TheChek News Flow

As we wrapped up our generative AI takeover list and our fifth annual most promising startups list, I got to thinking: how many of the companies will still be independent a year from now?

Already, this year has seen a flurry of acquisitions and quasi-acquisitions as startups look for a safe haven in the arms of larger tech companies. And there are likely more such deals to come—bankers tell me that many potential acquirers and targets are waiting for the results of next week’s presidential elections to suss out what the antitrust regulatory environment might look like before beginning deal talks.

When we think about what AI deals might make the most sense for next year, there are a couple likely takeover candidates that we’ve written at length about, such as embattled AI wearable-device maker Humane and Stability AI, which now develops AI image and video technology aimed at Hollywood. But here are theoretical pairings you might not have thought about:

Cohere + SAP, Oracle, Salesforce or Snowflake: Out of the major closed-source model developers, Cohere has so far trailed behind rivals when it comes to generating revenue. A few months before it received an investment that valued it at a whopping $5.5 billion, the Canadian model developer was making just $35 million in annualized revenue, compared to the $3.4 billion in ARR OpenAI was making just a few months later (and that figure is now around $4 billion). With rising competition, it might make more sense for Cohere to sell.

One possible buyer would be SAP, which has already invested in Cohere and has yet to make a major AI acquisition. As a German company, SAP may also be interested in Cohere’s recent work with multilingual models. Another option is Oracle, which has also invested in Cohere and uses the startup’s models to power AI features in its HR software.

While Cohere’s general-purpose large language models have fallen behind in evaluations such as the Lmsys leaderboard, some developers tell me they’re impressed by the retrieval augmented generation features Cohere offers, which can help to reduce bad results from LLMs by providing the models with access to data whose accuracy has been vetted. (More on RAG here.)

Those capabilities could also be helpful for two other potential acquirers, Salesforce and Snowflake, that want to help businesses use their private sales or enterprise data in conjunction with LLMs. Snowflake has already discussed partnerships with Cohere, as I’ve previously reported.

Mistral + Apple, Microsoft or Amazon: My colleague Kalley has written about the challenges Mistral faces in making money from open-source models. The list of companies that could acquire the Paris-based startup is short, given its $6 billion valuation. It could make sense for Apple to buy Mistral, given that the iPhone maker has been slow to develop its own conversational AI and is relying on outside providers, starting with OpenAI, to handle complex tasks that its own AI cannot. Mistral has also recently doubled down on small, lightweight models that can run on devices, which could make it especially interesting for applications on the iPhone or Macbook.

Another option is Microsoft, which yesterday continued its pattern of diversifying from OpenAI by incorporating models from the ChatGPT maker’s rivals into its GitHub Copilot coding assistant. Microsoft has clearly wanted to move closer to Mistral, as it invested $16 million in the startup earlier this year. Perhaps Mistral’s AI talent could help Microsoft in developing its MAI-1 model, which we haven’t heard much about since we first reported on it in May. (To be fair, a spokesperson for Microsoft said Mistral isn’t a candidate for acquisition.)

Mistral’s AI talent could also be attractive for Amazon, which is similarly developing its own LLM, Olympus, but whose past attempts at building its own conversational AI haven’t gone so well. For now, though, Amazon is relying in part on Anthropic’s technology to power AI products and features, so an Anthropic-Amazon mega-deal is not out of the question either, as my colleagues have previously covered. Amazon is, of course, a big investor in Anthropic.

Modular + Advanced Micro Devices: One of Nvidia’s biggest advantages is Cuda, its software for writing machine-learning apps that only works with its own chips. Cuda isn’t particularly easy to use, but it’s so ubiquitous that most developers working with AI chips have gotten used to its quirks and are essentially locked into using the software with Nvidia graphics processing units. The startup Modular, which was last valued at $600 million, is working to change that by developing a Cuda alternative that’s simpler to use and works with non-Nvidia hardware as well. That could be the secret weapon to help chipmaker AMD catch up with Nvidia, as it could theoretically make AMD chips easier for developers to use.

First, though, we’ll have to see how well that software works. CEO Chris Lattner recently told my colleague Akash that the product is set to launch sometime this fall.

Perplexity + Meta: Perplexity is no stranger to acquisition offers: as my colleague Kevin reported in this profile of its CEO from Friday, the search engine startup has received acquisition interest from companies including Microsoft, X, OpenAI and Notion. Though it seems like none of those deal talks have panned out (for now), another natural partner for the startup could be Meta, which my colleague Kalley reported this week is developing its own AI search engine to lessen its reliance on Google and Microsoft’s Bing.

Perplexity could accelerate that process for Meta by providing the company with its proprietary search engine bot that crawls and indexes websites or the algorithms it uses to rank and write out search results. Meta would first need to figure out how much of Perplexity’s results come from proprietary technology versus relying on back-end tech from other parties. And Meta is more likely to be more interested in a deal if its leaders believe Perplexity’s app will become the next big thing with consumers. Only that would justify paying anywhere near the $8 billion price that Perplexity wants from investors in an upcoming funding round.

Clay or 11x + Salesforce: If Clay and 11x sound familiar, it’s probably because both early-stage startups landed on our most recent promising startups list for helping salespeople automate the process of finding and reaching out to prospective customers. Though Salesforce is already developing AI software for salespeople, it has stopped short of saying it wants to replace salespeople by fully automating their roles. If it ever decides to move in that direction, these startups could be of interest. They have had early success in getting other software companies to pay for tech that replaces junior salespeople. (Check out Jon and Aaron’s piece from yesterday on the boom in AI sales tech startups.) Plus, Salesforce CEO Marc Benioff loves acquisitions more than almost any CEO we know!

Hugging Face + Meta: One of Meta’s biggest challenges in AI is getting big businesses to use its Llama open-source models. One way to solve that problem is to buy a popular community of open-source developers through an acquisition of Hugging Face, which was last valued at $4.5 billion.

Today, Meta lets Hugging Face host its Llama models for developers to download, and Hugging Face gives Meta some basic stats, like the number of times Llama models are downloaded and the geographic distribution of users, according to a person familiar with their relationship. However, Meta might benefit from Hugging Face’s data on non-Meta models and datasets, like which model types, architectures, datasets, libraries and modalities are trending more broadly in the open-source developer community. Meta could then use that data to improve its Llama models and understand how developers are using conversational AI.

Plus, an acquisition would extend the halo effect Meta has received from positioning itself as a supporter of the open-source AI community. One potential complication to this idea: Hugging Face on Monday publicly supported a new definition for open-source AI that wouldn’t include Llama because Meta requires apps with more than 700 million monthly active users to request a special license to use Llama models.

Cognition + Microsoft: One of the most successful early uses of generative AI has been generating code, which has consequently triggered VCs to back coding assistant startups. That’s also made the market increasingly competitive and crowded.

One of the startups worth paying attention to is Cognition, whose buzzy Devin coding assistant helped it raise multiple funding rounds in quick succession earlier this year, reaching a $2 billion valuation just six months after it was founded. Since then, though, I haven’t heard a ton about the startup, and it still has yet to make Devin widely available to users. It might make sense for Microsoft to acquire Cognition and use its tech—which helps Devin search the web, reference outside files and run autonomously to build websites and products without much prompting from the user—to improve GitHub Copilot, Microsoft’s AI coding assistant. Some developers say GitHub Copilot is less sophisticated than alternative coding assistants such as Cursor or even OpenAI’s ChatGPT, which works as a bona fide coding assistant too. And now, OpenAI is developing an even more sophisticated coding assistant.