Apple, Google and Microsoft, according to data from CBInsights.

Only two AI acquisitions were revealed year up to now by the tech giants. Apple in March acquired, for an undisclosed price, WaveOne, a small developer of algorithms for compressing video. Also in March, EU regulators cleared Google to acquire AI math vendor Photomath, reportedly for several hundred million dollars.

A bias towards partnerships

The pattern of partnership versus acquisition has been around for a while, according to some industry experts.

"There's definitely a natural bias towards partnerships," said Kjell Carlsson, head of AI strategy at independent MLOps platform vendor Domino Data Lab and a former tech analyst.

The tech giants have long been hesitant to make big AI acquisitions, even before the Microsoft-OpenAI partnership.

That is because the bigger vendors have been unwilling to confront the challenge of integrating and trying to gain value from acquired companies, Carlsson said.

Some of the acquisitions that the tech giants have revealed in recent years were for AI startups burning through cash and often looking to sell themselves, he said.

"There was an opportunity where they couldn't really partner with this company long term because they weren't going to be around long term," Carlsson said.

For example, in 2021, HPE acquired Determined AI, a startup with a software stack for training AI models faster.

Other acquisitions aimed to grab a capability that the tech vendor did not have.

For example, when Databricks acquired MosaicML earlier this year for about $1.3 billion, the vendor wanted to beef up its generative AI credentials, Carlsson continued. Often, tech vendors gain little value from acquisitions.

"It's rarer to find an instance where one of the acquisitions really did pay for itself than it is to find ones where it didn't," he said.

Survival necessity

Therefore, partnerships are usually a survival necessity for both tech vendors and startups, according to Kashyap Kompella, analyst and CEO of RPA2AI Research.

Sometimes, a new technology may lead a vendor to acquire a startup but that could lead to the failure of both the vendor and the startup.

A classic example of this was IBM Watson, the tech giant's early AI system.

"In the previous wave of AI, Watson was the first poster child of AI success," Kompella said.

Because of Watson's success, IBM pumped about $4 billion into building a Watson Health portfolio and acquired multiple startups.

Watson Health failed to deliver on its promise, and IBM sold it Watson Health technology assets for a reported $1 billion in 2022.

"The story does not always end well," Kompella said. "Being a first mover is not necessarily an advantage."

Therefore, the strategy of big tech vendors like Microsoft, Amazon and Oracle this year and the end of 2022 has been more about cautiously taking advantage of the technology these generative AI startups offer.

"It gives them a strategic optionality," Kompella added. If their investment succeeds, it will boost their existing technology offerings. However, they end up not suffering huge losses like they would if the technology did not turn out well and they had acquired the AI startup, he added.

The benefit for vendors and startups

Investing also enables the tech giants to innovate faster, Gartner analyst Jim Hare said.

"By supporting innovations more hands off, they're able to support faster innovation and develop capabilities faster than if they tried to do it the traditional way," he said.

The clearest example of this is the Microsoft-OpenAI alliance. Through Microsoft's support of OpenAI, the tech giant spent much of 2023 incorporating generative AI technology swiftly into its products and was even able to challenge Google's longstanding search model.

Meanwhile, the startups can still take advantage of the infrastructure and resources the hyperscalers have.

"It's so expensive to go in and train these [LLMs]," Carlsson said.

Building LLMs requires substantial computing resources, extensive storage and a big server capacity that the giant cloud providers possess.

Startups also gain access to the customer and business relationships that big tech giants have established, said Mark Beccue, Futurum Research analyst.

Since academic researchers launch start many AI startups, "they don't have the sales channels," he said. "It's a big deal."

This means that many startups have to make decisions about their go-to-market capability. Thus, having the guidance of a massive tech giant like Microsoft, Amazon or Oracle will help them establish the right go-to-market strategy.

For instance, OpenAI has been able to run and power its technology from the as part of the tech giant's investment relationship instead of having to pay for it directly.

Despite the success of ChatGPT, part of OpenAI's success and recent shift and pivot into the enterprise market is due to Microsoft's move to incorporate OpenAI's technology into all its office productivity software in addition to Bing.

