<\/p>\n
(Bloomberg) — In an AI chip industry that\u2019s almost entirely commanded by Nvidia Corp., a Google chip first developed more than 10 years ago especially for artificial intelligence tasks is finally gaining momentum outside its home company as a way to train and run complex AI models.<\/p>\n
Anthropic PBC on Thursday unveiled a deal with Alphabet Inc.\u2019s Google to supply the artificial intelligence startup with more than a gigawatt of additional computing power, valued in the tens of billions of dollars. The agreement gives Anthropic access to as many as 1 million of Google\u2019s tensor processing units, or TPUs \u2014 the company\u2019s chips that are custom designed to accelerate machine learning workloads \u2014 and expands its use of the internet giant\u2019s cloud services.<\/p>\n
Most Read from Bloomberg<\/p>\n
As AI industry contenders clamber to keep up with runaway demand, they\u2019ve been looking for ways to boost their computing power that don\u2019t hinge on access to Nvidia\u2019s accelerator chips \u2014 both to temper dependence on the chip giant\u2019s expensive products and to mitigate the impact of shortages. While Anthropic is already a TPU customer, the dramatically increased deployment is one of the strongest endorsements yet of Google\u2019s technology, and represents a win for its cloud business, which has long lagged behind Amazon.com Inc. and Microsoft Corp.<\/p>\n
A surge of interest in TPUs is likely to direct the attention of other AI startups and new customers toward Google\u2019s cloud, helping the company leverage its years of investment in the chip.<\/p>\n
Google\u2019s cloud business reported operating income of $2.8 billion in the second quarter, more than double the amount from the same quarter last year. Shares of Alphabet rose slightly in premarket trading on Friday.<\/p>\n
Google\u2019s deal with Anthropic is a \u201creally powerful validation of TPUs,\u201d which could get more companies to try them, said Seaport analyst Jay Goldberg. \u201cA lot of people were already thinking about it, and a lot more people are probably thinking about it now.\u201d<\/p>\n
<\/p>\n
Graphics processing units, or GPUs, the part of the chip market dominated by Nvidia, were created to speed the rendering of graphics \u2014 mainly in video games and other visual-effects applications \u2014 but turned out to be well-suited to training AI models because they can handle large amounts of data and computations. TPUs, on the other hand, are a type of specialized product known as application specific integrated circuits, or microchips that were designed for a discrete purpose.<\/p>\n
Google began working on its first TPU in 2013 and released it two years later. Initially, it was used to speed up the company\u2019s web search engine and boost efficiency. Google first began putting TPUs in its cloud platform in 2018, allowing customers to sign up for computing services running on the same technology that had boosted the search engine.<\/p>\n
It was also adapted as an accelerator for AI and machine learning tasks in Google\u2019s own applications. Because Google and its DeepMind unit develop cutting-edge AI models like Gemini, the company has been able to take lessons from the AI teams back to the chip designers, while the ability to customize the chips has benefited the AI teams.<\/p>\n
\u201cWhen we built our first TPU-based system a little bit over 10 years ago now, it was really about solving some internal scaling challenges we had,\u201d said Mark Lohmeyer, Google Cloud vice president and general manager of AI and computing infrastructure, in a conference speech in September. \u201cThen when we put that compute power into the hands of our researchers in Google DeepMind and others, that in many ways directly enabled the invention of the transformer,\u201d he said, referring to the pioneering Google-proposed AI architecture that has become the foundation for today\u2019s models.<\/p>\n
Nvidia\u2019s chips have become the gold standard in the AI market because the company has been making GPUs for far longer than anyone else, plus they are powerful, frequently updated, offer a full suite of related software, and are general-purpose enough to work for a wide array of tasks. Yet, owing to skyrocketing demand, they are also pricey and, for the past few years, chronically in short supply.<\/p>\n
