How Google put together the pieces for its AI comeback

When ChatGPT launched in 2022, Google was caught flatfooted, but the launch of Gemini 3 and the Ironwood AI chip this month has experts raving about Alphabet’s AI comeback. 

Google kicked off November by unveiling Ironwood, the seventh generation of its tensor processing units, or TPUs, that the company says lets customers “run and scale the largest, most data-intensive models in existence.” And last week, Google launched Gemini 3, its latest artificial intelligence model, saying it requires “less prompting” and provides smarter answers than its predecessors.

Salesforce CEO Marc Benioff captured the excitement around Gemini 3 with a Sunday post on X, saying that despite using OpenAI’s ChatGPT daily for three years, he wasn’t going back after two hours of using Gemini 3.

“The leap is insane,” wrote Benioff, whose company has partnerships with Google, OpenAI and other frontier AI model providers. “Everything is sharper and faster. It feels like the world just changed, again.”

Most tech stocks were down to start the week, except for one: Alphabet.

Shares of the Google parent surged more than 5% on Monday, adding to last week’s gain of more than 8%. Warren Buffett’s Berkshire Hathaway revealed earlier this month that it owns a $4.3 billion stake in Alphabet as of the end of the third quarter.

Alphabet shares are up nearly 70% this year and have outperformed Meta’s by more than 50 percentage points this year, and last week, Alphabet’s market cap surpassed Microsoft’s.

All of this came despite Nvidia reporting stronger-than-expect revenue and guidance in its third-quarter earnings last week.

“You may be asking why almost all of the AI stocks we cover are selling off after such good news from Nvidia,” Melius Research analyst Ben Reitzes wrote in a note Monday, referring to Nvidia’s positive third quarter earnings last week. “There is one real reason for worry and it is the ‘AI comeback’ of Alphabet.” 

But while Google appears to have regained the edge, its lead over rivals remains razor thin in the gruelingly competitive AI market, experts said.

“Three years ago, they were seen as kind of lost and there were all these hot takes saying they lost their way and Sundar is a failure,” Nathanson said. “Now, they have a huge leg up.”

The company had a number of AI product mishaps in its initial attempts to catch up with OpenAI. In 2024 alone, Google had to pull its image generation product Imagen 2 for several months after users discovered a number of historical inaccuracies. The launch of AI Overviews caused a similar reaction when users discovered it gave faulty advice, which the company later remedied with additional guardrails.

“There was a lot of fumbling, and they were scrambling,” said Gil Luria, managing director at technology research firm DA Davidson. “But they had the tech in the pantry, and it was just a matter of getting it all together and shipped.”

Of particular note is how quickly Google launched Gemini 3 after the spring release of Gemini 2.5, which was already considered an impressive model. The hyper-realistic image generation features of Nano Banana is another notch in Google’s belt. After the company initially launched the image generation tool, Gemini shot to the top of the Apple App Store in September, dethroning ChatGPT.

And after the launch of Gemini 3, Google released Nano Banana Pro last week.

Google’s ownership of YouTube and all the content on the video platform gives the company an edge when it comes to training models for image and video generation.

“The amount of video and current data that Google has, that’s really a huge competitive advantage,” said Mike Gualtieri, vice president and principal analyst for Forrester Research. “I don’t see how OpenAI and Anthropic can overcome that.”

Additionally, Google has successfully incorporated its AI models into its enterprise products, driving sales for the company’s cloud unit. In its third quarter earnings results last month, Google reached its first $100 billion quarter, boosted by its cloud growth. The company’s cloud unit, which houses its AI services, showed solid growth and a $155 billion backlog from customers.

And it’s not just the AI models. Google is also garnering attention with its AI chips.

Google says Ironwood is nearly 30 times more power efficient than its first TPU from 2018. Google’s ASIC chips are emerging as the company’s secret weapon in the AI wars and have helped it notch recent deals worth billions with customers such as Anthropic.

After a report said that Meta could strike a deal with Google to use its TPUs for the social media company’s data centers, Nvidia saw its stock drop 3% on Tuesday, prompting the chipmaker to post a response on social media.

With the rise of Google’s TPUs, Nvidia may no longer have the AI chips market cornered.

“The advantage of having the whole stack is you can optimize your model to work specifically well on a TPU chip and you’re building everything to a more optimally designed,” said Luria.

The company’s ability to serve AI enterprise customers with its TPUs and Google Cloud offerings as well as its incorporation of Gemini 3 throughout its consumer products is driving Wall Street’s enthusiasm.

Experts who spoke with CNBC said the competitive landscape is broader than just one AI winner, but they added that it’s become increasingly expensive for multiple companies to prove success.

“Having the state of the art model for a few days doesn’t mean they’ve won to the extent that the stock market is implying,” Luria said, pointing to Anthropic’s new Opus 4.5 model launched Monday.

Earlier this month, OpenAI also announced two updates to its GPT-5 model to make it “warmer by default and more conversational” as well as “more efficient and easier to understand in everyday use,” the company said.

“The frontier models still seem to be neck and neck in some ways,” Forrester Research’s Gualtieri said.

The competitive edge will likely go to the companies willing to spend more money given the expenses of the AI race, experts said. In their earnings reports last month, Alphabet, Meta, Microsoft and Amazon each lifted their guidance for capital expenditures. They collectively expect that number to reach more than $380 billion this year.

“These companies are spending a lot of money assuming there’s gonna be a winner take all when in reality we may end up with frontier models being a commodity and several will be interchangeable,” Luria said.

For Google, maintaining a lead in AI won’t be without challenges.

Company executives told employees earlier this month that Google has to double its serving capacity every six month to meet demand for AI services and run its frontier models, CNBC reported last week.

“The competition in AI infrastructure is the most critical and also the most expensive part of the AI race,” Google Cloud Vice President Amin Vahdat told employees.

Although Google’s in-house TPUs have gotten increased attention as viable alternatives to Nvidia’s Blackwell chips, Nvidia still holds more than 90% of the AI chip market.

In its post on Tuesday, Nvidia pointed out that its chips are more flexible and powerful than ASIC chips, like Google’s Ironwood, which are typically designed for a single company or function.

And despite getting Salesforce’s Benioff to switch to Gemini, Google also has a lot of catching up to do with its consumer chat product, experts said, citing hallucinations and lower user numbers than OpenAI’s.

The Gemini app has 650 million monthly active users and AI Overviews has 2 billion monthly users, Google said last month. OpenAI, by comparison, said in August that ChatGPT hit 700 million users per week.  

“Yes, Google has got its act together,” Luria said. “But that doesn’t mean they’ve won.”

WATCH: AI narrative is shifting towards Google with its complete stack, says Plexo Capital’s Lo Toney