TheChek News Flow

The Takeaway

Google’s improvements to its Gemini AI models are boosting the company’s top line.

Over the past year, Google’s business selling access to its Gemini AI models has skyrocketed, reflecting the improving quality of those models, according to three people with knowledge of Gemini’s sales. That’s likely to lift revenue from Google Cloud’s core business of server sales, as money customers spend on AI tends to lead to additional spending on other Google products, according to a person familiar with Google’s sales.

And it could also bolster Google’s still-nascent business selling software on top of those models, which has grown quickly in recent months but has garnered mixed reviews from customers.

The improvements are likely to show up in Google’s fourth-quarter earnings, which the company is scheduled to report on Feb. 4. Investors will be looking for signs that Google is making a return on the extraordinary investments it has made over the past year as it competes on AI. Google said last fall that it expected to spend between $91 billion and $93 billion on capital expenditures, including to support its AI business, nearly double the $52.5 billion it spent on capex in 2024. In addition, the company has been hiring specialized talent for Google Cloud and its AI research unit, Google DeepMind (for more on that, see this story).

Google sells access to its models through Google Cloud via an application programming interface. Requests sent to the Gemini API, known as API calls, more than doubled from around 35 billion in March, when Gemini 2.5 was first released, to around 85 billion in August, according to internal data reviewed by The Information.

The big challenge Google still faces is persuading businesses to pay up for sophisticated software it has developed to run on top of its AI models. That software offers it a way to raise the profit margins on its AI business, according to people who have worked at Google Cloud. The strategy’s centerpiece is Gemini Enterprise, which combines access to Google’s Gemini chatbot with the ability to search across company data as well as a platform for building and using AI agents—AI that can complete more complex tasks.

A Google spokesperson said Gemini Enterprise has grown to 8 million subscribers across 1,500 companies, as well as over 1 million subscribers who have signed up online, a sign it is gaining ground. Google is expected to highlight the growth of Gemini Enterprise when it reports earnings in February.

“We are seeing tremendous momentum throughout our Cloud business, particularly our AI applications,” a Google spokesperson said.

But according to interviews with several customers, cloud consultants and employees, customer feedback hasn’t been entirely positive. Simon Margolis, an executive at Sada, a consultant that specializes in advising customers on Google Cloud services, said a slim majority of customers he talked to had favorable experiences with Gemini Enterprise, but “it’s close to 50-50 of customers who like it or don’t.”

In some cases, the success of Google’s Gemini models—which customers use to build their own software—may hinder its ability to sell its software built on top of them, Margolis said.

“Google has always been much more of a builder’s cloud than a ‘buy a product’ cloud,” Margolis said. “You have a customer who could buy Gemini Enterprise or build a bunch of custom agents, and they will probably build a bunch of custom agents. Google makes it so easy to build something like that.”

Google’s business selling access to its models had a slow start, as its early Gemini models weren’t particularly well received. But after Google released Gemini 2.5 last spring, to enthusiastic reviews from developers, the API sales took off.

So many customers, including staff at coding tool startup Cursor, wanted to use the Gemini API after the 2.5 release that Google had to tweak how the model was delivered to make it more efficient. That adjustment freed up enough compute to accommodate the additional usage, according to a separate person with knowledge of the situation.

There was another usage surge after Gemini 3 was released in November, also to strongly positive reviews, another of the people said.

Google has worked to improve API margins over the past year. Google’s first Gemini models, Gemini 1.0 and Gemini 1.5, had negative profit margins, meaning they cost more to operate than Google charged, due to its heavy discounting, one of the people said.

Gemini 2 only sometimes had positive margins. Gemini 2.5 had positive margins, as the improved quality of the models meant Google could compete on quality, not just price, according to the person, but even that only accounted for the cost of serving the tokens itself, not any of the costs that went into model development.

As of the middle of last year, Google’s blended Gemini margins—across all models—were barely positive, far below margins on Cloud as a whole, according to another person with direct knowledge of the situation.

However, the success of the Gemini API boosted Google’s business beyond just direct API sales. Google calculated that spending on API calls also increased spending on other Google Cloud products like storage or databases, the person said.

Where Google still has work to do is in its software AI application suite, Gemini Enterprise. According to Margolis, the product has received mixed reviews all year from customers.

Niko Gruben, head of applied research at Harvey, a legal AI startup, said Gemini Enterprise was useful for him because it connected to company data. He uses it for deep research based on internal data, writing long-form documents and generating images, but he hasn’t used the agent builder.

Mark Shank, who leads consulting firm KPMG’s business selling Google Cloud to its clients, said 83% of KPMG employees are satisfied with Gemini Enterprise, according to internal surveys. He said it hasn’t been hard to drive adoption among KPMG clients.

But according to Margolis, the consultant who specializes in Google Cloud services, Gemini Enterprise isn’t good at building custom applications. Another consultant whose company uses Gemini Enterprise tried to build an agent to summarize his emails, one of Gemini Enterprise’s suggestions, and was unable to do so, according to a screenshot of the chatbot exchange viewed by The Information.

Chirag Mehta, a principal analyst at Constellation Research, said customers told him Gemini Enterprise was good at answering shallow, general-purpose questions based on enterprise data and at generating code based on their code repository, but it struggled when asked very specific questions or instructed to perform specific tasks.

Still, Mehta said, the customers he talked to liked the product enough that they haven’t given up on it yet, unlike with competitors like Copilot.

“People are saying, ‘We’re going to give it a shot,’” Mehta said.