Similarly, Cohere trains its generative AI models on Oracle Cloud Infrastructure, and that gives Oracle customers access to Cohere's language models. Anthropic is also gaining by having access to AWS' purpose-built AI chips. It also now has access to AWS customers.

An open relationship

However, within these mutually beneficial relationships, the big tech giants and the AI startups aren't locked into the deals.

While Anthropic uniquely relates to AWS in some ways, it also partners with Google, another investor. Despite Oracle's partnership with Cohere, the latter still offers its AI models on AWS.

This flexibility also benefits tech giants like AWS and Google in that they can offer more choices to enterprise clients, Beccue said. For example, AWS offers its own Amazon Titan foundation models but still invests in Anthropic while offering Cohere's two LLMs and other open source LLMs to its enterprise customers.

By supporting innovations more hands off, they're able to support faster innovation and develop capabilities faster than if they tried to do it the traditional way.

Similarly, Google has developed its own foundation models, such as Gemini, Codey and Imagen, but still invests in Anthropic. It also offers open source models such as Llama 2 from competitor Meta in its model garden.

"We think that it's of benefit to our customers to be able to use and test different models and decide which one is best for their purpose," said Lisa O'Malley, senior director of product management at Google Cloud, during an interview with TechTarget Editorial about the latest updates to Vertex AI search.

However, the open relationship has recently been rocky for Microsoft and OpenAI.

When OpenAI's board decided to fire CEO Sam Altman without alerting Microsoft, Microsoft saw a dip in its stock. This made Microsoft scramble to hire Altman so the tech giant could try to maintain supremacy in the AI arms race.

The turn of events led some to question why Microsoft, OpenAI's biggest investor, is not a part of its board. It also revealed that even with open partnerships, the two companies can have different goals.

This episode exposed the conflict inherent in OpenAI's structure, with a volunteer board overseeing a for-profit company.

The state of generative AI

The different options show that tech giants understand that their customers want choice. It also speaks to the state of the generative AI market.

"It's too early to tell who's going to win," Beccue said. "The LLM and foundation model companies are morphing. It's amazing how quickly they're changing."

Moreover, because the market is so early, there have been few proven enterprise-grade generative AI applications, Beccue added.

Many enterprises are still figuring out how the applications will benefit them beyond employees being able to use a consumer product like ChatGPT for basic tasks.

Many enterprises are still not using the voice, language, text or image applications that are now available, Carlsson said.

"A lot of the ways in which folks are looking to build businesses off of generative AI haven't been done before," he said. "They're new to mainstream businesses."

Due to this newness, despite the success of AI startups the tech giants are investing in, only some of the startups will prove successful over time.

"It's almost certain that much in the same that there's been an explosion of startups, there's going to at some point be a shakeout of them," Carlsson continued.

Open versus closed

Another dimension to the emerging AI partnership paradigm is the choice between open and closed source foundation models in the generative AI market.

While the closed models created by startups such as Anthropic, Cohere and OpenAI work when tested, when incorporated within an enterprise, their behavior could change. Most enterprises are not keen on this because they want to avoid surprises, Hare said.

Alternatively, when enterprises use open source models, they have more control over what they can change and ensure that their data is protected. Data security is a concern with some of OpenAI's tools.

Open source models are also more easily adapted for vertical applications that are fit for internal uses within an enterprise, Hare added.

"What we may see is a new crop of vendors emerge where they're building out more of these vertical-focused models that are more relevant for enterprise use cases," he said.

Therefore, while startups like OpenAI, Anthropic and Cohere might have a head start, especially with significant investments from tech giants, an influx of new startups means that other startups could quickly them as market conditions continue to change, Carlsson said.

Lofty valuations and hype

However, some, such as OpenAI, have achieved a status that could makes them too big to fail, Kompella said. For Microsoft, that means it will likely glean much value from its stake in the company,

OpenAI is reportedly talking to investors about a share sale that would bring its value to $80 to $90 billion. That kind of valuation means that it would be nearly impossible for any company to acquire OpenAI, Hare said.

"Maybe when some of the valuations come down and the GenAI hype sort of dies down, that will maybe create more of a normal market where acquiring some of these companies makes sense," he said. "It's going to be some time here until things start dying down and people are focused on the value of GenAI."

Esther Ajao is a TechTarget Editorial news writer covering artificial intelligence software and systems.