Google Unveils Data Center-Focused Axion CPU

(Bloomberg) -- Google unveiled a host of updates to its artificial intelligence offerings for cloud computing customers, emphasizing that the technology is safe and ready for use in the corporate realm, despite recent stumbles in consumer-facing tools. 

At the company’s annual cloud computing conference in Las Vegas on Tuesday, Cloud Chief Executive Officer Thomas Kurian showed off how Google’s most powerful AI model, Gemini, can be used to create advertisements, ward off cybersecurity threats and spin up short videos and podcasts. The company also touted a new chip designed to handle the massive AI workloads and control the associated rising costs.

Corporate customers will be able to peg Gemini’s query responses to reliable sources of information, known as grounding. The company is rolling out the use of Google search results as a source for the AI model’s answers, thereby providing greater accuracy and freshness, Kurian said.

“Enterprises have been piloting with us a number of scenarios with generative AI; now they’re deploying them in production,” Kurian said in an interview with Bloomberg ahead of the announcements. “The capabilities to do things like grounding, improving correctness of answers – all of those, step by step, people have gotten comfortable, they’re seeing value, and they’re deploying as a result.”

Related:TSMC Gets $11.6 Billion in US Grants, Loans for Chip Plants

Google, a unit of Alphabet, trails Amazon.com and Microsoft Corporation in cloud computing, but the market is one of the tech giant’s best bets for growth as its core search advertising business matures. Google reported the first full year of profitability at its cloud unit in 2023 and hopes to use its prowess in AI to close the gap with rivals. After OpenAI’s ChatGPT burst onto the scene in late 2022 and was quickly embraced by the general public, Google and its cloud competitors see 2024 as the year the technology conquers the corporate world.

Much of the battle for AI superiority relies on having powerful semiconductors that are able to handle all the data being processed. Google is following Amazon’s AWS and Microsoft in designing its own microprocessors using technology from UK-based chip designer Arm Holdings. The company announced Tuesday that it would roll out a new central processing unit, called Axion, that’s built with data centers in mind. Google said the chip will perform better and be more energy efficient than existing models. Kurian called the technology "a major new advance."

Google has been developing its own custom silicon since 2015. But Axion represents Google’s aim to control its own destiny in semiconductors and reduce dependence on third-party vendors. The move further narrows the opportunity for its traditional chip providers –  Intel Corporation and Advanced Micro Devices –  and adds to Arm’s foothold in the lucrative data center market. Google also announced that it would deepen its partnership with Nvidia Corp., which has established a commanding lead in AI chips.

Related:Google to Invest $1bn in New UK Data Center to Meet Demand

Google’s chief rival in artificial intelligence, the Microsoft-backed startup OpenAI, is also courting corporate customers. OpenAI now has more than 600,000 people signed up to use ChatGPT Enterprise, up from around 150,000 in January, Chief Operating Officer Brad Lightcap said last week.

Google’s enterprise push follows some embarrassing setbacks in the consumer market. In February, its flagship artificial intelligence product Gemini, which ingests enormous volumes of digital media to train software that predicts and generates content in response to a prompt or query, was roundly criticized after it spit out historically inaccurate images. CEO Sundar Pichai blasted the responses as “completely unacceptable,” and the Mountain View, California-based company stopped accepting prompts for people in its image generator while it works to address the concerns.

Yet Kurian presented generative AI in the enterprise space as a very different story. Businesses can use Gemini to create images for advertising campaigns, but the pro tool comes with 19 different controls to help marketers ensure that the content is in keeping with their brand, Kurian said. 

Despite the fallout over the Gemini images, Google Cloud has continued to allow corporate customers to generate images of people using the enterprise version of the tool – and no customers have complained about the results, Kurian said.

“We had zero, zero issues with the reported issues that people had on the consumer side with Gemini for Workspace,” he said. “There was not a single customer affected by it because we have capability in our enterprise platform for the company to control various elements of factuality, safety, model safety, responsibility.”

 Those controls will now be augmented by the ability for corporate clients to ground Gemini’s responses in Google search. When this feature is enabled, the AI model will produce citations for every sentence of its outputs, based on its retrieval of information from Google search results. In a demonstration with Bloomberg on Friday, hours after an earthquake struck New Jersey, a Google employee showed how the default version of the model stated that there had been no recent earthquakes in the area; the version of the model grounded in Google search results correctly gave the magnitude of the temblor and said there had been no major reports of damage.

Corporate clients can also ground the model’s responses in their company’s data, or even a specific portion of an employee manual – in contrast to the consumer version of Gemini, which is more a one-size-fits-all tool.  

Google Cloud’s app developer platform, called Vertex AI, is adding new features underpinned by Gemini 1.5 Pro, which Google has said has the “longest context window” of any large-scale AI model. Gemini 1.5 Pro can process up to 1 million “tokens” – essentially, words or pieces of words – at a time, according to the company, including audio. That means developers can ask the AI model for responses based on hundreds, or potentially thousands, of images, videos, documents and audio files. 

In a demonstration for Bloomberg, a Google Vertex AI product leader showed how Gemini 1.5 Pro works with Google Workspace. Cloud customers can upload marketing images and other media to Google Drive and ask the AI model to create new content such as a slideshow or a podcast based on a brand’s style. Users can also ask the AI model for “live images,” a four-second moving image showing a particular product within a scene. For example, Nenshad Bardoliwalla, the Vertex AI product leader, generated an image of a yellow camping tent against the backdrop of a gently babbling brook. Google said the images generated by its Vertex AI platform would include digital watermarks to signal they were generated by AI.

While last year Google touted how its AI tools could be used to complete everyday corporate tasks such as composing email and marketing copy, this year the company extended their capabilities to include more behind-the-scenes work.

The company also rolled out a series of Gemini applications for cybersecurity, which Google says will help clients analyze threats and address potential vulnerabilities. The features build on Google’s $5.4 billion acquisition of cybersecurity firm Mandiant in 2022.  Google’s AI-powered security features can help companies be more proactive about combating bad actors, said Eric Doerr, vice president of cloud security engineering. “What otherwise would be very manual research tasks” can be aided by AI, he said. 

Google was keen to point out how it works with the burgeoning crop of AI startups, which it sees as a key source of cloud business. Many of the most prominent young AI companies were founded by former Google employees, and they make for desirable clients given the tremendous amount of computing power they require. 

Google Cloud has seen an increase in business from startups using its platform to build generative AI apps and services, Kurian said. More than 60% of generative AI startups that have raised funding are now paying for Google’s cloud computing services, Kurian said. Of those valued at more than $1 billion – colloquially referred to as “unicorns” – 90% are Google Cloud customers, up from 70% in August, he said. New clients include AssemblyAI, a company building AI models for transcription, and Writer, a startup focused on custom generative AI apps like chatbots.  

Competition for such clients is fierce, and many of the startups use multiple cloud providers. Amazon said more than 5,000 generative AI startups were customers of its AWS platform as of September.

Kurian said the product updates flowed from close collaboration between his cloud unit and Google DeepMind, the company’s premiere AI lab, led by Demis Hassabis. The lab was the product of the merger of two research groups last year, a move Google made to bring its full talent to bear in the intensifying AI race. Engineers from the cloud and research organizations work closely together – in some cases sitting side by side – to sharpen product focus, Kurian said. 

“That collaboration is up and down the organization,” Kurian said. “We’ve got a super close working relationship in the Bay Area, in London, in Seattle with the DeepMind team.”