TPUs, meanwhile, can often perform better for AI workloads because they are custom designed for that purpose, said Seaport\u2019s Goldberg, who has a rare sell rating on Nvidia shares. That means the company can \u201cstrip out a lot of other parts of the chip\u201d that aren\u2019t tailored to AI, he said. Now in its seventh generation of the product, Google has improved performance of the chips, made them more powerful and lowered the energy required to use them, which makes them less expensive to run.<\/p>\n
Current TPU customers include Safe Superintelligence \u2014 the startup founded last year by OpenAI co-founder Ilya Sutskever, as well as Salesforce Inc. and Midjourney, alongside Anthropic.<\/p>\n
For now, businesses that want to use Google TPUs have to sign up to rent computing power in Google\u2019s cloud. But that may soon change \u2014 the Anthropic deal makes an expansion into other clouds more likely, said Bloomberg Intelligence analysts.<\/p>\n
\u201cGoogle\u2019s potential deal with Anthropic suggests more commercialization of the former\u2019s tensor processing units beyond Google Cloud to other neo-clouds,\u201d BI\u2019s Mandeep Singh and Robert Biggar wrote in a note Wednesday, referring to smaller companies offering computing power for AI.<\/p>\n
To be sure, no one \u2014 including Google \u2014 is currently looking to replace Nvidia GPUs entirely; the pace of AI development means that isn\u2019t possible right now. Google is still one of Nvidia\u2019s biggest customers despite having its own chips because it has to maintain flexibility for customers, said Gaurav Gupta, an analyst at Gartner. If a customer\u2019s algorithm or model changes, GPUs are better suited to handle a wider range of workloads.<\/p>\n
Key Banc analyst Justin Patterson agrees, saying tensor processing units are \u201cless versatile\u201d than the more general-purpose GPUs. But the Anthropic deal demonstrates both that Google Cloud is gaining share and that TPUs are \u201cstrategically important,\u201d Patterson wrote in a note to clients.<\/p>\n
The latest version of Google\u2019s TPU, called Ironwood, was unveiled in April. It\u2019s liquid-cooled and designed for running AI inference workloads \u2014 meaning using the AI models rather than training them. It\u2019s available in two configurations \u2014 a pod of 256 chips or an even larger one with 9,216 chips.<\/p>\n
Veterans of the TPU work at Google are now leading chip startups or key projects at other large AI companies. Inference-chip startup Groq is helmed by Jonathan Ross, who began some of the work that became TPU. Other people who worked on Google\u2019s TPU include Richard Ho, vice president of hardware at ChatGPT developer OpenAI, and Safeen Huda, who joined OpenAI to work on hardware and software codesign, according to his LinkedIn.<\/p>\n
By helping TPUs proliferate as AI workhorses, these former Googlers continue to spread the internet company\u2019s influence across the AI industry. Those at Google tout the years of work as a key driver of the success of their product.<\/p>\n
\u201cThere really is no substitute for this level of experience,\u201d Google\u2019s Lohmeyer said in September.<\/p>\n
(Updates with Google cloud figures and premarket shares in the fifth paragraph)<\/p>\n
Most Read from Bloomberg Businessweek<\/p>\n
\u00a92025 Bloomberg L.P.<\/p>\n","protected":false},"excerpt":{"rendered":"(Bloomberg) — In an AI chip industry that\u2019s almost entirely commanded by Nvidia Corp., a Google chip first…\n","protected":false},"author":2,"featured_media":98385,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[365,6166,363,364,8187,3830,367,13070,11751,111,139,69,72417,145,72416],"class_list":["post-98384","post","type-post","status-publish","format-standard","has-post-thumbnail","category-artificial-intelligence","tag-ai","tag-anthropic","tag-artificial-intelligence","tag-artificialintelligence","tag-cloud-services","tag-computing-power","tag-google","tag-google-cloud","tag-google-deepmind","tag-new-zealand","tag-newzealand","tag-nz","tag-processing-units","tag-technology","tag-tensor-processing-units"],"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/posts\/98384","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/comments?post=98384"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/posts\/98384\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/media\/98385"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/media?parent=98384"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/categories?post=98384"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/tags?post=98384"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}