To Build A.I. Technology, Start-Ups Turn to Bigger Rivals for Help

The tech industry loves garage startup stories. From Hewlett-Packard to Google, stories of bootstrapped companies becoming giants have inspired generations of entrepreneurs.

But the huge amount of money and computing power required for start-ups trying to make it with today’s hottest technology, the artificial intelligence used by chatbots like ChatGPT and Google Bard, can be makes inspirational stories a thing of the past.

In 2019, Aiden Gomez and Nick Frosst left Google to create an AI start-up in Toronto called Cohere that could compete with their former employer. A few months later, they returned to Google and were asked if they would sell the massive computing power they needed to build their own AI technology. After Google’s chief executive, Sundar Pichai, personally approved the arrangement, the tech giant gave them what they wanted.

“It’s ‘Game of Thrones.’ That’s it,” said David Katz, a partner at Radical Ventures, the first investor in Cohere. Big companies like Google, Microsoft and Amazon, he added, control the chips. “They control the power to control. compute,” he said. “They choose who gets it.”

Building a groundbreaking AI startup is difficult without enlisting the support of “hyperscalers,” who control large data centers capable of running AI systems. And that puts industry giants in the driver’s seat — again — for what many expect to be the most important shift for the tech industry in decades.

OpenAI, the start-up behind ChatGPT, recently raised $10 billion from Microsoft. This will return most of Microsoft’s money as it pays for time on the large clusters of computer servers that the larger company runs. Containing thousands of specialized computer chips, these machines are essential to improving and expanding the expertise of ChatGPT and similar technologies.

Competitors won’t be able to keep up with OpenAI unless they can get their hands on the same amount of computing power. Cohere recently raised $270 million, bringing its total funding to more than $440 million. It will use most of the money to buy computing power from the likes of Google.

Other start-ups have made similar arrangements, notably a Silicon Valley company called Anthropic, founded in 2021 by a group of former OpenAI researchers; Character.AI, founded by two leading researchers from Google; and Inflection AI, founded by a former Google executive. Inflection raised a $1.3 billion funding round last week, bringing the total to $1.5 billion.

At Google, Mr. Gomez is part of the small research team that designed Transformer, the underlying technology used to create chatbots such as ChatGPT and Google Bard.

Transformer is a powerful example of what scientists call a neural network — a mathematical system that learns skills by analyzing data. Neural networks have been around for a long time, helping to drive everything from talking digital assistants like Siri to instant translation services like Google Translate.

The Transformer took the idea into new territory. Running on hundreds or even thousands of computer chips, it can analyze vast amounts of data, much faster.

Using this technology, companies such as Google and OpenAI have begun to develop systems that learn from large amounts of digital text, including Wikipedia articles, digital books and chat logs. As these systems analyze more and more data, they learn to create text on their own, including term papers, blog posts, poetry and computer code.

These systems – called large language models – now underpin chatbots such as Google Bard and ChatGPT.

Before the advent of ChatGPT, Mr. Gomez left Google to start his own company with Mr. Frosst and another Toronto entrepreneur, Ivan Zhang. The goal is to create great language models that rival Google.

At Google, he and his fellow researchers have access to an almost unlimited amount of computing power. After leaving the company, he needed something similar. So he and his co-founders bought it from Google, which sells access to the same chips through cloud computing services.

Over the next three years, Cohere built a massive language model that rivaled almost any other. Now, this technology is being sold to other businesses. The idea is to give any company the technology they need to build and run their own AI applications, from chatbots to search engines to personal tutors.

“The strategy is to create a platform that others can build on and experiment with,” Mr. Gomez said.

OpenAI offers a service along the same lines called GPT-4, which many businesses already use to build chatbots and other applications. This new technology can analyze, create and edit text. But it will soon handle images and sounds as well. OpenAI has prepared a version of GPT-4 that can analyze a photo, describe it immediately and even answer questions about it.

Microsoft’s chief executive, Satya Nadella, says the company’s arrangement with OpenAI is the kind of mutually beneficial relationship small competitors have long nurtured. “I grew up in a company that was always doing these kinds of deals with other companies,” he told The New York Times earlier this year.

As the industry races to match GPT-4, traders, investors and experts debate who will be the winners. Most agree that OpenAI is leading the field. But Cohere and a small group of other companies are building similar technology.

The tech giants are in a strong position because they have more resources needed to maintain these systems than anyone else. Google also holds a patent on Transformer, the foundational technology behind AI systems built by Cohere and other companies.

But there is one wild card: Open source software.

Meta, another giant with the computing power needed to build the next wave of AI, has just opened the source of its latest multilingual model, meaning anyone can use it. again and build upon it. Many in the field believe that this type of freely available software will allow anyone to compete.

“Having the collective minds of every researcher on Earth can defeat any company,” said Amr Awadallah, chief executive of AI startup Vectara and a former Google executive. But they still have to pay for access to larger competitor data centers.