Big AI thins out the competition as startups quit the race to build large language models

The market for large language models lost another contender yesterday, while at the same time, its leader—OpenAI—closed a massive $6.6 billion funding round that represents the largest VC raise in history and will give it practically unrivaled resources to compete. 

AI startup Character.AI said it’s abandoning its efforts to build LLMs because it’s gotten “insanely expensive” to compete with the big guys, mainly Google and Microsoft-backed OpenAI. Dominic Perella, the company’s new interim chief executive, told The Financial Times that training frontier models has become “difficult to finance on even a very large start-up budget.” 

The market for generative AI models is barely two years old, and yet big tech has already eliminated much of the competition.

A new kind of acquisition thins out the model market

Character.AI secured $150 million in funding at a $1 billion valuation last year, but the respectable funding is only a dent in what’s needed to build expensive and resource-intensive models—let alone compete with companies like Google and OpenAI. This isn’t the only reason Character.AI is where it is today, however. The company is the latest AI startup to be kinda-sorta acquired by Big Tech.

In August, Character.AI and Google struck a deal in which co-founders Noam Shazeer and Daniel De Freitas rejoined Google (where they worked until 2021), as well as around 30 of the startup’s researchers. Character.AI still stands but said it will use external models and focus on developing its consumer chatbots that simulate conversations with celebrities and characters.

The structure of the deal mirrors the ones Microsoft made with Inflection and Amazon struck with both Adept and Covariant earlier this year. Critics have blasted these types of deals as a way for Big Tech to skirt antitrust scrutiny that could block a full acquisition deal. While Covariant will continue developing models for various customers, it’s still linked up with a tech giant in a big way. In other cases, like with Adept, what’s left of the startup is also ceasing to develop models (its deal with Amazon is now under FTC scrutiny). Overall, the result is a much thinner competitive market for model development and boosts for the largest players in the game.

Who can make a model?

At the same time Character.AI became the latest startup to throw in the towel on making AI models, OpenAI announced it had secured unprecedented funds while Nvidia released a new massive open-source model it says rivals OpenAI’s GPT-4o. Nvidia, of course, is among the few other big tech players with the resources to truly compete—thanks to the piles of money it makes selling the chips used for AI models. The timing of it all underscores who can compete in the LLM market and who barely stands a chance. 

As if OpenAI scoring the largest funding round in history (with an eye-popping $157 billion valuation) isn’t enough, the company is going a step further to make it more difficult for its competitors to raise funds. According to The Financial Times, CEO Sam Altman asked investors to not back rivals like Anthropic or Elon Musk’s xAI. 

In January, I published an issue of Eye on AI that asked “where did all the models go?” It discussed how enterprises aren’t making their own AI models anymore because it’s so resource-intensive to do so, and because plugging into a model developer’s API is so much easier. For customers, it’s a reasonable and strategic approach. But when it doesn’t even make sense for the startups that just recently poured millions of dollars and expert hours into developing models to continue doing so, well that says a whole lot about the state of competition in this market. 

Now, here’s more AI news. 

Sage Lazzaro
sage.lazzaro@consultant.fortune.com
sagelazzaro.com

AI IN THE NEWS

Recursion approved to begin clinical trials of AI-discovered cancer treatment. The FDA gave the AI drug discovery pioneer the green light for a phase 1/2 trial, where it will test its treatment targeted at solid tumors and lymphoma for safety, side effects, and effectiveness. The company used its AI-powered platform to identify an area of biology to target for treatment, match it with a drug candidate, and gain regulatory approval all in less than 18 months. You can read more from CNBC. 

Harvard students add facial recognition to Meta smart glasses to demo doxxing dangers. Their customized glasses combine Meta’s hardware with facial recognition service Pimeyes and other tools for looking up someone’s personal information. As 404 Media writes, the combination of the technologies “lets someone automatically go from face, to name, to phone number, and home address.” The students told 404 the project is only intended to raise awareness about the dangers of the technologies and that they won’t release the code

Another OpenAI cofounder joins Anthropic. Durk Kingma announced on X that he’s joining Anthropic, writing that “Anthropic’s approach to AI development resonates significantly with my own beliefs.” Neither he nor the company shared details about his role. Kingma left OpenAI in 2018 and was currently working at the company, but his hiring furthers a growing trend of rival Anthropic’s emergence of a landing pad for previous and recent key figures from OpenAI. You can read more from TechCrunch. 

Google DeepMind and BioNTech are developing AI research assistants for scientists. The BioNTech model, called Laila, was built on top of Meta’s open-source Llama 3.1 model, has a “detailed knowledge of biology, connects to lab devices to monitor ongoing experiments performed by robots, and can automate routine experiment tasks such as analysis and segmentation of DNA sequences, The Financial Times reported following a demo. The Google DeepMind model is still in development but will be similarly aimed at assisting researchers with experiments. 

FORTUNE ON AI

IT’S OFFICIAL: OpenAI is worth $157 billion —By Jessica Mathews

Hurricane Helene shuts down critical mines, endangering AI chip supply —By Marco Quiroz-Gutierrez

Why Aflac’s CIO won’t ever be first to market with generative AI—but is still finding ways to adopt the technology —By John Kell

AI CALENDAR

Oct. 22-23: TedAI, San Francisco

Oct. 28-30: Voice & AI, Arlington, Va.

Nov. 19-22: Microsoft Ignite, Chicago

Dec. 2-6: AWS re:Invent, Las Vegas

Dec. 8-12: Neural Information Processing Systems (Neurips) 2024 in Vancouver, British Columbia

Dec. 9-10: Fortune Brainstorm AI San Francisco (register here)

EYE ON AI NUMBERS

That’s how many billions market intelligence firm IDC predicts will be spent on generative AI software next year, according to Pitchbook’s Q3 Emerging Tech report released this week. The number is almost double the estimated $17 billion being spent in 2024 and a sharp rise from the $7.5 billion spent in 2023.

At the same time, IDC reports that generative AI LLMs still only represent a fraction of the AI market compared to legacy AI models. According to the firm’s estimates, generative AI software will contribute only about 14% of AI software spending in 2024 and is on pace to contribute only 32.3% of spending by 2028.